Entity Recognition Method for Fire Safety Standards Based on FT-FLAT

Abstract

The continuous advancement of fire protection technologies has necessitated the development of comprehensive safety standards, leading to an increasingly diversified and specialized regulatory landscape. This has made it difficult for fire protection professionals to quickly and accurately locate the required fire safety standard information. In addition, the lack of effective integration and knowledge organization concerning fire safety standard entities has led to the severe fragmentation of fire safety standard information and the absence of a comprehensive “one map”. To address this challenge, we introduce FT-FLAT, an innovative CNN–Transformer fusion architecture designed specifically for fire safety standard entity extraction. Unlike traditional methods that rely on rules or single-modality deep learning, our approach integrates TextCNN for local feature extraction and combines it with the Flat-Lattice Transformer for global dependency modeling. The key innovations include the following. (1) Relative Position Embedding (RPE) dynamically encodes the positional relationships between spans in fire safety texts, addressing the limitations of absolute positional encoding in hierarchical structures. (2) The Multi-Branch Prediction Head (MBPH) aggregates the outputs of TextCNN and the Transformer using Einstein summation, enhancing the feature learning capabilities and improving the robustness for domain-specific terminology. (3) Experiments conducted on the newly annotated Fire Safety Standard Entity Recognition Dataset (FSSERD) demonstrate state-of-the-art performance (94.24% accuracy, 83.20% precision). This work provides a scalable solution for constructing fire safety knowledge graphs and supports intelligent information retrieval in emergency situations.

1. Introduction

Fire safety industry standards are mandatory technical specifications and standards specifically formulated for the fire protection sector, covering areas such as firefighting equipment, emergency response procedures, and safety management [1,2,3]. The formulation and implementation of these standards are of great significance for ensuring public safety, improving rescue efficiency, standardizing industry practices, and promoting technological innovation and development [4]. They serve as a crucial guarantee of the scientific, standardized, and orderly conduct of fire rescue work [5]. As of the end of August 2024, there were 1061 current fire safety standard specifications in China, including 45 national engineering construction fire safety technical standards, of which 145 were mandatory national standards, 143 were recommended national standards, and three were national standardization guidance technical documents. In addition, there were 116 mandatory industry standards, 69 recommended industry standards, and 540 local standards. Furthermore, the National Fire Standardization Technical Committee (TC113) is responsible for managing 291 national fire standards and 185 industry standards, covering areas such as fireworks fire prevention, firefighting and rescue, fire protection products and quality supervision, and fire protection prerequisites [6], as shown in Figure 1. In summary, in the field of fire safety, there are numerous industry standards, but they lack effective integration and organization [7], resulting in severe information fragmentation [8] and a lack of comprehensive coordination [9]. With the systematic research on “Smart Fire Protection” and the rapid iteration and updating of AI and digital technologies, it is particularly important to establish a highly informatized sharing system and related standards, further rationalize the fire protection standards that have been issued and are being formulated [10], prevent various fire protection standards from overlapping and conflicting with each other, and lay a solid foundation for the long-term development of fire rescue informatization.

Fire safety standards exhibit unique linguistic characteristics, including dense technical terminology (e.g., “gas smoke fire detector”), nested entities (e.g., “GB/T 44481-2024: Fire Nozzle Specifications”), and long-range dependencies across clauses [11]. Traditional NER models struggle to handle these patterns for the following reasons:

• Local–global trade-off: The CNN ignores the context between sentences, while the Transformer performs poorly in terms of short-range morphological features.
• Positional sensitivity: Absolute position encoding cannot represent hierarchical relationships in standard documents.

This paper makes three main contributions.

Firstly, we introduce FT-FLAT (Feature Template TextCNN–Flat-Lattice Transformer), based on a hybrid CNN–Transformer architecture, for identifying fire safety standard entities. To better capture positional relationships and feature patterns, we introduce a Relative Position Embedding (RPE) mechanism. The model employs a Multi-Branch Prediction Head (MBPH) to enhance the multi-scale feature learning and hierarchical representation.

Secondly, this paper introduces a Fire Safety Standards Entity Recognition Dataset (FSSERD), which contains 2723 annotated sentences, 99,593 tokens and 3152 tags. The dataset aggregates key sentences from multiple standardized sources and is manually annotated using the BIO pattern. It covers five entity types: buildings, fire alarm products, fire protection products, fire extinguishing products, and laws/regulations.

Finally, benchmarked against seven NER models on our dataset, FT-FLAT performed exceptionally well, achieving an accuracy rate of 94.24% (0.21% higher than the second best) and a precision rate of 83.20% (an improvement of 1.88%). Cross-validation on the MSRA and Boson datasets further confirmed the model’s enhanced capability in relation to fire safety standard entity recognition and demonstrated its consistent performance advantages. The results demonstrated that the method achieved excellent accuracy in fire safety standard entity recognition and outperformed the baseline method in all the evaluation metrics.

The content of this paper is as follows. Section 2 introduces related work; Section 3 outlines the methodology; Section 4 describes the experiments and results; Section 5 presents the word clouds; and Section 6 draws conclusions.

2. Related Work

Recent advances in deep learning [12] have revolutionized the field of entity recognition, with neural architectures now becoming the dominant paradigm for both general and domain-specific extraction tasks. The non-linear modeling properties of deep neural network models are conducive to learning the digital features of entities, and these networks are capable of handling deep-level semantic combinations [13]. They can automatically learn complex hidden features from input data without the need for complex feature processing or in-depth domain knowledge, thereby reducing the need for human intervention and saving a substantial amount of time and effort. The evolution of deep learning for entity recognition has made several key advances. Early work by Collobert [14] (2019) pioneered the application of CNNs in NER tasks, while Cao [15] subsequently enhanced performance through CNN-CRF architectures. Subsequent innovations addressed specific challenges. Kong and Zhang [16] (2021) combined attention mechanisms, multi-scale convolutional kernels with residual blocks to handle long-range dependencies, and Bi-LSTM to capture temporal features. Parallel developments included Peng’s [17] character feature Bi-LSTM-CRF model optimized for small sample learning, and Li’s [18] adversarial active learning approach for joint entity relationship extraction in the field of network security.

The application of Transformer models in entity recognition tasks has seen significant research progress, with numerous studies demonstrating their advantages over traditional sequence labeling methods. The Transformer primarily achieves this goal through the attention mechanisms, and extensive research has shown that that using a Transformer can improve accuracy and significantly reduce training time. Two key advances have been made in the field of entity recognition based on Transformers: Yan’s TENER model [19] and Li Bo’s Transformer–CRF architecture [20], which have set new benchmarks in this field. The TENER model introduces an attention mechanism with directional and relative position information and abandons the original Transformer self-attention scale factor. The introduction of the scale factor is intended to obtain relatively uniform attention weights. The Transformer–CRF model is the first to combine the Transformer model with a conditional random field for entity recognition. Transformer models excel at capturing long-range dependencies and contextual information [21], while the CRF can help resolve dependencies between labels, thereby improving the accuracy of NER tasks. Extensive research has shown that the attention mechanism of Transformers has become a new opportunity and research direction in entity recognition tasks.

3. Methods

This paper introduces an FT-FLAT model for fire safety standard entity recognition, which integrates two complementary components (as shown in Figure 2). One is the TextCNN module [22], which extracts local text patterns through convolution operations. The other is the Flat-Lattice Transformer [23], which captures hierarchical structures and sequence relationships through graph structure attention mechanisms [24]. This hybrid architecture synergistically combines local feature extraction with global dependency modeling and demonstrates excellent performance in fire safety entity identification tasks. Experimental results confirm that the accuracy is significantly improved compared to baseline methods.

3.1. Feature Template

Feature templates are designed based on selected features, taking into account specific contextual information and the specific position of the information in the data. This paper demonstrates that the FT-FLAT model can fuse contextual information about characters in fire safety standard texts using feature templates. Contextual information refers to the “observation window” of the current character and several characters before and after it [25]. A larger observation window can contain more contextual information from the template, but an excessively large window will reduce the efficiency of the fire safety standard entity recognition model and lead to overfitting. On the other hand, smaller observation windows naturally utilize less information and reduce the recognition efficiency of the model. Therefore, it is crucial to choose a reasonable template observation window size.

This paper designs feature templates based on atomic templates, where the specified features are atomic features. The template is designed as follows. The first number in parentheses indicates its position relative to the current character, and the second number indicates the selected feature column, including features such as words and parts of speech. Based on the features specified in the feature template, a set of feature functions is defined (where y represents the label of the current character, yi-1 represents the next label, i represents the current position, and w represents the current character). Each feature function is assigned a weight. When a feature function is activated, its weight is added to a cumulative value, which is the score for that feature. The higher the score, the higher the score for the corresponding label, resulting in a more accurate prediction.

3.2. TextCNN

In convolutional neural networks (CNNs), local feature detection can be achieved by using multiple filters that vary in terms of the vertical dimensions while maintaining a consistent horizontal dimension [26]. This architectural choice allows each filter to focus on capturing word dependencies within different contextual ranges. Each filter contains a unique convolution kernel, enabling CNNs to independently optimize the parameters of each filter through gradient-based learning mechanisms [27]. Using multiple filter configurations simultaneously helps extract complementary feature representations. This multi-filter paradigm enhances the network’s capacity to recognize subtle patterns at different granularities, ultimately strengthening both its representational richness and its adaptive generalization capabilities.

$$C_i^j=soft\max (W^{(fc)}.z+b^{(fc)})$$

(1)

In the context of a CNN, using S independent convolution kernels on the input data generates S independent feature maps. Each convolution kernel captures unique features from the input data, thereby generating different feature maps. The fusion of these feature maps forms a feature vector, which is mathematically represented as:

$${\mathrm{Z}}^j=[{\stackrel{⌢}{C}}_1^j,{\stackrel{⌢}{C}}_2^j,\dots ,{\stackrel{⌢}{C}}_S^j]\in R^m$$

(2)

3.3. Multi-Branch Prediction Head, MBPH

Our model architecture integrates convolutional neural networks (CNNs) for local feature extraction and a Transformer encoder, enhanced with relative position encoding to capture long-range dependencies and structured span interactions. The key innovations include a Multi-Branch Prediction Head (MBPH) for parallel multi-scale feature fusion and a novel Relative Position Encoding (RPE) scheme.

Let $E\in R^{V\times d_e}$ denote the word vector matrix, where V denotes the vocabulary size and d_e denotes the vector dimension. For the input sequence $S=[w_1,w_2,\dots ,w_n]$, we obtain the vector matrix $X\in R^{n\times d_e}$, where each row x_i denotes the vector of word w_i. We process $X$ using a TextCNN layer with multiple filters in order to capture the n-gram features at different scales. Each filter bank contains f filters. The output is a feature map tensor $C\in R^{n\times f\times \left|k\right|}$, where |k| is the number of filter widths used. This tensor captures local contextual features at different n-gram scales around each word position. To integrate the original word vectors and the multi-scale features extracted by the CNN, we reshape X to $X^{\prime }\in R^{n\times 1\times d_e}$ and $C$ to $C^{\prime }\in R^{n\times \left|k\right|\times f}$. Then, we use the Einstein summation einsum to compute the fusion representation $F\in R^{n\times \left|k\right|\times d_h}$ (where d_h is a chosen hidden dimension), thereby effectively performing weighted combination across feature channels:

$$F_{fusion}=einsum{(}^{\prime }nid,njk->nk{d}_h^{\prime },X^{\prime },C^{\prime })$$

(3)

This operation allows interaction between the semantic information in $X$ and the local multi-scale patterns in $C$, thereby enriching the input representation before encoding.

The embedding vectors are processed in parallel by multiple attention heads in the MHA (Multi-Head Attention) layer. Each attention head performs the following independently: linear projection to generate query (Q), key (K), and value (V) vectors: queries ($Q\in R^{b\ast n\ast l\ast d}$), keys ($K\in R^{b\ast n\ast l\ast d}$), and values ($V\in R^{b\ast n\ast l\ast d}$). The scaled dot product attention (SDPA) calculation is as follows:

$$Q=t^{\prime }{W}^Q,K=t^{\prime }{W}^K,V=t^{\prime }{W}^V$$

(4)

Let $W^Q,W^K,W^V$ denote the trainable weight matrices for the query, key, and value projection, respectively. Given the input embeddings V, the attention mechanism is calculated as follows:

$$Attention(Q,K,V)=Softmax({\displaystyle \frac{Q{k}^T}{\sqrt{d_k}}})V$$

(5)

The outputs from multiple heads are connected together to obtain the final output $MBPH(Q,K,V)$. This process can be calculated using the following formulas:

$$hea{d}_{\mathrm{i}}=Attention(Q{W}_m^Q,Q{W}_m^K,Q{W}_m^V),m\in (0,n]$$

(6)

$$MultiHead(Q,K,V)=Concat(hea{d}_1,\dots ,hea{d}_n)W^o$$

(7)

3.4. Relative Position Embedding

The original Transformer model employs absolute position encoding to capture sequence information. However, this paper introduces a Relative Position Encoding method based on Lattice [28], which differs from the absolute position encoding method [29]. We propose a Relative Position Encoding (RPE) based on the relative distances between the span boundaries. We calculate four key relative distances [30]. For spans s_i (head_[i], tail_[i]) and s_j (head_[j], tail_[j]), the four relative distances are calculated using the following formulas:

$$d_{hh}=|hea{d}_{[i]}-hea{d}_{[j]}|$$

(8)

$$d_{ht}=|hea{d}_{[i]}-tai{l}_{[j]}|$$

(9)

$$d_{th}=|tai{l}_{[i]}-hea{d}_{[j]}|$$

(10)

$$d_{tt}=|tai{l}_{[i]}-tai{l}_{[j]}|$$

(11)

These distances capture the relative positions of the start and end points between spans s_i and s_j. The final relative position of the span pair (s_i,s_j) embedded in R_ij is obtained by applying a learnable non-linear transformation ϕ() to the concatenation of these distances:

$$R_{ij}=ReLU(W_r(P_{d_{ij}^{(hh)}}\oplus P_{d_{ij}^{(ht)}}\oplus P_{d_{ij}^{(th)}}\oplus P_{d_{ij}^{(tt)}}))$$

(12)

where $W_r\in R^{4\times d_k}$ is a learnable weight matrix that projects the connection distance onto the key vector space dimension d_k. $R\in R^{n\times n\times d_k}$ is the complete relative position embedding matrix.

The attention score for the head h is calculated by directly merging the relative position embeddings R_ij into the attention logic:

$${Attention}^{\left(h\right)}(Q^{\left(h\right)},K^{\left(h\right)},V^{\left(h\right)},R)=softmax({\displaystyle \frac{Q^{\left(h\right)}(K^{\left(h\right)}{)}^{\top }+S^{\left(h\right)}}{\sqrt{d_k}}})V^{\left(h\right)}$$

(13)

where $S^{\left(h\right)}\in R^{n\times n}$ is a matrix defined as $S_{ij}^{\left(h\right)}=Q_i^{\left(h\right)}\cdot R_{ij}^{\top }$ (a dot product between the relative position embedding of the i-th query vector and the span between i and j. This explicitly biases the attention score based on the relative spatial relationship between spans s_i and s_j, going beyond simple label distances.

3.5. Layer Normalization, FFN and Linear CRF

We used layer normalization, a feed-forward neural network (FFN) [31], and a linear chain conditional random field (linear CRF) [32] in our design. Layer normalization adjusts the input data to a mean of 0 and a variance of 1, which helps the data fit into the range of non-linear activation functions. This process ensures that the data remains in a non-saturated region, thereby enhancing the training speed and stability and avoiding issues such as gradient vanishing or explosion. Note that the FFN submodule in the encoder adopts the Transformer encoder design. It is also used in the model for linear CRF.

4. Experiment

4.1. Data Pre-Processing

To ensure high-quality data for fire safety standard analysis, we implemented the following pre-processing workflow. 1. Data cleaning: Handling missing or duplicate information in unstructured fire safety standard data to ensure data integrity and consistency. 2. Data integration: Integrating key statements from different standardization sources (

Table 1. The analysis of the source document collection.

Document Type	Number
Chinese National Standards	265
Mandatory Industry Standards	226
Recommended Industry Standards	249
Local Standards	159
Count	899

) to establish a unified dataset, providing comprehensive data sources. For example, “GB 50016-2014 Fire Protection Design Code for Buildings, GB 50140-2005 Design Code for Fire Extinguisher Layout, and GA 588-2012 Technical Specifications for Firefighting Personal Protective Equipment”. 3. Data transformation: Converting data into a text format suitable for analysis, including standardization and normalization operations, to facilitate subsequent processing and modeling. A total of 2723 sentences (99,593 tokens) were extracted from standard clauses, technical specifications, and regulatory provisions. 4. Data reduction: Processing data using data feature templates to reduce the data complexity while retaining key information, thereby improving the data processing efficiency and analysis accuracy. (e.g., “gas smoke fire detector” → “smoke fire detector”). 5. Manual annotation: The dataset follows BIO annotation rules and contains five entity categories (

Table 2. The analysis of the datasets.

	Data Type	Sentences	Tokens	Tags
Datasets	Train	1898	69,715	2208
	Validation	271	9960	275
	Test	542	19,918	669
	Count	2723	99,593	3152

). All the text is in Chinese and includes a large number of technical terms. The annotated entity types include buildings, fire alarm products, fire protection products, fire extinguishing products, and laws and regulations. To ensure the quality of the dataset, all the annotations have undergone rigorous multi-stage validation by domain experts, ultimately forming the Fire Safety Standard Entity Recognition Dataset (FSSERD).

4.2. Experimental Data

The Fire Safety Standard Entity Recognition Dataset (FSSERD) adopts a 7:1:2 ratio to divide the training set (70%), validation set (10%), and test set (20%) to ensure the comprehensiveness of the model development and evaluation. The training set is used for model learning, the validation set is used for hyperparameter tuning and training monitoring, and the test set follows established practices in machine learning research to objectively evaluate the final model performance. The dataset is partitioned as follows (Table 2). Training set: 1898 sentences (69,715 tokens), containing 2208 entity annotations. Validation set: 271 sentences (9960 tokens), containing 275 entity annotations. Test set: 542 sentences (19,918 tokens), containing 669 entity annotations.

The dataset uses a two-column format (word–label pairs), with sentences separated by spaces and line breaks. Entity distribution analysis reveals the following (Table 3):

• Fire extinguishing products: 1398 (39.98%).
• Fire alarm products: 676 (19.34%).
• Fire protection products: 872 (24.94%).
• Buildings: 372 (10.64%).
• Laws and regulations: 178 (5.10%).

Table 3. FSSERD entity distribution.

Tag Name	Count	Ratio
Fire extinguishing products	1398	39.98%
Fire alarm products	676	19.34%
Fire protection products	872	24.94%
Buildings	372	10.64%
Laws and regulations	178	5.10%

Table 3. FSSERD entity distribution.

Tag Name	Count	Ratio
Fire extinguishing products	1398	39.98%
Fire alarm products	676	19.34%
Fire protection products	872	24.94%
Buildings	372	10.64%
Laws and regulations	178	5.10%

4.3. Model Training

The model was implemented in PyTorch 2.3.1 using Adam optimization (learning rate = 0.001) and a cross-entropy loss function. Training was conducted for 50 epochs with a batch size of 64, and batch gradient descent was used for parameter updates. The model performance metrics (loss function and accuracy) were recorded for each epoch and used for the optimization analysis (

Table 4. Training process.

Training Procedure: FT-TCFLTransformer.

For epoch = 1… M do                  For batch = 1… T do                        Initialize Parameter                         TextCNN extracts entity feature vectors                       Word vectors and TextCNN output feature matrices are fed into the Flat-Lattice Transformer self-attention layer                       Flat-Lattice Transformer model passes backwards and automatically learns to extract features                        The CRF layer computes the global likelihood probability of the sequence                        Updating parameters                    End for            End for

).

4.4. Evaluation Metrics

Consistent with previous studies, we employ standard evaluation metrics, including precision (P), recall (R), F1-score, accuracy (Acc), macro-average, and micro-average [33]. For each subtask, we use strict matching criteria:

• Trigger identification: Exact offset and reference matching.
• Trigger classification: Exact offset and type matching.
• Argument identification: Exact offset and correct trigger association matching.
• Argument classification: Exact offset, role, and trigger correspondence.

4.5. Experimental Method

The experiments were conducted on a workstation with the following configurations:

• CPU: Intel Core i7-8700K (3.70GHz)
• RAM: 32GB DDR4
• GPU: NVIDIA GeForce GTX 1080 Ti (11GB GDDR5X)

To evaluate the performance of the fire safety standard entity recognition model based on FT-FLAT, experiments were conducted in this section and compared with different algorithm models. The comparison models included SVM [34], Naive Bayes [35], TextCNN [36], BiLSTM-CRF [37], Transformer-CRF [38], Flat-Lattice Transformer [39], Bert-BiLSTM-CRF [40], and the deep learning model based on FT-FLAT proposed in this paper. All the models were trained and validated on the same dataset. Throughout the experimental process, we focused on exploring the performance of different models in the task of fire safety standard entity recognition, with the aim of providing more effective solutions for this task. The goal was to reveal the advantages and limitations of the FT-FLAT model relative to traditional algorithm models and other deep learning models. During the training and validation phases, all the models were trained in the same environment to ensure the reliability and comparability of the evaluation results.

5. Results

We compared our method with traditional and contemporary baseline models.

Table 5. Benchmark results.

Model Name	ACC	Macro-Averaging			Micro-Averaging
		P	R	F	P	R	F
SVM	77.97	58.35	59.25	58.80	59.92	60.99	60.45
Naive Bayes	75.55	76.61	66.86	71.40	77.06	74.37	75.69
TextCNN	86.36	78.62	58.80	67.28	78.40	85.34	81.72
BiLSTM-CRF	93.28	75.47	58.48	65.90	93.80	90.29	92.01
Transformer-CRF	93.88	80.39	65.00	71.88	96.51	91.92	94.16
Flat-Lattice Transformer	94.01	80.68	65.51	72.31	96.68	93.61	95.12
Bert-BiLSTM-CRF	94.03	81.32	73.06	76.77	96.84	94.16	95.48
FT-TCFLTransformer	94.24	83.20	71.27	76.97	99.06	93.72	96.32

presents the comprehensive test set results, while Figure 3 provides a visual comparison of the model performance. Our initial comparison focuses primarily on traditional machine learning methods based on feature engineering. The experimental results demonstrate that our proposed method performs exceptionally well in fire safety standards entity recognition, achieving an accuracy of 94.24%, a recall rate of 71.27%, and an F1-score of 76.97%. In contrast, traditional methods such as SVM performed relatively poorly, with an accuracy rate of only 77.97%, a recall rate of 59.25%, and an F1-score of 58.80%. These results clearly demonstrate the advantages of deep learning models in the task of fire safety standard entity recognition. This performance gap stems from the fact that traditional methods rely on manual feature engineering to convert text into numbers prior to classification, a process that often fails to capture the semantic nuances of safety standard text. However, this method has limited expressive power for text data and struggles to capture and process complex semantic information, resulting in poor performance in fire safety standard entity recognition tasks. In contrast, the deep learning-based method proposed in this paper can better learn semantic information from text data and automatically learn feature representations through end-to-end learning, thereby improving the recognition accuracy and performance. These results highlight the superiority of the FT-FLAT model in text entity recognition and provide a more effective solution for fire safety standard entity recognition.

Next, our method demonstrates superior performance to existing deep learning models, achieving state-of-the-art accuracy of 94.24% (baseline model: 94.03%) and precision of 83.20% (baseline model: 81.32%), representing improvements of +0.21% and +1.88%, respectively. This advancement stems from three key innovations:

• Hybrid architecture: Combines TextCNN local feature extraction with the Flat-Lattice Transformer’s hierarchical modeling.
• Relative Position Embedding (RPE): Encodes precise positional context to improve entity boundary detection.
• Multi-Branch Prediction Head (MBPH): Enhances feature learning through parallel processing of branches.

These components synergistically capture contextual relationships (e.g., regulatory text dependencies) and intrinsic entity features (e.g., product terminology patterns), which has been validated through our experiments.

Finally, we further analyzed the model’s performance across four fire safety entity categories. Figure 4 displays the test results for the five entities across the different models. The F-value was highest for the fire extinguishing product category, demonstrating TextCNN’s unique advantage in extracting discriminative features for this entity type. The reason for this is that using character vectors as input features can better capture and represent morphological information in the text data, such as specific character combinations, fixed formats, and frequency of occurrence. The results indicate that this method can effectively utilize morphological features in text data, making it easier for the model to recognize entities with fixed formats.

6. Discussion

6.1. Confusion Matrix Analysis

To diagnose the model behavior, we provide the confusion matrix for FT-FLAT (Table 6). The key observations are as follows:

• Major errors: Only 18% of “Fire Protection Products” were misclassified as “Fire Extinguishing Products” (e.g., “fireproof sealant” → “fire hose”).
• Nested entities: Here, 22% of errors occurred in laws/regulations with nested structures (e.g., “GB/T 44481-2024: Fire Hose Specifications”).
• Terminology ambiguity: Only 15% of errors were caused by ambiguous terminology (e.g., “detector” in “alarm system” vs. “detector” in “fire extinguishing system”).

Table 6. Confusion matrix.

Actual/Predicted	Building	Alarm	Protection	Extinguishing	Law
Building	0.92	0.01	0.04	0.03	0.00
Fire Alarm Product	0.00	0.88	0.07	0.05	0.00
Fire Protection Prod	0.02	0.03	0.79	0.18	0.00
Fire Extinguishing	0.01	0.02	0.10	0.87	0.00
Laws/Regulations	0.00	0.00	0.11	0.11	0.78

Table 6. Confusion matrix.

Actual/Predicted	Building	Alarm	Protection	Extinguishing	Law
Building	0.92	0.01	0.04	0.03	0.00
Fire Alarm Product	0.00	0.88	0.07	0.05	0.00
Fire Protection Prod	0.02	0.03	0.79	0.18	0.00
Fire Extinguishing	0.01	0.02	0.10	0.87	0.00
Laws/Regulations	0.00	0.00	0.11	0.11	0.78

6.2. Ablation Study

We analyzed the various components of FT-FLAT and quantified their contributions, with the results shown in Table 7:

• RPE removal: Macro-F1 ↓ 3.21%—confirming its role in the encoding layer hierarchy.
• MBPH removal: Macro-F1 ↓ 2.58%—confirming the necessity of multi-scale fusion.
• TextCNN removal: Precision ↓ 4.15%—highlighting the importance of local pattern extraction.

Table 7. Ablation study.

Variant	Macro-P	Macro-R	Macro-F1	ΔF1
Full FT-FLAT	83.20	71.27	76.97	-
w/o RPE	79.85	68.91	73.76	−3.21
w/o MBPH	80.12	69.04	74.39	−2.58
w/o TextCNN	79.05	70.33	74.42	−2.55
CNN-Only	76.81	65.28	70.62	−6.35

Table 7. Ablation study.

Variant	Macro-P	Macro-R	Macro-F1	ΔF1
Full FT-FLAT	83.20	71.27	76.97	-
w/o RPE	79.85	68.91	73.76	−3.21
w/o MBPH	80.12	69.04	74.39	−2.58
w/o TextCNN	79.05	70.33	74.42	−2.55
CNN-Only	76.81	65.28	70.62	−6.35

6.3. Error Analysis

We categorized 100 random errors from the test set (

Table 8. Error type distribution.

Error Type	Value	Example
Nested Entities	32%	“GB/T 44481-2024 [Law] → Fire hose [Extinguishing]”
Domain Polysemy	28%	“Detector [Alarm] → Fire detector [Protection]”
Long-Range Dependencies	22%	“When smoke density exceeds …”
Annotation Ambiguity	18%	Boundary mismatches (“fireproof glass window” → window vs. glass)

).

The key findings are as follows:

• First, 60% of errors stem from structural complexities (nested/long-range).
• Compared to absolute encoding (e.g., clause references), RPE reduced the positional errors by 41%.
• Error analysis reveals that the remaining 60% of errors are related to nested entities and terminology polysemy, suggesting that future should focus on syntax-aware embedding studies. Ablation studies confirm that RPE and MBPH contribute +3.21% and +2.58% to F1, respectively.

6.4. Word Clouds

This paper trained a fire safety standard entity recognition model to extract fire safety standard entities from text data. Figure 5 shows the word cloud formed by the five categories of fire safety standard entities extracted.

As can be seen from Figure 5, among the five major categories of fire safety standard entities, the “fire extinguishing products” entity accounts for the largest proportion. Examples of fire extinguishing products include gas fire extinguishing controllers, dry powder fire extinguishing controllers, fire sprinklers, fire pumps, and fire hose reels. Following this, “fire protection products” (e.g., gas-based smoke detectors, steel fire-proof windows and fire protection electric devices) are particularly prominent. Subsequently, “fire alarm products” are mentioned, such as fire emergency broadcast equipment, fire sound and light alarms, point-type smoke detectors, etc. There are relatively few entities related to “Buildings” and “Laws and Regulations”, such as factories and GB/T 44481-2024.

7. Conclusions

We propose FT-FLAT, a novel hybrid CNN–Transformer architecture for fire safety standard entity recognition, to address the challenge of information fragmentation in complex regulatory texts. Our main contributions are four-fold. (1) Architectural innovations: We integrate TextCNN for local feature extraction and the Flat-Lattice Transformer for global dependency modeling to synergistically capture short-range terminology patterns and long-range clause dependencies. (2) Technological advances: The proposed Relative Position Embedding (RPE) dynamically encodes the hierarchical relationships between entities, while the Multi-Branch Prediction Head (MBPH) enhances the multi-scale feature fusion through Einstein summation. (3) Resource contribution: We released the Fire Safety Standard Entity Recognition Dataset (FSSERD) containing 2723 annotated sentences covering 3152 entities across five categories. (4) Experimental results demonstrate that FT-FLAT achieves state-of-the-art performance: 94.24% accuracy (+0.21% improvement over best baseline), 83.20% precision (+1.88% improvement), and the highest F1-score (76.97%) on the FSSERD, which is excellent for fire extinguishing products.

Whilst this study demonstrates promising results, there are still some noteworthy limitations. The current dataset, whilst carefully annotated, is still limited in terms of the size (3152 entities) and diversity of fire safety standards. Expanding the corpus to cover international standards and multilingual sources could enhance the generality of the model. While GPT-4 [41] and RWKV [42] represent state-of-the-art general-purpose language models, the domain-specific design of FT-FLAT offers distinct advantages for fire safety standards. (1) Specialized architecture: The hybrid CNN–Transformer structure with RPE explicitly addresses unique challenges such as nested entities (“GB/T 44481-2024: Fire Hose Specifications”) and long-range dependencies across clauses, which may be ignored by general-purpose models. (2) Computational efficiency: FT-FLAT requires significantly fewer resources (trained on a single GTX 1080 Ti) than billion-parameter LLMs, making it suitable for resource-constrained emergency management systems. (3) Interpretability: Feature templates and MBPH provide transparency in capturing domain-specific patterns (e.g., morphological features of “gas smoke fire detector”), whereas black-box LLMs lack this fine-grained control. Future work will benchmark FT-FLAT against these models to quantify the trade-offs between generality and domain precision. These limitations provide concrete ways to advance fire safety NLP research while maintaining the interpretability advantage of our method over black-box alternatives.