Patent Recommendation Based on Enterprise Demand Classification and Supply-Demand Matching

Abstract

Effective patent recommendation plays a crucial role in bridging the gap between enterprise technological demands and patent supply. However, semantic mismatches and incomplete demand expressions often hinder accurate supply–demand matching. This research proposes a demand-driven patent recommendation method. First, content analysis and topic clustering were used to construct an enterprise demand element system, dividing the demand content into five elements: materials, methods, efficacy, products, and applications. Based on the completeness of these elements, enterprise demands were further classified into explicit and implicit types. Second, an enterprise technical problem space and a patent solution space were established, identifying ten types of enterprise technical problems and fifteen types of patent solution categories. These were connected through supply–demand elements to build corresponding correlation systems for explicit and implicit demands. Finally, according to different types of supply–demand correlations and demand characteristics, differentiated patent recommendation methods were designed. Taking various demands in the lithium battery industry as empirical cases, the results show that the proposed method based on demand classification and supply–demand element association effectively achieves accurate patent matching and addresses the challenges caused by incomplete demand information. The study provides an intelligent, content-based recommendation pathway for enterprise technology acquisition and patent transformation, offering theoretical and practical significance for enhancing patent commercialization and improving the efficiency of technological achievement transformation.

1. Introduction

In recent years, the accelerating pace of technological innovation has driven an unprecedented surge in global patent applications, with annual filings now exceeding several million worldwide. This explosive growth has created both vast opportunities and considerable challenges for enterprises seeking technological solutions [1]. As essential carriers of technical knowledge, patents contain rich information and innovative ideas that can serve as valuable sources for addressing enterprise technology needs [2]. Yet, many enterprises lack the expertise to effectively navigate patent databases or to articulate their requirements using standardized patent terminology [3]. Consequently, there is an urgent need for intelligent, demand-driven patent recommendation systems that can adapt to varying levels of demand expression completeness. Such systems would enable enterprises to efficiently identify relevant patents, reduce screening time, and improve recommendation precision, thereby facilitating effective technology acquisition, accelerating innovation-driven industrial development, and enhancing the commercialization and economic value of patents.

Recent research on patent recommendation systems can be grouped into three main approaches: content-based methods that use patent document features, collaborative filtering methods, and hybrid techniques that combine multiple algorithms [4,5,6]. Methodologically, studies have evolved from graph-based models to deep learning approaches and hybrid frameworks [7]. While these studies demonstrate that semantic modeling and neural architectures can improve recommendation accuracy, most existing systems neglect a crucial aspect—the varying completeness of enterprise technology demand expressions [8]. In practice, enterprise demands often contain non-standardized terminology and incomplete descriptions, whereas patent texts use formal, highly structured language [9]. This mismatch creates a semantic gap that hinders effective supply–demand alignment. The conventional one-size-fits-all approach fails to consider the heterogeneity of enterprise needs.

Despite extensive research on patent recommendation algorithms, three critical research gaps remain. First, existing studies lack systematic frameworks for categorizing and processing different types of technology demands based on expression completeness, treating all demands uniformly regardless of their information richness [10]. Second, the semantic gap between enterprise expressions and patent language has not been adequately addressed through differentiated processing strategies—current approaches either rely solely on text similarity calculations with word embeddings or adopt overly complex ontology-based methods, neither of which effectively accommodates the diversity of demand expressions [11]. Third, the theoretical foundation linking enterprise technical problems to patent solutions remains underdeveloped, with insufficient integration of supply-demand theory and invention principles to guide recommendation design.

To address the aforementioned research gaps, this study develops a demand-driven patent recommendation framework with three primary objectives:

(1)

to establish a systematic demand classification method based on the completeness of content elements;

(2)

to construct a supply–demand mapping system that links ten categories of enterprise technical problems with fifteen categories of patent solutions;

(3)

to design differentiated recommendation algorithms—using a BERT-based approach to evaluate content similarity and element coverage for explicit demands, and a dual-pathway strategy combining BERT and BM25 for implicit demands.

Through empirical validation in the lithium battery domain, this study demonstrates that the proposed differentiated strategies significantly enhance the matching accuracy for heterogeneous demand expressions.

This research makes several theoretical and practical contributions to the field of patent recommendation. First, it proposes a novel element-based classification method for enterprise technological demands. By distinguishing explicit and implicit demands according to element completeness, the framework supports the development of targeted recommendation strategies tailored to specific demand characteristics. Second, grounded in supply–demand theory and TRIZ inventive principles, the study constructs a comprehensive supply–demand mapping system that categorizes ten types of enterprise technical problems and fifteen types of patent solutions, and establishes their structural associations. Third, the proposed dual-pathway recommendation algorithm, through differentiated processing strategies, bridges the semantic gap between enterprise demand expressions and patent language, thereby improving the accuracy and relevance of patent recommendations.

2. Literature Review

2.1. Theoretical Foundations of Patent Supply-Demand Matching

In existing studies on patent supply–demand matching, three core components are typically involved: the technology demand side, the patent supply side, and the matching process. The 5W1H analytical framework (Why, What, Who, When, Where, How) provides a systematic approach to examining patent supply–demand matching problems [12]. This framework aims to address several key questions: who is the demander (Who), why the demand arises (Why), who is the supplier (Who), what the demand entails (What), what the supply provides (What), how the matching process is achieved (How), when the demand occurs (When), and where to obtain information about the demand and supply sides (Where).

First, it is essential to identify the demand actors and their corresponding technological problems. In real production settings, most enterprises find it difficult to independently complete complex technological innovations, particularly small and medium-sized enterprises, whose innovation activities are constrained by limited human capital, funding, and R&D capacity. Consequently, acquiring advanced technologies, knowledge, and resources from external environments has become an important means for enterprises to enhance their innovation capability and competitiveness. According to open innovation theory [13], organizations can achieve technological innovation through three main external approaches:

(1)

Imitative innovation, which involves introducing advanced foreign technologies, products, or models and then re-innovating through assimilation and adaptation;

(2)

Collaborative R&D, in which enterprises, universities, and research institutes jointly conduct R&D, leveraging the knowledge strengths of research institutions and the industrial transformation capabilities of enterprises to achieve breakthroughs;

(3)

External technology acquisition, which refers to obtaining technologies from other organizations or individuals through introduction, transfer, or purchase. Patents, in this context, serve as the primary carriers and commercialization forms of technological achievements.

Therefore, enterprises are the principal initiators of technological demand. Nascimento and Zawislak integrated innovation theory with transaction cost theory and proposed the concept of innovation capability complementarity [14], suggesting that enterprises can overcome the limitations of internal innovation by purchasing technologies or engaging in cooperative innovation to acquire external technologies that generate synergistic performance.

Second, it is necessary to clarify the theoretical basis of supply–demand matching. The matching process is fundamentally grounded in supply–demand theory. In neoclassical economics, mathematical models are used to quantify and evaluate supply–demand relationships, analyze determinants, equilibrium conditions, and interactions from an economic perspective, and make supply–demand analysis a powerful tool for explaining economic phenomena [15]. In the field of recommendation systems, supply–demand matching is regarded as an extension and application of supply–demand theory within information systems. The core function of recommendation systems is to achieve intelligent matching between supply and demand through technical means. This optimization process is analogous to the price mechanism that balances supply and demand in economics, but in recommendation systems, the “equilibrium” is realized through algorithmic optimization rather than price adjustment.

Essentially, the supply–demand matching process is also a problem-solving process. From this perspective, an enterprise’s technology demand represents a statement of a technical problem, while patents provide corresponding technical solutions. This process aligns with the Theory of Inventive Problem Solving (TRIZ), which holds that specific technical problems can be abstracted into standardized inventive problems. By applying general inventive principles, feasible solutions can be identified and then refined into targeted innovative schemes, thus achieving a systematic transformation from problem identification to solution generation [16].

Building upon this understanding, achieving dynamic balance in supply–demand matching has become a key issue in patent recommendation research. Two-sided Matching Theory provides the theoretical foundation for maximizing satisfaction and achieving stable matching between two parties. Originating from decision-making problems in market allocation, this theory examines how to consider the preferences of both sides in the matching process to reach a stable equilibrium and maximize overall satisfaction [17]. Patent recommendation, as a representative application of knowledge and technology supply–demand matching, involves three key entities: the patent supplier, the demander, and the candidate patents. Based on two-sided matching theory, the patent recommendation process should take matching degree as the central criterion and fulfilling enterprise technological needs as the ultimate objective, ensuring that recommendation results are rational and acceptable to both sides.

Supply data primarily come from patent databases such as Derwent Innovation, IncoPat, and PatSnap, which provide structured and searchable patent information. However, there is currently a lack of authoritative and centralized platforms for collecting demand data, as enterprise technology needs remain fragmented and difficult to obtain. This fragmentation poses a significant challenge to research on supply–demand matching. With the implementation of China’s Intellectual Property Power Strategy, various provinces have established public service platforms for technology transfer and commercialization, aimed at collecting both enterprise technology demands and transferable patent achievements. These platforms form a robust foundation for data acquisition. For example, the Henan Province Technology Transfer and Commercialization Public Service Platform (http://www.nttzzc.cn/cement/index.html#/Index, accessed on 1 January 2025) has built comprehensive databases for technological achievements and enterprise technology demands, covering 15 technological domains such as new energy, electronic information, and modern agriculture, and includes more than 6000 enterprise technology demand records. The data are authentic, standardized, and accessible. In this study, data from such provincial-level public service platforms are used as the primary source for enterprise technology demands to support the research on patent supply–demand matching and recommendation.

2.2. Methodological Advances in Patent Supply-Demand Matching

(1)

Patent recommendation methods integrating NLP and semantic similarity.

Techniques such as word embeddings, topic modeling, and sentiment analysis in NLP are effective tools for extracting information from patent texts. With the assistance of machine learning algorithms, models can learn patent relationships and perform personalized recommendations [18]. Helmers et al. compared models such as Word2vec and Doc2vec for semantic extraction of Chinese patent texts and their impact on patent recommendation [19]. Chen et al. proposed a word-embedding–enhanced method that leverages patent text similarity for recommendation [20]. Trappey et al. applied the Doc2vec model to vectorize patent texts and calculate semantic similarity, automatically recommending highly relevant patents to target users and ranking results based on similarity [11]. This category of methods can reduce expert involvement in the recommendation process and typically produces more focused outcomes.

(2)

Network analysis–based patent recommendation methods.

Network analysis, encompassing both social networks and heterogeneous networks, focuses on constructing patent information networks to uncover deep relationships among patents. Community detection, path search, and link prediction are then applied to generate recommendations. Heterogeneous patent information networks incorporate multiple node types and relationships [21]. Recommendations based on such networks are mainly applied to patent transactions and information retrieval. For instance, He et al. constructed a heterogeneous patent transaction network using ten types of relationships, including technical proximity, geographic proximity, and citation links, and designed a recommendation model based on entity relationship sequences [22]. This model can identify potential transaction paths that other models cannot, thereby increasing the diversity of recommendation results. Similarly, Wang et al. integrated data from patent transactions, inventions, citations, ontologies, and content to build a heterogeneous patent information network, and calculated enterprise–patent associations using meta-path similarity [23]. This method more accurately captured user transaction preferences, improving recommendation personalization. However, these approaches depend on manually constructed meta-paths or meta-structures, and their comprehensiveness, accuracy, and the expertise of the designers directly affect the quality of recommendations.

(3)

Deep learning–based patent recommendation methods.

In patent studies, deep learning has been applied to information extraction, classification, citation prediction, similarity assessment, and technology evolution analysis, but its application to patent recommendation remains limited [24]. Lu et al. combined bidirectional RNNs, CNNs, and multilayer perceptrons to build a citation classification model to assist examiners in quickly identifying similar patents [25]. On this basis, Choi et al. proposed a two-stage citation recommendation model: the first stage encodes patent features into dense embeddings of textual and preprocessed metadata, and the second stage uses a trained deep learning model with citation information to rank candidate patents, thereby improving accuracy [26]. Li et al. proposed a semi-supervised learning algorithm to automatically classify patent functional information, extract scenario and technical terms using entity recognition, and generate a patent knowledge space [27]. By mapping functions to contexts, their approach enables cross-domain patent recommendation. Despite these advances, deep learning models often function as black boxes with complex structures and numerous parameters, making the recommendation process difficult to interpret.

In summary, existing patent recommendation studies show three major deficiencies. First, most methods focus solely on patent features and treat all enterprise demands uniformly. They lack a structured framework for classifying demand expressions by completeness, overlooking the heterogeneity between explicit (complete) and implicit (incomplete) technology needs. Second, existing systems rely on one-size-fits-all strategies that fail to adapt to different types of enterprise demands, resulting in limited accuracy and applicability. Third, most algorithms function as data-driven black boxes, without theoretical grounding in supply–demand theory or TRIZ inventive principles, which are essential for bridging the semantic gap between enterprise language and patent terminology.

3. Materials and Methods

3.1. Research Framework

This study proposes a demand-driven patent recommendation method aimed at providing enterprises with patent supply solutions that effectively meet their technological needs. The method addresses key challenges in the patent recommendation process, including semantic misalignment between supply and demand and incomplete demand content. The framework consists of four sequential stages, as illustrated in Figure 1.

Figure 1. Overall framework.

Stage 1: Identification of supply–demand elements.

In this stage, enterprise technology demand texts and patent supply texts are collected separately. A topic clustering method is applied to identify the main constituent elements of both demand and supply content. Based on these results, an enterprise demand element system and a patent supply element system are constructed to provide a structured foundation and a semantic bridge for subsequent supply–demand correlation analysis.

Stage 2: Construction of the enterprise technical problem space and the patent solution space.

On the demand side, real enterprise technology demand texts are analyzed using a combination of content analysis and large language models to automatically extract and classify enterprise technical problems, forming a taxonomy of ten types of enterprise technical problem categories. On the supply side, patent texts are analyzed based on the forty inventive principles of TRIZ. These principles are optimized and reconstructed to establish a representative and discriminative classification system of patent solution types, forming a structured correspondence between enterprise problems and patent solutions.

Stage 3: Supply–demand correlation analysis based on element relationships.

According to the correspondence of supply–demand elements and the characteristics of demand types, the study categorizes supply–demand correlations into explicit correlations and implicit correlations. A hierarchical and progressive correlation pathway is constructed to reveal the matching patterns and logical mechanisms of different demand types, clarifying the direct matching mechanism for explicit demands and the reasoning-based matching mechanism for implicit demands.

Stage 4: Demand-driven patent recommendation.

In this stage, differentiated patent recommendation strategies are designed for explicit and implicit demands. Specific evaluation indicators and scoring methods are developed according to the characteristics of each demand type to quantitatively assess candidate patents and select those that best meet enterprise requirements as recommendation results. Finally, representative enterprise demand cases are used to validate the proposed method and evaluate its recommendation performance, confirming the model’s effectiveness and practical applicability.

3.2. Classification and Identification of Enterprise Technology Demands

According to structural functional theory, elements are the basic units that constitute objective phenomena [28]. Different elements perform specific functions and form a complete system through relatively stable interconnections. Although the specific technological demands of enterprises are diverse, from the perspective of element decomposition, these demands can be broken down into a set of fundamental units with intrinsic logical relationships.

Existing studies have not clearly defined or classified the elements of enterprise technological demands. Current academic research mainly relies on patents and scientific papers as analytical texts, extracting technical elements from them and constructing general classification systems that include dimensions such as products, methods, materials, components or parts, efficacy or performance, technical attributes, application fields, and influencing factors [29]. The technological element systems derived from such scientific and technical literature are characterized by academic rigor and specialization, focusing primarily on explaining the principles of technological innovation and implementation mechanisms.

To reveal the content structure and expression characteristics of enterprise technological demands, this study conducts content analysis based on authentic enterprise demand data. Accordingly, the provincial-level Technology Transfer and Commercialization Public Service Platforms established across China were selected as the primary data source. These platforms are constructed and managed under the supervision of provincial government departments and possess three key advantages: authenticity, comprehensiveness, and public welfare orientation.

First, compared with commercial patent transaction platforms or user-generated information on social media, provincial platforms offer higher data authenticity. Each platform enforces strict qualification review procedures, allowing only legally registered enterprises to submit and publish their technological demands. This mechanism ensures the traceability and reliability of enterprise information.

Second, these platforms demonstrate remarkable advantages in terms of data coverage and industrial representativeness. They aggregate a large number of enterprise technology demand records from multiple key sectors—such as energy, chemical engineering, power, electronic information, equipment manufacturing, and modern agriculture—covering enterprises of various sizes and types. This wide coverage provides a comprehensive reflection of the technological demand characteristics of China’s industrial sectors.

Third, the enterprise demand texts collected from these provincial platforms primarily focus on practical technological problems, transformation bottlenecks, and application objectives. As such, they objectively capture the real needs of enterprises in technological innovation and industrial upgrading, offering high policy relevance and strong application-oriented value.

Therefore, this study collected demand texts from provincial-level public service platforms for technology transfer and commercialization established in Anhui, Fujian, Gansu, Hebei, Henan, Heilongjiang, Hunan, Jilin, Jiangxi, Shanxi, Shaanxi, and Yunnan. These enterprise demand texts were compared with scientific and technical literature such as patents and research papers. The comparison revealed that enterprise technological demands place greater emphasis on practical application issues, focusing on technical challenges, directions for improvement, and expected outcomes, while paying relatively little attention to the principles of innovation or technical implementation paths. This difference in content focus suggests that element systems constructed from scientific literature are inadequate for accurately representing the content characteristics of enterprise technological demands. Therefore, it is necessary to refine and optimize existing technical element classification systems by integrating enterprise demand texts, in order to construct a classification framework better suited to describing enterprise technological demands.

In this study, 5000 enterprise technological demand texts in the new energy field were selected as the research sample. The BERTopic model was employed to perform topic modeling and semantic clustering analysis for extracting the content elements of enterprise technology demands. BERTopic integrates Transformer-based semantic encoding with the class-based TF-IDF (c-TF-IDF) weighting algorithm, which allows the model to preserve high-value semantic information during topic modeling and enhances both topic distinctiveness and interpretability [30]. The model parameters were configured as follows: the text embedding model paraphrase-multilingual-MiniLM-L12-v2 was used, with n_neighbors = 15, n_components = 2, min_cluster_size = 80, and min_samples = 10. These settings achieve a balance between topic coherence and clustering granularity, ensuring that the resulting clusters accurately reflect meaningful patterns in enterprise technological demands.

Based on the clustering results, this study further decomposed enterprise technological demands into five fundamental content elements, which provide a clear structural foundation for subsequent analyses, including demand classification, demand identification, and supply–demand matching. Table 1 summarizes the thematic clustering results of enterprise technological demands, while Table 2 presents the definitions and corresponding meanings of each element category.

Table 1. Thematic clustering results of enterprise technology demands ¹.

Topic	Representative Keywords	Topic	Representative Keywords
1	‘reduction furnace’, ‘magnesium metal’, ‘neodymium–iron–boron’, ‘oil shale’, ‘product quality’	8	‘high-temperature resistance’, ‘thermal stability’, ‘anti-aging’, ‘flame retardancy’, ‘glass wool’
2	‘copper alloy’, ‘rare earth metals’, ‘cemented carbide’, ‘non-ferrous metals’, ‘diamond’	9	‘aluminum alloy’, ‘aluminum profile’, ‘bauxite’, ‘alumina’, ‘gibbsite’
3	‘dimethyl sulfoxide ‘, ‘boron nitride’, ‘calcium carbonate’, ‘magnesium alloy’, ‘caprolactam’	10	‘lithium battery’, ‘lithium-ion’, ‘metallic lithium’, ‘power battery’, ‘electrolyte’
4	‘tensile strength’, ‘boron carbide’, ‘elongation’, ‘elongation at break’, ‘recovery rate’	11	‘concrete’, ‘accelerator’, ‘steel fiber’, ‘fly ash’, ‘admixture’
5	‘sewage treatment’, ‘wastewater treatment’, ‘biodegradation’, ‘water-reducing agent’, ‘harmless treatment’	12	‘polylactic acid (pla)’, ‘copolymer’, ‘acrylic acid’, ‘tetrahydrofuran’, ‘lactide’
6	‘waterproof coating’, ‘waterproof materials’, ‘protective layer’, ‘polyurethane’, ‘sealant’	13	‘solar energy’, ‘magnetofluid’, ‘electromagnet’, ‘bevel gear’, ‘generator’
7	‘sensor’, ‘robot’, ‘automation’, ‘subsystem’, ‘fully automatic’	14	‘glass fiber’, ‘carbon fiber’, ‘cellulose’, ‘ceramic fiber’, ‘fibrous materials’

¹ Enterprise technology demand data were collected from provincial-level public service platforms for technology transfer and commercialization established in Anhui, Fujian, Gansu, Hebei, Henan, Heilongjiang, Hunan, Jilin, Jiangxi, Shanxi, Shaanxi, and Yunnan. Topics were derived from the clustering results of the BERTopic model.

Table 2. Elements of enterprise technology demands: composition, definition and examples ¹.

Demand Element Type	Definition	Example Themes
material	key raw materials, auxiliary materials, or functional components required by enterprises	Topic 2, 3, 9, 11, 12, 14
method	technical implementation paths, including operational processes, improvement measures, or control procedures	Topic5
efficacy	expected functional improvements or performance enhancements	Topic 4, 8
product	final deliverables of technology development, including physical devices (equipment, components) or logical products (software platforms, algorithms)	Topic 6, 10
application	intended deployment scenarios, emphasizing industrial applicability and business relevance	Topic 7, 13

¹ The demand element types and definitions were determined through manual classification, while the example themes were identified based on the clustering results of the BERTopic model.

Ideally, a complete and well-defined enterprise technology demand should include all five types of elements, thereby providing a comprehensive representation of the enterprise’s technological requirements. In practice, however, the level of detail in such demand disclosures varies considerably, and some demands may lack one or more elements. In the field of economics, Luo Yongtai classified consumer demand into explicit, semi-implicit, and implicit types based on consumers’ levels of information cognition and price sensitivity when analyzing market demand. Building upon this theoretical framework, the present study adapts and extends it to the context of enterprise technological needs by considering the completeness of demand content elements [31]. Accordingly, enterprise technology demands are categorized into explicit and implicit types. Explicit technological demands contain complete descriptions of all five elements, clearly articulating the technical problem, implementation method, required materials, performance indicators, target product, and application scenario. In contrast, implicit technological demands are characterized by missing elements and provide only partial information, reflecting an enterprise’s incomplete understanding or imprecise expression of its technological requirements.

For explicit technology demands, enterprises can provide detailed descriptions of the five content elements: material, method, efficacy, product, and application. The terms corresponding to each element carry clear technical meanings and effectively reflect the enterprise’s focal concerns and expected objectives regarding the required technology. Therefore, this study adopts a direct extraction method to identify representative technical terms within the five element categories and constructs a complete demand expression framework.

For implicit technological demands, the absence of one or more content elements often leads to identification results that are imprecise, unclear, or incomplete. The objective of identifying such demands is not to reconstruct the full technical details but rather to infer the potential categories of technological needs by leveraging multi-source data related to enterprise technological activities. To support the identification of implicit technological demand categories, this study incorporates three types of auxiliary data that are closely associated with enterprise technological activities: (1) Enterprise basic information. This refers to background data such as the industry to which the enterprise belongs and its main business activities. These data serve as the basis for determining the enterprise’s technological background and related fields. By analyzing the enterprise’s industry position and core business direction, it becomes possible to infer the likely application scenarios and technological domains, thereby providing contextual information for understanding enterprise technological needs. (2) Enterprise patent data. Patent data reflect an enterprise’s R&D achievements and indirectly reveal its technological innovation focus. They represent the technological foundation of what an enterprise is capable of accomplishing (“what it can do”). Such data provide valuable references for understanding an enterprise’s technological capabilities and strategic R&D orientation. (3) Enterprise technology recruitment information. This refers to publicly released information on job postings that express an enterprise’s demand for specific technical expertise. These recruitment postings typically list required technical skills and competencies, which reveal the technologies that enterprises are currently prioritizing. Moreover, they indirectly indicate potential deficiencies or capability gaps in the enterprise’s ongoing R&D activities.

(1)

Enterprise basic information: identifies the technological direction category.

Enterprise technological directions are closely related to their industrial attributes. Based on the national standard Industrial Classification for National Economic Activities (GB/T 4754–2017) [32], this study proposes a text semantic matching method that combines TF-IDF and cosine similarity to identify the technological direction categories of enterprises [33]. The information on each enterprise’s industry and main business was obtained from its official website and the National Enterprise Credit Information Publicity System. The enterprise text was vectorized and compared with the text entries of the national industry classification using cosine similarity, as shown in Equation (1). Here, $A$ and $B$ denote the TF-IDF vectors of the enterprise text and the industry classification text, while $a_i$ and $b_i$ represent the i-th dimension values of the vectors. The top three industry categories with the highest similarity scores were selected as candidate results. These were further verified manually according to the enterprise’s actual business scope to determine its final technological direction category.

$$Sim\left(A,B\right)={\displaystyle \frac{\sum _{i=1}^n a_i{b}_i}{\sqrt{\sum _{i=1}^n{a}^2}\cdot \sqrt{\sum _{i=1}^n{b}^2}}}$$

(1)

(2)

Enterprise patent data: identifies the technological foundation category.

Enterprise patent data record past R&D investment and innovation activities, providing the basis for identifying technological foundations. In this study, patent data were collected from the IncoPat patent database, using enterprise names as retrieval keywords to obtain all published patents. The titles and abstracts of the retrieved patents were extracted, and the Chinese text was segmented using the jieba tool. A stopword list specific to technological domains (e.g., “method,” “system,” “device,” and other high-frequency general terms) was applied to remove non-informative words, while nouns and noun phrases were retained through part-of-speech tagging. Weighted TF-IDF was used to calculate the term importance scores, and the top ten technical keywords with the highest weights were selected. These keywords reflect the enterprise’s technological foundation and are used in this study as categorical delimiters for identifying implicit technological needs.

(3)

Enterprise technical capability requirements: identify the capability demand category.

This study also uses publicly available recruitment information as a data source, obtained from enterprise websites and major recruitment platforms. Technical recruitment data refer specifically to positions related to R&D, engineering, technical support, and system development. Managerial, administrative, and sales positions were excluded to ensure that the extracted information accurately reflects the enterprise’s technological capability needs. The recruitment texts were analyzed to identify sentences containing patterns such as “responsible for/conduct/complete + [technical task]” and “proficient in/familiar with/capable of/possess + [technical tool or method].” The extracted technical tasks and tools/methods were recorded as terms representing the enterprise’s technical capability requirements.

3.3. Analysis of Supply-Demand Relationships

3.3.1. Construction of the Enterprise Technical Problem Space and the Solution Space

From the perspective of supply–demand theory, “supply” refers to the willingness and ability of the supplier to provide products or services that meet specific needs under the premise that demand exists. In the context of patent supply–demand matching, enterprises, as the demand side, put forward technological requirements, while patents, as the supply side, provide technological solutions that can serve as references or adoption options for enterprises. The relationship between the two is established around technical problems. This section explores the associations between patent supply and enterprise technological demands, taking the technical problem as the bridge connecting supply and demand.

As discussed in Section 3.2, from the perspective of element decomposition, enterprise technological demands consist of five content elements: material, method, efficacy, product, and application. Different types of elements correspond to various technological problems—for instance, substitutability problems of materials, process complexity problems of methods, and performance enhancement problems of efficacy. According to TRIZ theory, inventive problems can be categorized into six major types:

(1)

Contradiction problems—how to improve product quality or functionality without increasing resource consumption;

(2)

Diagnostic problems—how to identify and prevent system defects or deficiencies;

(3)

Trimming problems—how to simplify system structure or functions to reduce costs;

(4)

Analogical problems—how to transfer existing knowledge, technologies, or processes to new contexts;

(5)

Combinatorial problems—how to integrate different existing solutions to form a more optimal one;

(6)

Generative problems—how to propose entirely new technological solutions to satisfy unmet needs.

These types of inventive problems not only guide technological invention and innovation but also correspond to the core issues enterprises face when expressing their technological demands, thereby forming the theoretical foundation for demand identification and classification. However, in practice, enterprises encounter more specific and diverse technical problems.

Building on the six categories of inventive problems in TRIZ theory, this study combines real enterprise technological demands with a large language model–based automatic extraction and human-assisted classification method to identify and categorize enterprise technical problems, thereby revealing the underlying problem types within enterprise technology demands. A total of 9000 enterprise technology demand texts were collected from 12 provincial-level technology transfer and commercialization service platforms, covering the fields of new energy, new materials, advanced manufacturing, biomedicine, and information technology. These data comprehensively represent the technological needs of Chinese enterprises across diverse fields, ensuring the breadth and representativeness of the sample.

Since enterprise technical problems are expressed in sentence form, a large language model was used to extract problem statements from the demand texts. Specifically, a prompt was designed to guide the DeepSeek-R1 model in automatically identifying technical problem sentences within enterprise demands. Filtering rules were constructed using characteristic keywords such as “difficult,” “bottleneck,” “impact,” “poor,” “insufficient,” “low efficiency,” “high cost,” and “poor adaptability” to improve the accuracy of technical problem identification.

Following the automatic classification of enterprise technical problems by the large language model, human-assisted classification was conducted to refine and validate the results. This study expanded upon the six original inventive problem types in TRIZ theory and developed ten categories of technical problems that better align with enterprise technological demands. These problem categories correspond to the five content elements of enterprise technological demands, collectively forming the Enterprise Technical Problem Space (Question-Space), which characterizes the contradictions and needs underlying enterprise technology demands, as shown in Table 3.

Table 3. Classification of enterprise technology requirements and corresponding elements ¹.

ID	Type of Technical Problem	Corresponding Enterprise Demand Elements	Connotation of Technical Problem
Q1	material property problem	material	issues related to the structure and performance of materials, affecting overall product durability and functional realization.
Q2	process complexity problem	method	problems concerning process organization and operational difficulty during implementation, leading to low efficiency or poor consistency.
Q3	efficacy and performance limitation problem	efficacy	limitations in achieving the final functional output or performance indicators of the technical solution, making it difficult to meet enterprise objectives.
Q4	functional integration and system optimization problem	product, method	issues in the coordination of functional units and structural design within a system, restricting overall solution efficiency.
Q5	adaptability and environmental compatibility problem	application, material	problems concerning the usability and stability of technical solutions under different working environments.
Q6	insufficient intelligence and automation problem	method, efficacy	low levels of intelligence in control, recognition, and decision-making processes, limiting improvements in production efficiency.
Q7	safety and risk control problem	method, efficacy	safety risks that may arise during the operation of the technical solution.
Q8	green, environmental protection, and energy-saving problem	material, efficacy, method	inadequate response to requirements for energy saving, emission reduction, and green manufacturing.
Q9	compatibility problem in emerging applications	application, product	lack of adaptability of the technical solution to emerging scenarios or cross-domain applications.
Q10	resource acquisition and material substitution problem	material	difficulty in acquiring key materials, core components, or special resources, or risks of unsustainability.

¹ The types of technical problems, corresponding enterprise demand elements, and connotations of technical problems were determined through a combination of large language model identification and manual classification and refinement.

In the supply side analysis, this study aims to identify the content elements of patent supply. To ensure consistency with the method used for identifying enterprise demand elements, the BERTopic model was applied to perform topic clustering analysis on patent texts. A total of 100,000 patents in the field of lithium battery technology were selected as the sample data. Through model-based clustering, the study systematically identified and extracted the main element types of patent supply as follows: (1) Material element: the material basis entities on which patent technologies rely; (2) Method element: the operational processes, technical means, or implementation paths used to achieve technological objectives; (3) Efficacy element: the quantifiable performance outcomes achieved by patent solutions; (4) Product element: the functional carriers delivered by patent technologies; and (5) Application element: the specific application scenarios, deployment environments, or service domains of patent technologies.

In the process of supply–demand matching, patents provide technical responses to enterprise demands at the level of content elements, thereby forming the patent solution space (Solution-Space). This space aggregates patents that embody different inventive principles, reflecting the innovative paths and methodological features through which patents address enterprise technological problems. According to TRIZ theory, patent inventive principles can be summarized into forty classical paradigms, each representing a distinct mode of innovation. However, this classification framework presents two main limitations when applied to the categorization of patent solutions. First, the forty inventive principles are relatively abstract and simplistic, making it difficult to directly align them with the actual technical content of patents. Second, an individual patent often integrates multiple inventive principles, and applying the original TRIZ classification directly can lead to category dispersion, high overlap, and insufficient differentiation. Therefore, based on TRIZ theory and the actual technical content of patents, this study refines and reconstructs the system of inventive principles to develop a more representative and discriminative classification framework for patent supply solutions.

Using the aforementioned patents as the analytical dataset, this study conducts a classification analysis of patent supply solutions. First, a structured prompt template was designed to guide the large language model DeepSeek-R1 in interpreting each patent abstract and automatically matching it with the forty inventive principles of TRIZ, thereby labeling each patent with its corresponding inventive principle category. Second, based on the model’s automatic matching results, patents with semantically or conceptually similar inventive principles were merged into unified categories. Finally, with human-assisted review and inductive analysis, the merged categories were refined and named according to their technical characteristics, resulting in fifteen representative and distinct types of patent supply solutions, thus constructing the patent solution space, as shown in Table 4 [34].

Table 4. Classification of patent supply solutions based on TRIZ and corresponding elements ¹.

ID	Type of Patent Supply Solution	Connotation of Solution	Corresponding Patent Element(s)	TRIZ Inventive Principles Basis
S1	new material compounding and modification	enhancing material performance through doping, alloying, or composite coatings	material	extraction, curvature, flexible shells, porous materials, composite materials
S2	functional material substitution and green raw material design	employing non-toxic, renewable, or high-performance substitute materials	material	low-cost substitution, discard or regeneration, change of physical/chemical parameters, inert environment
S3	process simplification and module reconfiguration	optimizing manufacturing steps, simplifying processes, and removing redundant links	method	segmentation, merging, nesting, dynamization
S4	intelligent control and automatic adjustment algorithms	applying ai, sensors, or models for automatic monitoring, regulation, and response	method	periodic action, feedback, self-service, copying
S5	multi-scenario coupled process optimization	realizing composite processing through coordinated control under thermal, electrical, magnetic, or mechanical fields	method	equipotentiality, change of physical/chemical parameters, phase transition
S6	precision parameter regulation and performance enhancement	improving product performance through precise control of input variables	efficacy	asymmetry, preliminary counteraction, anticipation, continuity of useful action
S7	energy-saving and efficiency management technologies	optimizing system energy pathways, designing heat recovery, and reducing input-output losses	efficacy	turning harm into benefit, discard or regeneration, thermal expansion
S8	functional integration and structural fusion design	integrating multiple functional modules into a single structural unit to improve overall performance	product	combination, merging, dimensional change, porous materials
S9	lightweight and flexible structural innovation	designing thin-walled, hollow, or foldable structures to achieve lightweight, high-strength, or flexible features	product	asymmetry, curvature, dimensional change, flexible shells
S10	fault diagnosis and safety control schemes	introducing fault tolerance, intelligent detection, and emergency response to enhance system reliability	product	preliminary counteraction, anticipation, continuity of useful action, inversion
S11	cross-scenario adaptation and environmental adjustment mechanisms	adjusting system design to suit different application environments	application	mechanical vibration, pneumatic and hydraulic structures, phase transition, inert environment
S12	general platforms and standardized interface design	developing universal interfaces, modules, or protocols for multi-device and multi-platform applications	application	universality, weight compensation/neutralization, use of intermediary, low-cost substitution
S13	application migration and multi-domain expansion	extending existing technical solutions to new industries, scenarios, or application domains	application	incomplete or excessive action, inversion, turning harm into benefit
S14	service function enhancement and human–machine collaborative design	introducing auxiliary operations, remote monitoring, or collaborative interaction to improve user experience	product, application	self-service, copying, periodic action
S15	lifecycle closed-loop management and recycling design	providing full-process design from manufacturing to recycling to promote green lifecycle development	application	discard or regeneration, change of physical/chemical parameters, low-cost substitution, inert environment

¹ Patent supply data were analyzed using TRIZ inventive principles as the theoretical foundation. The types of patent supply solutions, connotations of solutions, and corresponding patent elements were determined through a combination of TRIZ-based classification and manual optimization and definition.

Based on the decomposition of supply-demand elements and the analysis of technical problems and solutions, a correspondence exists between the demand side of enterprises and the supply side of patents in terms of element structure and problem-solving logic. The essence of supply-demand matching lies in establishing a logical correspondence between technical problems and solutions through content elements, thereby constructing a complete mapping system from demand problems to patent solutions.

3.3.2. Supply-Demand Correlation Based on Element Relationships

Integrating the logic of supply–demand matching, the concept of element decomposition, and the TRIZ theory, this study classifies supply–demand correlations into two categories, explicit correlation and implicit correlation, according to the correspondence of supply–demand elements and the type of demand. A hierarchical and progressive correlation path is established to illustrate how the completeness of enterprise demand expression influences the pattern of supply–demand matching.

(1)

Explicit supply-demand correlation

The explicit supply–demand correlation focuses on explicit technological demands of enterprises. Such demands possess a complete structure of five content elements, allowing for point-to-point alignment with patent supply elements at the content level.

Explicit supply–demand correlation is established on the basis of element correspondence. Both enterprise demands and patent supplies contain five categories of content elements, forming direct matching relationships in terms of content, as illustrated in Figure 2. In terms of element characteristics, this correlation emphasizes the symmetry and correspondence between supply and demand elements across content dimensions. Each demand element proposed by the enterprise can find a corresponding supply element in the patent data, while the patent supply provides responses within the same element dimension, thereby achieving direct content-level matching from the technical problem to its corresponding solution.

Figure 2. The logical path of explicit correlation between supply and demand.

From a theoretical perspective, explicit supply–demand correlation reflects the fundamental principle of supply–demand theory and aligns with the logical framework of technical problem solving. Both sides achieve effective matching through element-level mapping between enterprise technological demands and patent supplies.

(2)

Implicit supply-demand correlation

Implicit correlation arises when enterprise technology demands are incomplete in terms of elements. In such cases, semantic matching between supply and demand is achieved through category supplementation, problem identification, and solution reasoning. This type of correlation targets implicit technology demands, whose central characteristic is the absence of direct element correspondence. Thus, multi-source information and reasoning mechanisms are required to uncover potential supply categories for matching.

In practice, to effectively address implicit technology demands, the original demand expressions must undergo category supplementation, problem abstraction, and element mapping, gradually transforming fuzzy expressions into content that can be matched by patent supply. This study further divides implicit correlation into two types: demand expansion-driven correlation and technical problem-driven correlation.

➀

Demand expansion-driven implicit correlation

This type relies on multi-source data to supplement implicit demand categories and construct demand content that can be responded to by patent supply. The correlation chain is expressed as “Enterprise implicit technology demand → Limited element identification → Demand category supplementation → Patent supply response,” as shown in Figure 3.

Figure 3. Implicit correlation between supply and demand driven by demand expansion.

In this pathway, the supply–demand correlation consists of four main steps. First, element information is extracted from the enterprise’s implicit technology demand texts. Second, based on the implicit technology demand category supplementation method established in Section 3.2 the categories of implicit technological demands are defined. Third, the supplemented demand categories are combined with the identified elements to serve as the basis for patent supply identification. Finally, patents that meet the enterprise’s needs are matched by using both the demand categories and elements as filtering conditions, thereby achieving a responsive link between fuzzy enterprise demands and patent supply.

In terms of element characteristics, the demand expansion–driven implicit correlation pathway exhibits asymmetry between supply and demand elements. Enterprise implicit technology demands usually contain only partial demand elements, making it difficult to achieve one-to-one correspondence with the five categories of patent supply elements. Therefore, it is necessary to compensate for the missing elements through demand category expansion.

From a theoretical perspective, the demand expansion–driven implicit correlation pathway reflects the core concept of “demand identification and supply adaptation” in supply–demand theory. This pathway emphasizes supplementing unexpressed technological needs through external information and transforming them into a basis for patent supply–demand matching.

②

Technical problem-driven implicit correlation

This type applies when enterprise implicit demands are expressed in vague terms. The process first identifies potential types of technical problems, and then uses these problem types together with demand elements as intermediaries to infer corresponding patent solutions. The correlation path follows progressive steps: “Enterprise implicit technology demand → Technical problem identification → Element attribution → Solution mapping → Patent supply matching,” as illustrated in Figure 4.

Figure 4. Implicit supply-demand correlation driven by technical issues.

In this pathway, the establishment of the supply–demand correlation involves four main steps. First, the types of technical problems are identified from the enterprise’s implicit technology demand texts. Second, these technical problems are mapped to their corresponding demand content elements, achieving an initial alignment between problems and demand elements. Third, once the demand elements are defined, their corresponding supply elements on the patent side are identified and mapped to the fifteen categories of patent supply solution types, enabling reasoning and inference of potential patent solutions. Finally, patents with the capability to respond to enterprise demands are identified based on the corresponding patent supply solution types, thus achieving a reasoning-based matching process from demand problems to technological solutions.

In terms of element characteristics, the technical problem–driven implicit correlation pathway also exhibits asymmetry between supply and demand elements. However, this pathway is primarily driven by technical problems, which serve as intermediaries connecting the demand and supply sides. By locating the technical problem, the related demand elements can be determined, and subsequently, the corresponding patent supply elements can be matched. This process forms a supply–demand reasoning chain of “demand problem → demand element → supply element → solution”.

From a theoretical perspective, this pathway embodies the core concept of TRIZ theory, which focuses on deriving inventive principles and finding solutions based on identified technical problems. The technical problems summarized from enterprise implicit demands are addressed through element correspondence and solution inference, thereby providing logical support and interpretability for the matching of implicit technological demands.

In summary, the correlation between enterprise technology demands and patent supply is jointly influenced by the explicitness of demand expression, the completeness of content elements, and the responsiveness of patent supply. Explicit correlation applies to explicit enterprise demands, where the supply side can respond directly through element correspondence. Implicit correlation applies to implicit enterprise demands, which require information supplementation, problem identification, and reasoning mechanisms to construct a viable supply-demand matching chain.

3.4. Research on Demand-Driven Patent Supply Information Retrieval and Recommendation Methods

3.4.1. Patent Supply Identification and Retrieval Under Demand Orientation

Before implementing patent recommendation, it is necessary to construct a supply set of patents tailored to specific technological demands. From a large pool of patents, a subset with potential semantic relevance to a particular demand must be identified, serving as the foundation for subsequent supply-demand correlation scoring and recommendation. Centered on the core logic of “selecting supply based on demand,” this study, in combination with supply-demand mapping relationships, the content element framework, and patent solution classifications, proposes identification and retrieval strategies for patent supply targeting different types of enterprise technological demands.

(1)

Patent supply identification for explicit technological demands

Explicit demands enable enterprises to clearly articulate their specific requirements in terms of materials, methods, efficacy, products, and applications. Such demands provide the necessary basis for direct identification of corresponding patents. For this type of demand, this study introduces a demand organization method centered on the “Demand Lexical Tree.” This method constructs a hierarchical and extensible terminology system of technological demands to support patent supply identification driven by explicit demands.

The “Demand Lexical Tree” is grounded in lexical field theory and knowledge ontology hierarchy. It takes the five core content elements of enterprise technological demands as the primary branches and builds multi-level sets of terms organized by hypernym-hyponym relations, synonymous expressions, and semantic associations. Together, these form a multi-layered network of demand terminology, as shown in Figure 5 [35].

Figure 5. Demand lexical tree based on explicit technological needs of enterprises.

Construction of the Demand Lexical Tree combines manual annotation with large language model-assisted classification. First, demand terms appearing under the five content elements are manually organized and preliminarily classified to ensure accuracy and representativeness. Second, a large language model is employed to expand the terminology, guided by prompts to understand the semantic and contextual attributes of each demand. This expansion generates synonymous expressions, scenario-specific substitutes, and hierarchical extensions of demand terms, assisting in determining hypernym-hyponym or co-level classification relationships.

The primary purpose of constructing the Demand Lexical Tree is not to expand the coverage of potential patent supply, but to clarify and constrain the semantic boundaries of patent retrieval. This improves the adaptability and efficiency of patent filtering while enhancing the consistency of demand expressions within the same technological field.

(2)

Patent supply identification for implicit technological demands

Implicit demands are characterized by vague expressions and missing elements, making it difficult to directly identify corresponding patents through keyword matching. Based on the two types of supply-demand correlations, this study designs a patent supply identification strategy for implicit demands, which integrates category supplementation, content element mapping, and solution reasoning.

First, at the category recognition level, implicit demands are expanded through the category supplementation method, using external data sources such as enterprise basic information, historical R&D records, and capability requirements. This step establishes the semantic boundaries for supply retrieval.

Second, at the supply content expansion level, the limited element information already identified from implicit demands is transformed into corresponding term subsets as a retrieval basis. In parallel, the technical problem types underlying enterprise demands are identified, thereby enabling the conversion of demand expressions into problem statements. Once identified, these problems are mapped to corresponding patent solution categories following the path: “Enterprise implicit technological demand → Technical problem identification → Element attribution → Solution mapping → Patent supply matching.” Based on these mappings, patent keyword sets are constructed according to the TRIZ inventive principles associated with each solution type.

Finally, the method for identifying patent supply targeting implicit demands integrates three key sources of information: (1) limited element information extracted directly from implicit demands; (2) demand categories supplemented through external multi-source data; and (3) patent solution keyword sets derived from TRIZ inventive principles. Collectively, these form a retrieval-oriented keyword system for implicit demands, transforming vague demand expressions into actionable conditions for patent supply identification.

3.4.2. Research on Patent Recommendation Methods Based on Supply-Demand Elements

In view of the differences in element completeness between explicit and implicit technological demands, this study proposes a dual-path patent recommendation framework. For explicit technological demands, the BERT model is employed to construct semantic embedding vectors for the five core content elements, namely materials, methods, efficacy, products, and applications, and to calculate both content similarity and element coverage. BERT is chosen because it effectively captures contextual and bidirectional semantic relationships within technical texts, enabling a more accurate representation of domain-specific language and improving the precision of supply–demand semantic matching. For implicit technological demands, where element expressions are incomplete, a dual strategy combining BERT-based semantic modeling and BM25 keyword retrieval is adopted to compute content similarity and category matching, respectively.

In both types of demand, content similarity serves as the primary indicator. It measures the degree of semantic correspondence between the patent text and the enterprise’s technological demand, reflecting the extent to which the patent content aligns with the element-level semantics of the demand. Because explicit demands contain all five content elements and thus allow direct element-level alignment, element coverage is introduced to evaluate whether a patent comprehensively addresses all five dimensions of enterprise needs, thereby indicating the completeness and adaptability of the proposed technological solution. By contrast, implicit demands lack complete element information. As defined earlier in this study, their demand categories are supplemented through multi-source data. Consequently, category matching is employed to assess the consistency between patents and the supplemented demand categories, including TRIZ-based solution keywords, thereby reflecting the rationality of supply–demand matching at the domain and directional levels.

(1)

Patent recommendation method for explicit technological demands

Explicit demands are expressed with complete and clear coverage of all five elements, making them suitable for element-level matching based on semantic embedding models. In this study, a semantic vector space model is constructed using the BERT model, and the responsiveness of patent supply is quantified from two perspectives: content accuracy and element coverage [36].

➀ 

Embedding construction for explicit demand elements

Based on the five types of content element information extracted from enterprise demand texts in the earlier stage (materials m, methods a, efficacy e, products p, and applications u), an explicit demand element set is constructed. In this set, $Q_i(i\in \{m,a,e,p,u\left\}\right)$ denotes the set of terms under the i-th element type in the enterprise technology demand.

$$Q=\left\{Q_m,Q_a,Q_e,Q_p,Q_u\right\}$$

(2)

The BERT model is employed to perform contextual semantic encoding for each term $t\in Q_i$, yielding its vector representation:

$${\mathrm{v}}_t=\mathrm{B}\mathrm{E}\mathrm{R}\mathrm{T}\left(t\right)$$

(3)

To construct a unified semantic representation for each element type, a mean pooling strategy is applied to aggregate the set of term vectors [37], thereby obtaining the element-level vector representation:

$${\mathbf{v}}_{Q_i}={\displaystyle \frac{1}{\left|Q_i\right|}}\sum _{t\in Q_i}{\mathbf{v}}_t$$

(4)

Finally, each explicit technological demand can be represented as a set of five embedding vectors $\{{\mathbf{v}}_{Q_m},{\mathbf{v}}_{Q_a},{\mathbf{v}}_{Q_e},{\mathbf{v}}_{Q_p},{\mathbf{v}}_{Q_u}\}$, which serves as the core representation basis for supply-demand semantic matching.

➁

Semantic representation construction for patent texts

To obtain the semantic embedding representation $V_{D_j}$ of a patent document $D_j$, this study adopts a deep semantic representation approach. Specifically, the title and abstract fields of each patent $D_j$ are concatenated and then fed into the BERT model to generate a unified semantic embedding vector [38]. This vector is subsequently used for semantic matching with the demand element embeddings. The formulation is as follows:

$$V_{D_j}=\mathrm{B}\mathrm{E}\mathrm{R}\mathrm{T}\left(Concat\left(Title,Abstract\right)\right)$$

(5)

➂

Supply-demand semantic matching score: content accuracy (Accuracy)

Cosine similarity is employed to compute the semantic matching score between the five types of demand elements and each patent text [39]. Specifically, the semantic similarity between the embedding of element $Q_i$ and the semantic representation of patent $D_j$ is calculated as follows:

$$S_{\left(i,j\right)}=\mathit{cos}\left(V_{Q_i},V_{D_j}\right)={\displaystyle \frac{V_{Q_i}\cdot V_{D_j}}{\parallel V_{Q_i}\parallel \cdot \parallel V_{D_j}\parallel }}$$

(6)

Following multi-metric score fusion methods and their applications in patent semantic matching, the similarity scores of the five elements are averaged with equal weights to obtain the overall content accuracy of patent $D_j$, denoted as ${\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}}_j$. This metric evaluates the patent’s overall responsiveness to all element categories at the semantic level and reflects the precision of its content matching:

$${\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}}_j={\displaystyle \frac{1}{5}}\sum _{i\in \left\{m,a,e,p,u\right\}}{S}_{\left(i,j\right)}$$

(7)

➃

Element integrity assessment: coverage

Referring to the definition and evaluation methods of coverage in recommendation studies, this study designs the “Coverage” indicator to evaluate how well a patent responds to the five demand element categories. Coverage is defined as the ratio of the number of element categories for which a patent provides an effective match to the total number of elements:

$${\mathrm{C}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}}_j={\displaystyle \frac{1}{5}}\sum _{i\in \left\{m,a,e,p,u\right\}}\mathbb{I}\left[S_{\left(i,j\right)}>\theta \right]$$

(8)

where $\mathbb{I}[·]$ is an indicator function. If the similarity score $S_{(i,j)}$ exceeds the threshold $\theta$, the element is considered to be effectively matched. This indicator measures the breadth of coverage of enterprise technological demands by a given patent, reflecting the patent’s ability to serve as a solution.

A threshold that is set too high may filter out potentially relevant elements, reducing the comprehensiveness of coverage. Conversely, a threshold that is set too low may introduce noise with insufficient semantic relevance, weakening the discriminative power of the coverage metric. Following common practices in recommendation systems and semantic similarity measures based on cosine similarity [40], this study sets $\theta =0.7$.

⑤

Comprehensive scoring of patent supply

By combining the two indicators, content accuracy (${\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}}_j$) and element coverage (${\mathrm{C}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}}_j$), this study constructs a comprehensive evaluation function for patent supply under explicit demand scenarios. Here, $\alpha$ is a weight parameter. To balance matching precision and element coverage, $\alpha$ is set to 0.5 in this study. This value indicates that both indicators are equally important in the comprehensive evaluation. It ensures that the recommended patents achieve high semantic matching precision while maintaining complete coverage of the five demand elements, thereby enhancing the overall comprehensiveness and reliability of the patent supply evaluation.

$${\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}_j=\alpha \cdot {\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}}_j+\left(1-\alpha \right)\cdot {\mathrm{C}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}}_j$$

(9)

(2)

Patent recommendation method for implicit technological demands

Implicit demands cannot be effectively addressed through direct element-level matching. To tackle this challenge, this study proposes a dual strategy that integrates BERT semantic modeling with the BM25 algorithm, performing selection and recommendation from two perspectives: content accuracy and category matching.

Content accuracy measures the extent to which patent supply responds to the already identified content elements of an implicit demand. Category matching, in turn, is based on the category labels of implicit demands and the keyword sets of patent supply solutions, using the BM25 algorithm to calculate the relevance between patent texts and category keywords.

For the limited element information contained in implicit demands, BERT embedding is employed with a mean pooling strategy to obtain element-level semantic vectors. Semantic similarity scores are then computed between demand element vectors and patent vectors in the BERT semantic space, and their equal-weight average is taken as the content accuracy.

The BM25 algorithm is applied to compute the relevance score between the keyword set of implicit demands and the patent supply texts [41]:

$${\mathrm{B}\mathrm{M}25}_j=\sum _{t\in K}\mathrm{I}\mathrm{D}\mathrm{F}\left(t\right)\cdot {\displaystyle \frac{f\left(t,D_j\right)\cdot \left(k_1+1\right)}{f\left(t,D_j\right)+k_1\cdot \left(1-b+b\cdot \frac{\left|D_j\right|}{avgdl}\right)}}$$

(10)

Here, $f\left(t,D_j\right)$ denotes the frequency of keyword ttt in patent document $D_j$, and $\mathrm{I}\mathrm{D}\mathrm{F}\left(t\right)$ represents its inverse document frequency. The empirical parameters are set as $k_1=1.2,b=0.75$ [42]. This metric reflects the degree of alignment between patent content and domain semantics as well as solution categories, serving as a supplement to content similarity evaluation.

By integrating semantic similarity and category matching, this study constructs a comprehensive scoring function for patent supply under implicit technological demands, denoted as ${\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}_j$.

$${\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}_j=\alpha \cdot {\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}}_j+\left(1-\alpha \right)\cdot \mathrm{B}\mathrm{M}{25}_j$$

(11)

Considering that implicit demands, due to missing elements, rely more heavily on semantic understanding and alignment to achieve effective supply-demand matching, this study increases the weight of semantic similarity and sets $\alpha =0.7$. This setting maintains a certain degree of category coverage while emphasizing the central role of semantic alignment in matching implicit demands.

Finally, all candidate patents are ranked in descending order according to their comprehensive scores, and the top ten patents are selected as the recommendation results. The choice of ten patents is based on three main considerations. First, from the perspective of experimental analysis, selecting ten patents provides sufficient samples to fully verify the effectiveness of the proposed recommendation method while maintaining a concise set that facilitates comparative evaluation. Second, from the perspective of content comparison, the Top-10 range ensures adequate diversity and representativeness, clearly illustrating the quantitative differences among the recommended patents. Third, from a practical application standpoint, enterprises typically focus on the top-ranked patents when assessing recommendation results; therefore, presenting the top ten patents aligns well with actual enterprise decision-making behavior.

4. Experiments

To validate the proposed methods for patent supply identification, retrieval, selection, and recommendation based on supply–demand element associations, this study selects the lithium battery technology field as a representative application scenario for empirical analysis.

On the one hand, lithium battery technology is a core component of the new energy industry, encompassing multiple interdisciplinary domains such as materials science, chemical engineering, energy science, and equipment manufacturing. It features a long technological chain and highly complex element structures, which effectively reflect the diversity of enterprise technological demands.

On the other hand, enterprises in the lithium battery sector demonstrate high levels of technological innovation in their production practices, and their demand expressions are highly diverse—ranging from explicit demands with complete and clearly described elements to implicit demands characterized by missing information and vague semantics. This diversity provides an ideal context for evaluating the applicability and robustness of the proposed methods in identifying different types of technological demands and completing their categorical information.

Therefore, the lithium battery technology field is chosen as the experimental domain, as it is both representative of complex industrial technologies and sufficiently challenging to verify the methodological validity of the proposed framework.

4.1. Patent Supply Retrieval Based on Enterprise Technological Demands

4.1.1. Retrieval of Patent Supply Based on Explicit Technological Demands

Following the core logic of “selecting supply based on demand,” this study selects an explicit technological demand in the lithium battery field as a case example. The five core content elements are extracted from the enterprise demand text and used as the basis for patent supply retrieval, as shown in Table 5.

Table 5. Examples of explicit technology needs and elements¹.

Explicit Technological Demand Case
Demanding Enterprise: Jiangxi **** Cobalt Industry Co., Ltd. 2. (Ganzhou City, China) Demand Content: Joint R&D on key technologies for the green, full-component recycling of spent ternary lithium-ion batteries. With the widespread adoption of new energy vehicles, research on key technologies for the recycling of spent lithium-ion batteries is not only essential for environmental protection but also a critical approach to alleviating shortages of strategic resources, carrying significant practical importance. However, the recycling of spent ternary lithium-ion batteries still faces a series of challenges. Studies on targeted separation of cathode and anode materials and high-value short-process conversion remain insufficient. Therefore, it is urgent to develop key technologies for the targeted separation and full-component recycling of retired ternary lithium battery materials. The objectives include developing one set of precision disassembly equipment for spent battery cathodes and anodes with a processing capacity of ≥1.5 t/h; establishing 3–4 key technologies for green, full-component recycling of spent lithium-ion batteries; achieving a separation rate of ≥98.5% and a purity of ≥99% for cathode and anode active materials; restoring the specific capacity of regenerated graphite anode materials to ≥98%; and reducing the cost of regenerated ternary materials by 20%. Experts in hydrometallurgy and ternary recycling are required for this research.
Material	Spent lithium-ion batteries; retired ternary lithium battery materials
Method	Targeted separation of cathode and anode materials; high-value short-process conversion; hydrometallurgy; ternary recycling
Efficacy	Processing capacity ≥ 1.5 t/h; separation rate of cathode and anode active materials ≥ 98.5%; purity of cathode and anode active materials ≥ 99%; specific capacity recovery of regenerated graphite anodes ≥ 98%; cost reduction in regenerated ternary materials by 20%
Product	Precision disassembly equipment for cathode and anode of spent lithium batteries
Application	Recycling of spent ternary lithium-ion batteries

¹ Data were obtained from the Jiangxi Provincial Online Integrated Service Platform for Scientific and Technological Achievement Transformation (https://kjgl.kjt.jiangxi.gov.cn/cgzyzh/#/DemandHall, accessed on 1 January 2025. ^2. **** indicates that specific information has been omitted to protect enterprise privacy and confidential data.

This study constructed a “Demand Lexical Tree” for explicit technological demands using a strategy that combines manual annotation with large language model-assisted classification. The five content elements were used as the first-level classification nodes to form the main structure of the tree. The large language model DeepSeek-R1 was introduced to perform semantic expansion of demand terms, generating alternative expressions, synonyms, and hypernyms, thereby building a multi-level hierarchical demand tree, as shown in Figure 6.

Figure 6. Examples of explicit technology demand words used by companies.

Based on the Demand Lexical Tree, this study developed retrieval expressions and searched for patent information in the IncoPat patent database. The search was limited to Chinese patents, with the retrieval date set to 1 August 2025. A total of 4318 patent supply records were collected.

4.1.2. Patent Supply Retrieval Based on Implicit Technological Demands

For implicit technological demands, this study also follows the principle of “selecting supply based on demand.” Building on demand expansion-driven and technical problem-driven implicit correlations, patents are retrieved and selected according to implicit demand categories and identified technical problems. The implicit demand case selected in this study and its element recognition results are presented in Table 6.

Table 6. Examples of implicit technical requirements and elements ¹.

Implicit Technological Demand Case
Demanding Enterprise: Henan ** Zinc Industry Co., Ltd. 2. (Jiyuan City, China) Demand Content: Recycling and reuse of new energy vehicle batteries. It is recommended that the company proactively plan to enter the field of new energy battery recycling. In China, new energy vehicles use various types of batteries, such as ternary lithium and lithium iron phosphate, with diverse specifications. After recovery, the recycling and processing technologies are complex and involve lengthy procedures. Many enterprises lack both experience and capability in battery recycling, as well as specialized recycling equipment. Relying on traditional manual disassembly methods may lead to safety risks, including short circuits, fires, or even explosions. Industry experts point out that secondary use of batteries after disassembly, inspection, and classification could reduce the cost of new energy vehicle batteries by up to 60%. The urgency of battery recycling for new energy vehicles can no longer be ignored.
Material	New energy vehicle batteries; ternary lithium; lithium iron phosphate
Method	Not specified
Efficacy	Not specified
Product	Not specified
Application	Recycling of new energy vehicle batteries

¹ Data were obtained from the Henan Provincial Technology Transfer and Commercialization Public Service Platform (http://www.nttzzc.cn/cement/index.html#/Index), accessed on 1 January 2025. ^2. ** indicates that specific information has been omitted to protect enterprise privacy and confidential data.

This implicit demand explicitly specifies only the “material” and “application” elements, while the other elements are missing. According to the implicit demand identification method, its technological demand categories are supplemented through multi-source data, drawing on the following three support layers:

• Enterprise basic information: identifies the technological direction category.

From the National Enterprise Credit Information Publicity System, the basic business scope of the case enterprise was obtained: non-ferrous metal smelting, precious metal smelting, rare and rare-earth metal smelting, manufacturing and sales of non-ferrous metal alloys, production and sales of chemical products, and production and operation of hazardous chemicals. Referring to the industry classification system in the national standard Industrial Classification for National Economic Activities (GB/T 4754—2017) [32], and applying text semantic matching methods, the enterprise’s technical directions were identified as “Non-ferrous Metal Smelting (C32)” and “Basic Chemical Raw Material Manufacturing (C26).”

2.

Enterprise patent data: identifies the technological foundation category.

From the patent database, 187 patents of the enterprise were retrieved. Using the TF-IDF method, the top ten high-weight technical keywords were extracted: non-ferrous metals, wastewater treatment, electromagnetic induction, cooling sleeve, sulfur dioxide, insulating materials, temperature sensors, wastewater treatment, nitrogen oxides, and toxic gases.

3.

Enterprise technical capability requirements: identify the capability demand category.

Job postings released by the enterprise on public platforms were collected, and technical positions were screened and retained, yielding eight valid records. Through syntactic parsing and keyword extraction, the enterprise’s technical capability requirements were identified as metallurgy, applied chemistry, materials chemistry, and metallic materials.

Based on technical problem-driven implicit correlations, the identified technical problem types were mapped to corresponding patent supply solutions. This demand involves three types of technical problems: Q2 (Process Complexity Problem), Q7 (Safety and Risk Control Problem), and Q8 (Green, Environmental Protection, and Energy-Saving Problem). These were mapped to solution types S3–S7, further associated with specific TRIZ inventive principles. The mapping results are shown in Table 7.

Table 7. Patent supply plan mapping results ¹.

Technical Problem Type	Supply-Demand Elements	Patent Supply Solution Type	TRIZ Inventive Principles Basis
Q2 process complexity problem	Method	S3 process simplification and module reconfiguration	segmentation, merging, nesting, dynamization
		S4 intelligent control and automatic adjustment algorithms	periodic action, feedback, self-service, copying
		S5 multi-scenario coupled process optimization	equipotentiality, change of physical/chemical parameters, phase transition
Q7 safety and risk control problem	Method, Efficacy	S6 precision parameter regulation and performance enhancement	asymmetry, preliminary counteraction, anticipation, continuity of useful action
		S7 energy-saving and efficiency management technologies	turning harm into benefit, discard or regeneration, thermal expansion
Q8 green, environmental protection, and energy-saving problem	Material, Method, Efficacy	S1 new material compounding and modification	extraction, curvature, flexible shells, porous materials, composite materials
		S2 functional material substitution and green raw material design	low-cost substitution, discard or regeneration, change of physical/chemical parameters, inert environment

¹ The mapping results in this table were obtained from the integration of multi-source enterprise data. The technical problem types, supply–demand elements, and patent supply solution types were identified through the technical problem–driven implicit correlation method proposed in this study and mapped to the corresponding TRIZ inventive principles.

Finally, based on the integration of multi-source information, a keyword set for patent supply identification was constructed for the enterprise’s implicit technological demand. This set consists of three components: (i) element terms extracted from the original demand text; (ii) category terms obtained from multi-source data; and (iii) solution-related keywords derived from TRIZ theory. Together, these three parts form a complete retrieval system supporting the identification and acquisition of patent supply solutions.

Using the constructed keyword set, retrieval expressions were developed and applied in the IncoPat patent database to search for Chinese patents. A total of 1864 patent supply records were collected, with the retrieval date set to 1 August 2025.

4.2. Patent Selection and Recommendation Based on Supply-Demand Element Associations

4.2.1. Patent Selection and Recommendation for Explicit Technological Demands

Given the characteristics of explicit technological demands, which are complete in elements and clearly expressed, this study employs the BERT model to construct a semantic vector space model. In this study, the BERT-base-Chinese pre-trained model was employed to construct the semantic embedding space for patent–demand matching. The model consists of 12 Transformer encoder layers, 12 self-attention heads, and a hidden size of 768. The tokenizer and model weights were loaded from the Hugging Face checkpoint bert-base-chinese. For missing elements, a zero vector (shape: [768]) was returned to maintain dimensional consistency during semantic encoding. The matching degree of patent supply is comprehensively quantified from two dimensions: content accuracy (Accuracy) and element coverage (Coverage).

For content accuracy evaluation, both the demand and supply sides are considered. The five element terms of enterprise explicit demands are encoded semantically using the BERT model, with mean pooling applied to obtain element-level vector representations. The titles and abstracts of 4318 candidate patents are concatenated and input into the BERT model. Cosine similarity is then calculated between each demand element vector and the patent vectors. The equal-weight average of the five element similarity scores is taken as the Accuracy metric.

For element coverage evaluation, the extent to which patent supply responds to all demand elements is measured. The number of elements with similarity scores greater than the threshold of 0.7 is counted, and the ratio of this number to the total five elements is defined as the Coverage metric. A coverage value of 1.0 indicates that a patent fully responds to all five demand elements of an enterprise’s technological requirement.

Empirical analysis shows that the mean comprehensive score of the 4318 candidate patents is 0.8157, with a standard deviation of 0.0611, a maximum score of 0.9008, and a minimum score of 0.5498. The score distribution is illustrated in Figure 7. The distribution demonstrates a clear concentration of high scores: 97.11% of patents score ≥ 0.75, mainly concentrated in the ranges of 0.75–0.80 (53.33%) and 0.85–0.90 (43.75%). Only 0.02% of patents score ≥0.90, indicating that the vast majority of patents perform excellently in both semantic matching and element coverage.

Figure 7. Patent score distribution for explicit technology needs.

In the comprehensive scoring calculation, content accuracy and element coverage were integrated using a weight coefficient of 0.5. Each patent was assigned a comprehensive score and ranked in descending order. Considering the balance between user acceptance and decision-making efficiency in patent recommendation practice, this study selected the top 10 patents as the recommendation results. This ensures recommendation quality while providing enterprises with sufficient options and comparative references.

The scores of the top 10 candidate patents are presented in Table 8. The results show that high-scoring patents achieve consistently high similarity levels across all five elements, with coverage generally reaching 1.0. This reflects the “dual excellence” feature of semantic matching and element coverage.

Table 8. Top 10 patent supply selection scores for explicit technology requirements ¹.

Rank	Patent_id	Patent Number	Score	Accuracy	Coverage
1	1153	CN103606719A	0.9008	0.8016	1
2	2988	CN113696749B	0.8974	0.7948	1
3	1034	CN114824542A	0.8945	0.7889	1
4	1630	CN107706480A	0.8943	0.7887	1
5	2711	CN117936961A	0.8936	0.7872	1
6	1427	CN105406146A	0.8926	0.7851	1
7	2260	CN218132149U	0.8920	0.7839	1
8	3000	CN118431451B	0.8920	0.7839	1
9	1532	CN110371526A	0.8918	0.7835	1
10	410	CN120221835A	0.8916	0.7832	1

¹ Patent data were retrieved from the IncoPat patent database, and the scoring results were generated through the explicit-demand patent recommendation model developed in this study.

4.2.2. Patent Selection and Recommendation for Implicit Technological Demands

Given the characteristics of implicit technological demands, which are incomplete in elements and insufficient in information expression, this study evaluates and selects candidate patent supply from two dimensions: content accuracy (Accuracy) and technical category matching (BM25_norm).

On the demand side, the limited element terms identified in implicit technological demands are semantically encoded using the BERT model, and element-level vector representations are obtained through mean pooling. According to the technical category keywords corresponding to implicit demands, combined with patent supply solution keywords refined from TRIZ inventive principles, the BM25 algorithm is applied to score candidate patents and generate category matching values.

For the 1864 candidate patents, titles and abstracts were concatenated, and cosine similarity with the demand-side element vectors (Accuracy) as well as BM25 normalized scores with the category keywords (BM25_norm) were calculated. These two metrics were integrated with a weight coefficient of $\alpha =0.7$ to obtain comprehensive scores, which were then ranked in descending order.

Empirical results show that the mean comprehensive score of the 1864 candidate patents is 0.5380, with a standard deviation of 0.0394, a maximum of 0.7520, and a minimum of 0.4483. The distribution of scores exhibits a clear concentration in the medium-to-high range (Figure 8). Among them, 65.56% of patents scored ≥ 0.55, primarily concentrated in the intervals of 0.55–0.60 (45.28%) and 0.60–0.65 (17.80%). Only 1.39% of patents scored above 0.70. Compared with explicit demand scenarios, the overall score levels of patents under implicit demands are lower, reflecting the impact of incomplete element information on supply-demand matching performance.

Figure 8. Patent score distribution for implicit technology needs.

Considering the exploratory nature of implicit demands and the diversified requirements of enterprises in technology selection, the top 10 patents were also chosen as recommendation results, providing enterprises with ample space for comparing technical solutions, as shown in Table 9.

Table 9. Top 10 patent supply selection scores for implicit technology requirements ¹.

Rank	Patent_Id	Patent Number	Score	Accuracy	BM25_Norm
1	356	CN116730315A	0.7520	0.6457	0.9998
2	657	CN116387696A	0.7054	0.6350	0.8698
3	615	CN116514178A	0.6888	0.6291	0.8281
4	639	CN110745802A	0.6784	0.6185	0.8181
5	1218	CN117581396A	0.6687	0.6286	0.7621
6	493	CN106328941A	0.6630	0.6301	0.7396
7	1524	CN118954461A	0.6591	0.6545	0.6697
8	115	CN117263160A	0.6590	0.6448	0.6922
9	544	CN114671424B	0.6581	0.6567	0.6616
10	927	CN114400393A	0.6570	0.6578	0.6553

¹ The ranking and scoring results were generated using the dual-pathway recommendation model proposed in this study.

4.3. Comparison and Evaluation of Patent Supply Selection Methods

To evaluate the effectiveness and practical applicability of the proposed patent recommendation method, this section establishes a multi-dimensional and systematic evaluation framework that assesses the quality of recommendation results from three perspectives: content analysis, expert validation, and method comparison.

First, a comparative analysis is conducted between high-scoring and low-scoring patent recommendation cases to examine the differences in technical content and degrees of alignment with enterprise demands. This analysis not only reveals the internal mechanisms by which patent solutions respond to enterprise technological needs but also explains how each evaluation indicator influences the final recommendation outcomes, thereby providing interpretability for the algorithmic decision-making process.

Second, an independent expert evaluation mechanism is introduced. Domain experts assess the recommendation results from three dimensions—technical relevance, solution feasibility, and recommendation rationality—thus overcoming the limitations of internal algorithmic scoring and ensuring that the recommendations are professionally validated in practical contexts.

Finally, a series of systematic comparative experiments are designed to benchmark the proposed method against multiple baseline approaches. By quantitatively analyzing the performance differences across key evaluation metrics, this study verifies the superiority of the proposed fusion strategy over single-dimensional methods and demonstrates its effectiveness in addressing various types of technological demands.

4.3.1. Content Analysis of Patent Recommendation Results

• Content analysis of recommendation results for explicit demands

For explicit technological demands, both the top-ranked patent CN103606719A and the tenth-ranked patent CN120221835A exhibit a high degree of alignment with the enterprise’s needs. The score difference between the two patents is only 0.0092, indicating that all of the top ten recommended patents effectively address the enterprise’s requirements. A closer content analysis, however, shows that the Top 1 patent demonstrates slightly better consistency between its technical route and the target demand.

(1)

Material element matching analysis

The Top 1 patent explicitly uses “spent lithium-ion batteries” as raw material, which directly corresponds to the demand elements “spent ternary lithium-ion batteries” and “retired ternary cathode materials,” achieving precise material-level alignment. The Top 10 patent also focuses on recycling spent lithium-ion batteries but introduces “spodumene” as an additional raw material, creating a hybrid processing system composed of cathode powder, anode carbon powder, spodumene, and carbon source. Although this multi-material co-processing strategy shows innovation in comprehensive resource utilization, it diverges slightly from the enterprise’s primary goal of “targeted separation of ternary lithium materials.”

(2)

Method element matching analysis

The enterprise demand emphasizes two key technical paths: “targeted separation of anode and cathode materials” and “high-value short-process conversion.” The Top 1 patent applies a sol–gel method using citric acid as both leaching and gelling agent. The process is simple and employs a single solvent system to achieve selective metal leaching and direct reconstruction of cathode material, fully embodying the “short-process conversion” concept. In contrast, the Top 10 patent involves multiple processing stages, including copper separation, ball milling, activation, calcination, and leaching. Although this achieves comprehensive recovery of anode and cathode components, its multi-step process chain is less consistent with the short-process requirement.

(3)

Efficacy element matching analysis

The enterprise demand specifies five performance indicators—processing capacity, separation rate, purity, capacity recovery, and cost reduction. The Top 1 patent notes that “the prepared lithium manganese oxide cathode material can be directly reused in production,” implying that the product meets industrial-grade quality standards, thus aligning with the demand expectations. The Top 10 patent, by contrast, focuses on “comprehensive recovery” and “environmental sustainability” but responds less directly to specific quantitative indicators.

Both patents emphasize environmental friendliness but adopt different technical approaches. The Top 1 patent replaces conventional inorganic acid leaching with citric acid, avoiding the release of harmful gases such as sulfur, nitrogen, and chlorine, and eliminates byproduct treatment in metal separation. The Top 10 patent achieves its environmental objectives through “integrated recovery” and “resource maximization,” though its multi-stage process may entail higher energy consumption and intermediate waste treatment.

(4)

Product and application element matching analysis

The demand explicitly calls for the development of “precision dismantling equipment for spent lithium-ion batteries” and its application in “spent ternary lithium battery recycling.” The Top 1 patent produces a “lithium manganese oxide cathode material,” a high-value product that can be directly reused in manufacturing, aligning with the demand’s goal of “high-value conversion.” The Top 10 patent extends its application scope further by recycling both electrode materials and valuable metals, and by converting copper and silica residues into catalysts, demonstrating broader potential for resource utilization.

In summary, the score difference between the Top 1 and Top 10 patents is minimal, confirming that both are high-quality recommendation results with excellent performance in semantic similarity and element coverage. The Top 1 patent excels in its focused technical pathway, streamlined process design, and precise alignment with the enterprise’s core demand. Meanwhile, the Top 10 patent achieves comprehensive five-element coverage and offers broader value in resource utilization, making it a meaningful alternative for enterprise decision-making.

2.

Content analysis of recommendation results for implicit demands

For implicit technological demands, the score gap between the top-ranked patent CN116730315A and the tenth-ranked patent CN114671424B is more pronounced, although both address the enterprise’s needs in different dimensions. Compared with explicit demand scenarios, implicit demand recommendations exhibit greater variability in technical pathways, reflecting the algorithm’s reliance on semantic expansion and category matching to achieve broader coverage under incomplete information conditions.

(1)

Material element matching analysis

The original demand mentions three material categories: “new energy vehicle batteries,” “ternary lithium,” and “lithium iron phosphate.” The Top 1 patent focuses on “preparing lithium iron phosphate” but takes “non-ferrous metallurgical tailings” as the starting raw material. This material source aligns well with the enterprise’s background in zinc smelting, demonstrating how multi-source data augmentation enables the algorithm to associate “metallurgical residues” with the enterprise’s resource capabilities. The Top 10 patent directly processes “anode and cathode powders,” which aligns more closely with the explicit material expressions in the demand but does not reflect the enterprise’s inherent strengths in metallurgical recycling.

(2)

Method element matching analysis

Because the original demand did not specify a technical method, the algorithm identified three latent technical issues through problem-driven inference: Q2 (process complexity), Q7 (safety and risk control), and Q8 (green and energy-efficient processing). The Top 1 patent adopts a five-step process—acid leaching, solid–liquid separation, metal precipitation, ferric phosphate synthesis, and lithium iron phosphate production—which effectively addresses issues Q2 and Q8. Its reuse of metallurgical tailings resonates with the enterprise’s focus on “comprehensive resource utilization.” The Top 10 patent employs a “hydrothermal treatment” method for separating anode and cathode powders. This approach features a shorter and simpler process chain, directly responding to Q2 and Q7.

(3)

Efficacy element matching analysis

In terms of efficacy, the Top 1 patent emphasizes “improving metal resource utilization efficiency” and “reducing production costs,” semantically aligning with the enterprise’s goal of “reducing new energy battery costs by up to 60%.” The Top 10 patent focuses on “enhancing the quality of regenerated lithium iron phosphate cathode materials” and “improving the utilization of graphite carbon in anodes,” providing solutions for product quality improvement and material reuse but with less emphasis on cost reduction.

(4)

Product and application element matching analysis

The Top 1 patent’s end product is lithium iron phosphate, a core cathode material for new energy batteries, establishing a logical “recycling-to-regeneration” link through TRIZ-based solution mapping. The Top 10 patent produces both “carbon-free lithium iron phosphate powder” and “synthetic gas byproducts” for reuse. While this approach demonstrates product diversity, its application scenarios are less explicitly aligned with the enterprise’s strategic objectives.

The score gap between the Top 1 and Top 10 patents is notably larger than in the explicit demand scenario, underscoring the inherent challenge of implicit demand recommendations. When element information is incomplete, the algorithm relies more heavily on external data enrichment and semantic inference, which increases variability in outcomes. The Top 1 patent’s advantage lies in its deep alignment with the enterprise’s industry background and technical capabilities, demonstrating the effectiveness of multi-source data enhancement. The Top 10 patent, though achieving higher content accuracy, shows lower category matching, resulting in a lower overall ranking. Nevertheless, its advantages in process simplicity and safety make it a valuable technical reference option.

4.3.2. Expert Evaluation of Patent Recommendation Results

Twelve domain experts were invited to conduct the evaluation. Their background characteristics were as follows: 8 males (66.7%) and 4 females (33.3%); mean age 40.7 ± 6.7 years, including 4 experts under 40 (33.3%) and 8 experts aged 40–49 (66.7%). The mean length of professional experience was 10.2 ± 4.6 years, with 5 experts having <10 years (41.7%) and 7 having 10–20 years (58.3%). In terms of education, 7 held doctoral degrees (58.3%) and 5 held master’s degrees (41.7%). Regarding professional titles, 4 were senior (33.3%), 3 were associate senior (25.0%), and 5 were intermediate or below (41.7%). All experts had >5 years of experience in new energy technologies, patent analysis, and intelligence studies, and were familiar with patent recommendation and technology supply–demand matching, thereby providing professional and objective assessments. Because this study focuses on technical matching at the method level, the panel primarily comprised research-institute experts, emphasizing judgments on technical relevance and methodological soundness. This composition ensures technical depth while retaining academic perspective and patent analysis capability.

Drawing on evaluation practices in technology transfer and expert-based criteria used in patent recommendation research, this research designed a three-dimension framework:

(1)

Technical Relevance (TR): the degree to which the patent aligns with the enterprise demand across materials, methods, efficacy, products, and application scenarios.

(2)

Solution Feasibility (SF): implementability in enterprise settings, considering technological maturity, implementation difficulty, fit with existing capabilities, and economic viability.

(3)

Recommendation Rationality (RR): the overall appropriateness of recommending the patent to the demand enterprise, considering solution completeness, potential technical risks, and suitability for introduction or collaboration.

Each dimension was rated on a five-point Likert scale: 1 (strongly disagree) to 5 (strongly agree). The evaluation proceeded in four stages. First, experts were briefed on objectives, indicator definitions, and scoring criteria, with sample demonstrations. Second, experts independently completed two rounds within two days (Round 1: explicit demand; Round 2: implicit demand, separated by two days); each round covered 10 patents and required approximately 1–2 h. Third, inter-rater reliability (Krippendorff’s alpha) was computed to ensure consistency. Finally, authors calculated the mean and standard deviation for each dimension per patent and derived an overall expert score. The items evaluated comprised 10 recommended patents for the explicit demand and 10 for the implicit demand.

Table 10 reports expert scores for the 10 patents recommended for the explicit demand. Overall, expert ratings closely mirrored algorithmic ranking (Spearman’s ρ = 0.915, p < 0.001), indicating a strong association between algorithmic recommendations and expert judgments.

Table 10. Expert scores for recommendations under the explicit demand ¹.

Rank	Patent Number	Mean TR (SD)	Mean SF (SD)	Mean RR (SD)	Expert Overall Score	Algorithm Score
1	CN103606719A	4.67(0.49)	4.42(0.51)	4.83(0.39)	4.64	0.9008
2	CN113696749B	4.67(0.49)	4.33(0.49)	4.75(0.45)	4.58	0.8974
3	CN114824542A	4.50(0.52)	4.25(0.45)	4.67(0.49)	4.47	0.8945
4	CN107706480A	4.50(0.52)	4.08(0.67)	4.50(0.52)	4.36	0.8943
5	CN117936961A	4.33(0.49)	4.25(0.45)	4.42(0.51)	4.33	0.8936
6	CN105406146A	4.33(0.49)	4.00(0.43)	4.33(0.49)	4.22	0.8926
7	CN218132149U	4.25(0.45)	3.92(0.51)	4.25(0.45)	4.14	0.8920
8	CN118431451B	4.25(0.45)	4.25(0.45)	4.25(0.45)	4.25	0.8920
9	CN110371526A	4.08(0.51)	4.00(0.43)	4.17(0.39)	4.08	0.8918
10	CN120221835A	4.17(0.39)	3.92(0.29)	4.25(0.45)	4.11	0.8916

¹ Data source: Evaluation results from 12 domain experts. Expert Overall Score = (TR + SF + RR)/3, SD = Standard Deviation.

Table 11 reports expert scores for the 10 patents recommended for the implicit demand. Compared with the explicit case, the overall level of expert ratings was lower, yet the correlation between algorithmic ranking and expert judgments remained moderate (Spearman’s ρ = 0.758, p = 0.011).

Table 11. Expert scores for recommendations under the implicit demand ¹.

Rank	Patent Number	Mean TR (SD)	Mean SF (SD)	Mean RR (SD)	Expert Overall Score	Algorithm Score
1	CN116730315A	4.08(0.67)	3.83(0.58)	4.17(0.58)	4.03	0.7520
2	CN116387696A	3.92(0.51)	3.58(0.51)	3.83(0.39)	3.78	0.7054
3	CN116514178A	3.75(0.45)	3.50(0.52)	3.75(0.45)	3.67	0.6888
4	CN110745802A	3.67(0.49)	3.33(0.49)	3.58(0.51)	3.53	0.6784
5	CN117581396A	3.67(0.49)	3.25(0.45)	3.42(0.51)	3.44	0.6687
6	CN106328941A	3.50(0.52)	3.33(0.49)	3.42(0.51)	3.42	0.6630
7	CN118954461A	3.33(0.49)	3.50(0.52)	3.33(0.49)	3.39	0.6591
8	CN117263160A	3.33(0.49)	3.33(0.49)	3.33(0.49)	3.33	0.6590
9	CN114671424B	3.25(0.45)	3.08(0.29)	3.17(0.39)	3.17	0.6590
10	CN114400393A	3.42(0.51)	3.67(0.49)	3.33(0.49)	3.47	0.6581

¹ Data source: Evaluation results from 12 domain experts.

To ensure reliability and validity, authors conducted consistency, validity, and difference tests. Reliability analysis showed Krippendorff’s alpha values of 0.72 (explicit) and 0.68 (implicit), both exceeding the 0.67 acceptability threshold, indicating credible evaluation data. Validity analysis revealed a strong positive correlation between algorithmic ranking and expert scores for the explicit demand (ρ = 0.915, p < 0.001) and a moderate positive correlation for the implicit demand (ρ = 0.758, p = 0.011), confirming the effectiveness of the recommendation framework. A paired t-test showed that expert overall scores for the explicit demand (4.32) were significantly higher than for the implicit demand (3.52), t(11) = 10.85, p < 0.001, Cohen’s d = 3.13, demonstrating that completeness of demand elements has a significant impact on recommendation quality and supporting the necessity of the proposed differentiated strategy.

In summary, the expert evaluation corroborates the effectiveness of the demand-driven patent recommendation approach from multiple perspectives. For explicit demands, experts rated the recommendations highly (mean 4.32/5.0), and algorithm–expert agreement was strong (ρ = 0.915), indicating that the fusion of BERT-based semantic modeling with element coverage effectively identifies high-quality solutions aligned with clearly expressed needs. For implicit demands, although overall ratings were lower (mean 3.52/5.0), expert recognition remained evident and algorithm–expert agreement was still significant, suggesting that the dual-path strategy based on multi-source data enhancement and category matching mitigates information gaps and improves practical value. The significant difference between explicit and implicit performance underscores the decisive role of demand-element completeness and validates the rationale for differentiated recommendation strategies. Collectively, these findings show that the proposed framework delivers strong recommendation quality and expert endorsement in practical settings, providing an effective methodological basis for precise matching between enterprise technological demands and patent supply.

4.3.3. Comparative Validation of Patent Supply Selection Methods

This section validates the effectiveness of patent supply selection methods through comparative experiments, using 4318 patents corresponding to explicit technological demands and 1864 patents corresponding to implicit technological demands. The core idea of the comparative experiment is to remove or alter key components of the method, observe their impact on the final results, and thereby verify the necessity and contribution of each component. Four comparative methods were designed for each type of demand, as shown in Table 12.

Table 12. Patent supply selection method dissolution experiment design ¹.

Demand Type	Comparative Method	Core Strategy	Computational Approach
Explicit Technological Demands	Method A	Keyword-based selection only	TF-IDF algorithm to compute similarity between demand element keywords and patent texts
	Method B	BERT-based content accuracy only	Using only the content accuracy metric, computing semantic similarity of five elements with the BERT model
	Method C	Element coverage only	Using only the element coverage metric, counting the completeness of demand element coverage by patents
	Method D	Fusion of content accuracy and element coverage	The proposed strategy in this study combining content accuracy and element coverage
Implicit Technological Demands	Method E	Keyword-based selection only	TF-IDF algorithm to compute similarity between demand element keywords and patent texts
	Method F	BERT-based content accuracy only	Using only the content accuracy metric, computing semantic similarity of five elements with the BERT model
	Method G	BM25-based category matching only	Using only the category matching metric, calculating the degree of alignment with implicit demand category keywords via the BM25 algorithm
	Method H	Fusion of content accuracy and category matching	The proposed strategy in this study combining content accuracy and category matching

¹ Data source: Experimental data were obtained from 4318 patents corresponding to explicit technological demands and 1864 patents corresponding to implicit technological demands, all retrieved from the IncoPat patent database.

To ensure objectivity and comparability of the results, all methods were ultimately evaluated and ranked using the unified comprehensive evaluation function ${\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}_j$. Even for methods focusing on a single component, the missing metric was supplemented to generate a complete comprehensive score.

• Comparative experiments for explicit technological demand selection

Based on an explicit technological demand case and its corresponding 4318 candidate patents, comparative experiments were conducted to validate the effectiveness of different selection strategies. Four combinations of methods were designed to systematically analyze the contribution of each component to selection quality. Each method was applied to score and rank the 4318 candidate patents, and the top 10 patents under each method were selected for comparative analysis. The results are presented in Table 13, which shows the distribution of comprehensive scores and ranking outcomes for different methods.

Table 13. Comparative experiment results of the explicit demand selection method¹.

Rank	Method A	Method B	Method C	Method D
1	0.7245	0.8456	0.8123	0.9008
2	0.7128	0.8321	0.8089	0.8974
3	0.7089	0.8287	0.8045	0.8945
4	0.6987	0.8156	0.7998	0.8943
5	0.6945	0.8134	0.7945	0.8936
6	0.6897	0.8098	0.7889	0.8926
7	0.6854	0.8067	0.7845	0.8920
8	0.6823	0.8023	0.7823	0.8920
9	0.6789	0.7989	0.7789	0.8918
10	0.6756	0.7954	0.7756	0.8916
Average Score	0.6951	0.8148	0.7930	0.8940

¹ The comparative results were produced using the four selection strategies (Methods A–D) designed in this study.

The results show that the baseline method (Method A) yields significantly lower comprehensive scores, with an average score of only 0.6951 and a top-ranked patent score of just 0.7245. This indicates that traditional keyword matching methods are inadequate for deep semantic understanding. Method B, which applies BERT-based semantic matching, improved the average score to 0.8148, representing a 17.2% increase over the baseline, confirming the positive impact of deep semantic understanding on selection quality. Method C, focusing on element coverage, achieved an average score of 0.7930, slightly lower than Method B, but its top-scoring patent performed strongly in structural completeness, highlighting the value of element coverage in patent selection.

Method D, which applies the fusion strategy, achieved the best results across all evaluation dimensions, with an average score of 0.8940 and a top score of 0.9008, representing a 28.8% improvement over the baseline. Moreover, the score distribution of the top 10 patents under Method D was concentrated at the high end: the lowest score was 0.8916, significantly higher than the maximum scores of other methods. This demonstrates that the fusion strategy not only improves selection accuracy but also enhances the stability of the results.

2.

Comparative experiments of selection methods for implicit technological demands

Based on the implicit demand case and its corresponding 1864 candidate patents, this section designs comparative experiments to verify the effectiveness of the selection strategies for implicit demands. Four different method combinations were constructed to systematically analyze the contribution of each component to selection quality. Each method was applied to score and rank the 1864 candidate patents, and the top 10 patents under each method were selected for comparative analysis. The experimental results are presented in Table 14, which illustrates the performance characteristics of different methods in the context of implicit demands.

Table 14. Comparative experimental results of selection methods for implicit demands ¹.

Rank	Method E	Method F	Method G	Method H
1	0.5892	0.6834	0.6445	0.7520
2	0.5756	0.6721	0.6323	0.7054
3	0.5689	0.6645	0.6287	0.6888
4	0.5634	0.6589	0.6198	0.6784
5	0.5578	0.6523	0.6145	0.6687
6	0.5534	0.6478	0.6089	0.6630
7	0.5489	0.6434	0.6045	0.6591
8	0.5445	0.6389	0.5998	0.6590
9	0.5398	0.6345	0.5954	0.6590
10	0.5356	0.6298	0.5912	0.6581
Average Score	0.5578	0.6526	0.6140	0.6792

¹ The comparative results were generated using the four selection strategies (Methods E–H) designed in this study.

The experimental results for implicit demand selection exhibit characteristics distinct from those of explicit demands. The baseline method produced consistently low comprehensive scores, with an average score of only 0.5577 and the highest patent score reaching just 0.5892. This reflects the difficulty of achieving effective patent selection when relying solely on limited element information. The result confirms the inherent challenge of supply-demand matching under incomplete demand expressions.

With BERT-based semantic matching (Method F), the average score increased significantly to 0.6526, representing a 17.0% improvement over the baseline, while the highest patent score reached 0.6834. This improvement magnitude is comparable to that observed in the explicit demand scenario, validating the stable performance of BERT’s semantic understanding capability across different demand types. Even with limited element information, the deep semantic model effectively captures the latent associations between demands and patents.

Method G, focusing on category matching, achieved an average score of 0.6140, which is lower than Method F but still represents a 10.1% improvement compared with the baseline.

Method H, which applies the fusion strategy, achieved the best performance under implicit demand scenarios. The average score reached 0.6792, and the highest patent score attained 0.7520, representing a 21.8% improvement over the baseline. The strength of this method lies in its ability to maintain high semantic accuracy while simultaneously expanding retrieval coverage through effective category matching.

5. Discussion

5.1. Theoretical and Methodological Contributions

This research contributes to the theoretical development of patent recommendation and technology transfer by defining enterprise technology demand as a structured and multi-dimensional concept rather than a simple text description. The proposed five-element framework, which includes materials, methods, efficacy, products, and applications, provides a systematic model for analyzing enterprise needs. It clarifies how the completeness of these elements affects the accuracy of supply–demand matching. The distinction between explicit and implicit demands also improves understanding of information asymmetry in technology markets. Explicit demands occur when enterprises can clearly describe their technical requirements, reflecting relatively symmetric information. Implicit demands appear when enterprises recognize technical problems but lack the knowledge to describe solutions, creating information asymmetry. This conceptual distinction deepens theoretical insight into how cognitive and informational differences shape technology acquisition.

Methodologically, the study shows that effective patent recommendation depends on processing strategies that fit the characteristics of the demand. For explicit demands, semantic modeling combined with structured element analysis enables precise alignment between enterprise needs and patent content. For implicit demands, semantic reasoning supported by additional information sources helps reconstruct missing elements and improve interpretation accuracy. These findings show that combining semantic understanding with structured reasoning enhances both the adaptability and transparency of recommendation mechanisms. The integrated framework balances data-driven learning with knowledge-based inference and provides a foundation for improving the reliability of patent recommendation systems.

Although this study is based on Chinese enterprise technology demand data, its core processes—including element extraction, semantic modeling, and supply–demand correlation analysis—are not restricted by language. With appropriate linguistic preprocessing and model adaptation, the proposed framework can be applied to multilingual patent and demand datasets, thereby demonstrating theoretical transferability across different languages and regions. The supply–demand data used in this research were obtained from publicly accessible provincial technology transfer and commercialization platforms. The models and parameter settings employed have been described in detail in the paper to ensure the transparency and reproducibility of the research.

5.2. Enhancing Enterprise Technology Acquisition Through Patent Recommendation

This study also examines how the proposed framework helps enterprises overcome practical challenges in technology acquisition. The findings show that recommendation effectiveness depends on how well the system recognizes and adapts to the structural characteristics of enterprise demands. For enterprises that can clearly express their technical requirements, the framework combines BERT-based semantic matching with element coverage analysis to ensure comprehensive identification of relevant patents across materials, methods, performance goals, product forms, and application contexts. This multi-dimensional approach avoids the common weakness of traditional keyword searches, which often capture only surface-level similarities and overlook patents describing equivalent solutions in different technical language. Validation results indicate that this method enables enterprises to obtain highly relevant patent portfolios with minimal manual filtering, saving time while maintaining selection accuracy.

Importantly, this study adopts a demand-driven research perspective, positioning enterprises as the central actors that initiate the recommendation process. The proposed framework is designed to proactively recommend patents that best align with the explicitly stated or implicitly inferred technological needs of enterprises. Unlike traditional patent retrieval or filtering approaches, the proposed method emphasizes a personalized, intelligent, and adaptive demand-oriented matching mechanism. It actively searches for patents based on the characteristics of enterprise technological needs and recommends the most relevant results to enterprises, thereby supporting informed decision-making in technology acquisition and innovation management.

The framework also provides an effective solution for enterprises with incomplete problem descriptions. Such situations often occur in small and medium-sized enterprises that lack dedicated R&D teams or sufficient technical expertise to specify requirements precisely. By integrating multiple data sources, including industry classifications, historical patent records, and technical capability indicators from recruitment information, the framework compensates for missing demand elements that typically hinder conventional systems. The dual-pathway strategy, which combines semantic understanding with category-level reasoning, enables meaningful patent recommendations even when enterprise needs are only partially defined. This capability broadens the accessibility of technology transfer to organizations that were previously constrained by information barriers.

5.3. Limitations and Future Research Directions

This study focuses on improving patent supply–demand matching by developing a demand-driven recommendation framework. While the results demonstrate promising theoretical and practical value, several aspects warrant further exploration to enhance the framework’s generality and applicability. The current research mainly examines the matching process at the element level, emphasizing how content completeness and semantic association influence recommendation performance. However, patent matching outcomes in practice also depend on broader contextual factors such as enterprise technological capability, absorptive capacity, organizational scale, and R&D intensity. Institutional and market environments may further shape how enterprises interpret and apply recommendation results. Future theoretical models of patent supply–demand matching should therefore consider these contextual influences alongside text-based semantic analysis.

The empirical validation conducted in the lithium battery domain provides a focused yet narrow test case. While the five-element classification framework is theoretically supported by system theory and TRIZ principles, its adaptability to other technology fields requires further investigation. Future work should examine how element definitions and completeness thresholds vary across domains and how these variations affect recommendation outcomes.

Future research can extend this work in three directions. First, validate the proposed framework across diverse technology domains to identify its boundary conditions. Second, develop adaptive mechanisms that adjust algorithmic parameters through user feedback and online learning to improve personalization. Third, integrate additional data sources, such as financial constraints, technology roadmaps, and collaboration networks, to strengthen recommendation precision while maintaining computational efficiency.