Named Entity Recognition with Feature-Enhanced BiLSTM and CRF for Fine-Grained Aspect Identification in Large-Scale Textual Reviews

Abstract

Named Entity Recognition (NER) plays a crucial role in Aspect-Based Sentiment Identification (ABSI), enabling the extraction of domain-specific aspects and their associated sentiment expressions from unstructured textual reviews. In complex domains such as movie reviews, sentiment is frequently conveyed through references to named entities (e.g., actors, directors, or movie titles) and other contextual cues. However, many existing ABSI approaches treat NER as a separate preprocessing step, limiting the effective modeling of entity–aspect–opinion relationships. Integrating NER directly into the ABSI framework, allows entity-specific opinions to be more accurately identified, overlapping aspects to be disambiguated, and contextual sentiment expressions to be captured more effectively. To address these challenges, this study proposes an integrated NER-based aspect identification model built on feature-enhanced LSTM and BiLSTM architectures. Linguistic features, including Parts-of-Speech (POS) tags and chunking information, are incorporated to enrich contextual representations, while a Conditional Random Field (CRF) decoding layer models inter-label dependencies for coherent sequence-level predictions of named entities, aspects, and associated opinion expressions. Compared with large transformer-based models, the proposed BiLSTM-CRF architecture offers lower computational complexity, fewer parameters, and allows explicit integration and analysis of linguistic features that are often implicitly encoded in transformer attention mechanisms. The model is evaluated through multiple experimental variants across three domains. Four configurations are applied to movie-review data to jointly extract person names, movie titles, and aspect-opinion pairs, while six configurations assess cross-domain robustness on restaurant and laptop review datasets. Results show that the BiLSTM-CRF model augmented with POS features consistently outperforms baseline configurations in the movie domain and remains competitive across domains, achieving an F1-score of 0.89. These findings demonstrate that explicit linguistic feature integration within a CRF-based sequence modeling can provide an effective and computationally efficient alternative to large-scale transformer fine-tuning for structured, entity-linked ABSI tasks.

1. Introduction

In today’s digital era, organizations, including businesses, government agencies, and social institutions, increasingly rely on public surveys and user-generated reviews to understand public sentiment toward their products and services. However, manually analyzing large volumes of unstructured comments is both complex and time-consuming. To overcome this challenge, sentiment analysis techniques have been introduced as automated solutions for extracting meaningful insights from textual data [1,2]. These insights enable decision-makers to evaluate performance, identify weaknesses, and make timely, evidence-based decisions.

Although traditional sentiment analysis focuses on determining the overall polarity of a review (positive, negative, or neutral) [3,4], such document-level analysis is insufficient for fine-grained decision support. In real-world scenarios, opinions are often directed toward specific aspects or entities within a product or service. For example, a movie review may praise an actor’s performance while criticizing the storyline, or a restaurant review may commend service quality but criticize pricing. Therefore, effective ABSI requires not only detecting sentiment but also accurately identifying the entities and aspects to which that sentiment refers.

In many domains, particularly the movie domain, opinions are closely tied to domain-dependent named entities such as actors, directors, movie titles, brands, or organizations. NER thus plays a critical role in transforming unstructured reviews into structured, analyzable information. Accurate NER enables the system to distinguish entity-specific opinions, disambiguate overlapping aspects, and improve the granularity of sentiment assignment. However, many existing ABSI approaches either overlook NER entirely or treat it as an independent preprocessing step. Such approaches fail to fully capture the interdependencies among entities, aspects, and sentiment expressions, thereby limiting performance in complex domains.

While numerous ABSI methods have been proposed for product [5,6,7] and service [8,9] domain, these datasets typically contain short and relatively straightforward reviews. In contrast, movie reviews are longer, more descriptive, and semantically richer, often involving multiple entities and nuanced contextual cues. Existing movie-domain studies have largely focused on document-level or aspect-level sentiment classification [10,11,12,13], primarily identifying generic aspects such as “plot” or “scene,” while neglecting entity-aware modeling. As a result, opinions directed toward specific persons or titles are not effectively captured.

To address these limitations, this research proposes an NER-based movie aspects identification model utilizing variants of LSTM and BiLSTM, with the aim of developing a feature-enhanced deep learning framework capable of jointly identifying domain-specific aspects, named entities, and associated opinion expressions. The term “variant” refers to the input of LSTM or BiLSTM models with additional features, where the output layer is modified accordingly. For example, one variant incorporates a concatenation of the POS (Part of Speech) layer and word layer, while another variant adds a CRF layer as the output. Four variants of LSTM and BiLSTM are applied to movie reviews, while six variants are applied to restaurant and laptop reviews, with the additional two variants incorporating a chunker layer to further enhance entity detection. The following are the main contributions of this research:

• Integrated NER–ABSI Framework: Unlike traditional approaches, the proposed model integrates NER directly into the aspect identification process, enabling joint modeling of named entities, domain-specific aspects, and associated opinion expressions.
• Feature-Enhanced Deep Learning Architecture: The study systematically investigates LSTM and BiLSTM variants enriched with linguistic features (POS and chunking) and evaluates the impact of CRF-based sequence decoding on entity prediction performance.
• Entity-Aware Movie Review Analysis: The model addresses the underexplored challenge of NER in the movie domain by identifying person names, movie titles, and movie-specific aspects (e.g., plot, scene) along with their associated opinion words.
• Cross-Domain Validation: To assess robustness and generalization capability, the model is evaluated across three heterogeneous domains: movies, restaurants, and laptop products.

The remainder of this paper is organized as follows. Section 2 reviews related work on aspect-based sentiment analysis, named entity recognition, and deep learning approaches for aspect extraction. Section 3 details the proposed methodology, including dataset construction, annotation procedures, and the architectures of the LSTM–CRF and BiLSTM–CRF variants enriched with linguistic features. Section 4 describes the experimental setup and model training configurations. Section 5 presents the experimental results, including ablation studies and a detailed error analysis. Section 6 evaluates the model across multiple domains and provides a comparative analysis with state-of-the-art approaches. Finally, Section 7 concludes the paper and outlines directions for future research.

2. Literature Review

Recent research has explored various strategies for improving aspect-based sentiment analysis (ABSA) and NER across multiple domains. For example, Zhong et al. [14] proposed a knowledge-graph-augmented hybrid architecture combining contextual, syntactic, and knowledge-graph features for laptop, restaurant, and Twitter datasets. Although effective in extracting semantic information, the approach relies on general-purpose embeddings and parsers that may introduce irrelevant domain-independent features and lack explicit mechanisms for identifying domain-specific entities. Similarly, Wu et al. [15] introduced the Phrase Dependency Relational Graph Attention Network (PDRGAT), integrating phrase-structure and dependency trees using graph attention mechanisms. Despite its syntactic modeling capability, the model depends on general parsing tools and does not explicitly address sequential entity labeling required for NER-driven aspect extraction.

Wang et al. [16] introduced a dual graph neural network integrating graph attention and graph convolutional mechanisms to model relationships between aspects and opinion expressions. They combine BERT embeddings with BiLSTM representations and graph neural networks (IDGNN + BERT) to capture both syntactic and semantic dependencies. In that approach, syntactic information from dependency trees and semantic relationships from attention-based representations are fused through graph neural networks, demonstrating that pretrained transformers alone may not fully capture structural linguistic information. Instead, they often benefit from additional architectures that explicitly encode syntactic knowledge. Although effective in aspect identification in restaurant and laptop review datasets, the framework was not designed to handle named entities specific to the movie-review domain.

Likewise, Jian et al. [17] reviewed the evolution of ABSA techniques, highlighting the growing effectiveness of deep learning approaches and proposing a contrastive-learning-based framework that constructs aspect-specific knowledge resources. While the method improved semantic representation learning, its evaluation was limited to general datasets such as restaurant, laptop, and Twitter reviews, leaving domain-specific challenges—particularly entity-centric aspect extraction in movie reviews largely unaddressed.

Deep learning–based sequence models have also been explored. For instance, Bensoltane and Zaki [18] developed a BiLSTM–CRF framework using pretrained embeddings for Arabic ABSA, demonstrating improved performance but still requiring manual annotation procedures that were not fully described. Likewise, Al-Smadi et al. [19] applied BiLSTM–CRF architectures for Arabic aspect extraction, showing the effectiveness of deep learning while highlighting the dependency on large annotated datasets.

Cross-domain NER has also been explored by Jiang et al. [20], who proposed a hybrid Transformer–BiLSTM framework incorporating an adaptive text-sequence enhancement module to enable entity recognition across heterogeneous datasets. While the model demonstrated improved cross-domain adaptability, it showed limitations in identifying domain-specific entities such as movie titles, which are critical for opinion mining in entertainment-related textual reviews.

Oswal et al. [21] combined NER tools, lexicon-based sentiment scoring, and machine-learning classifiers for Twitter data; however, the separation between entity recognition and sentiment analysis limited the ability to capture interdependencies between entities, aspects, and opinions. Likewise, Onalaja, Romero and Yun [12] proposed an aspect extraction method using semantic similarity and handcrafted lexicons, but did not explicitly address named entity identification, reducing the precision of entity-specific opinion mining.

With the emergence of large language models (LLMs), Cheng et al. [22] investigated prompt-based methods for NER and proposed refined prompting strategies, including task-definition prompts, few-shot demonstrations, and output-format prompts, along with a prompt-pruning mechanism to improve prediction quality. Despite these improvements, the study highlighted that LLM-based prompting techniques struggle to generalize across application domains and require careful prompt engineering before deployment in domain-specific tasks such as fine-grained aspect identification. Similarly, Stilwell and Inkpen [23] proposed a prompt-learning-based sentiment analysis approach enhanced with XAI techniques for interpretability. However, the approach remained limited to document-level sentiment classification and did not address aspect-level sentiment analysis or entity-aware opinion extraction.

Domain-specific applications have also been examined in related studies. Musto et al. [24] developed an NLP-driven movie recommendation system that extracts subjective and objective features from structured and unstructured data to support user preference elicitation. Although the system employed NER techniques, the training methodology and dataset specifications were not clearly documented, and the model did not explicitly focus on identifying movie titles as named entities.

The above studies demonstrate that although significant advances have been made in NER-based systems, many studies (i) treat NER as a preprocessing step, (ii) rely heavily on general-purpose linguistic tools, or (iii) lack domain-specific entity modeling. These limitations motivate the proposed research, which integrates NER directly into an end-to-end ABSI framework using feature-enhanced BiLSTM/LSTM architectures with POS, chunking, and CRF decoding layers to jointly model entities, aspects, and sentiment expressions across domains.

3. Proposed NER Model

Figure 1 presents the overall architecture of the proposed NER identification model by addressing key limitations in prior approaches, such as inadequate entity boundary detection and a lack of linguistic feature integration. It mainly consists of two phases: the Annotation Phase and the Training and Identification Phase. The Annotation Phase systematically augments input data with linguistic features, including POS tags and chunking information to enhance contextual representations. The Training and Identification Phase then employs a BiLSTM–CRF hybrid architecture to capture bidirectional contextual dependencies and enforce sequence-level prediction through structured decoding.

Although recent advances in natural language processing have been dominated by transformer-based architectures such as BERT, DeBERTa [17], and hybrid models that integrate transformers with graph neural networks (e.g., IDGNN + BERT) [16]. The selection of a feature-enhanced BiLSTM–CRF architecture over transformer-based alternatives is motivated by three technical considerations.

First, the central contribution of this work is the systematic investigation of how explicit linguistic features (POS tags and chunk boundaries) individually and jointly influence named-entity-driven aspect identification across domains. The BiLSTM–CRF architecture provides an analytically tractable framework in which each feature type can be isolated and evaluated independently. Transformer models encode syntactic and semantic signals implicitly within multi-head attention weights distributed across hundreds of layers, making it extremely difficult to attribute prediction changes to specific linguistic inputs. Our feature-ablation study would not be replicable under a transformer paradigm without a substantial additional interpretability strategy.

Second, the CRF decoding layer explicitly models dependencies between consecutive output labels, enabling globally consistent IOB label sequences. This is particularly important for our nine-class annotation schema (B/I-Person, B/I-Movie, B/I-Feature, B/I-Opinion, O), where invalid transitions, for instance, I-Movie following B-Person must be penalized. Transformer-based sequence labeling models typically apply independent softmax classifiers per token; adding a CRF head on top of a transformer is possible but complicates the fine-tuning pipeline and introduces additional hyperparameters.

Third, the BiLSTM–CRF architecture is substantially more parameter-efficient than large transformer models (BERT-base: ~110 M parameters; DeBERTa-large: ~900 M parameters), enabling the ten experimental configurations reported in this study to be trained and cross-validated on standard academic hardware. This efficiency supports rigorous reproducibility and facilitates systematic cross-domain comparison without requiring specialized computational infrastructure.

This integrated architecture combines linguistic feature engineering, efficient recurrent modeling, and structured sequence decoding, providing a computationally efficient yet effective framework for NER-based aspect identification in sentiment analysis tasks.

During the annotation phase, the dataset is augmented with linguistic features, including word embeddings, POS tags, and chunking information to enrich contextual representations, facilitate precise entity boundary detection, and address limitations in handling noisy data and nested entities in traditional NER models [20,24,25]. In the training and identification phase, these representations are concatenated into a unified input vector and fed into a BiLSTM–CRF hybrid deep learning architecture, which has demonstrated superior performance in sequence labeling tasks by capturing bidirectional contextual dependencies [18] and inter-label predictions via CRF decoding [19]. The model is trained end-to-end on the annotated dataset and applied to input text, predicting single or multiple entity types based on the annotation schema, as illustrated by outputs such as Entity 1, Entity 2, and others in Figure 1. The following sections detail the LSTM- and BiLSTM-based architectural variants explored in this study, along with data annotation strategies.

3.1. LSTM and BiLSTM Variants for the Named Entity Identification

LSTM model, originally proposed by Gers et al. [26], were specifically designed to address the vanishing-gradient problem in recurrent neural networks, enabling the effective learning of long-range dependencies in sequential data [25]. This capability makes LSTM particularly suitable for sequence labeling tasks such as NER or aspect extraction, where contextual information from earlier tokens strongly influences later predictions.

BiLSTM architecture extends this capability by processing sequences in both forward and backward directions, allowing the model to simultaneously capture past and future contextual information. Research by Bensoltane and Zaki [18] demonstrated that bidirectional recurrent models significantly improve sequence modeling performance by incorporating full contextual awareness. For entity extraction tasks, this bidirectional context is especially important because the classification of a token often depends on both preceding and succeeding words.

Subsequent studies, including a sequence-labeling model developed by Jiang, Jin, Chen, Liu, Pang and Yuan [20], showed that BiLSTM models combined with CRF decoding achieve state-of-the-art results in structured prediction tasks by jointly modeling contextual representations and label dependencies. Similarly, work by Al-Smadi, Talafha, Al-Ayyoub and Jararweh [19] demonstrated that BiLSTM-CRF architectures outperform traditional token-level classifiers due to their ability to learn contextual representations while enforcing sequence-level tagging constraints, yielding superior F1-scores in aspect-based sentiment analysis tasks.

Therefore, LSTM and BiLSTM variants are selected in this study because they (i) effectively capture long-range contextual dependencies in textual sequences [19,25], (ii) provide bidirectional contextual understanding essential for accurate entity boundary detection, and (iii) integrate with CRF decoding layers to model inter-label dependencies, resulting in improved entity prediction in aspect identification across domains.

Algorithm 1 illustrates the training and entity identification procedure for the proposed NER framework. The algorithm systematically constructs multiple model variants by combining different linguistic feature embeddings (word, POS, and chunking) with either a Softmax or CRF decoding layer.

Algorithm 1. Training and Entity Identification for the Proposed NER Model.

Input:   $\mathrm{Word} \mathrm{embeddings} W$$, \mathrm{POS} \mathrm{embeddings} P$$, \mathrm{Chunk} \mathrm{embeddings} C$$, \mathrm{Training} \mathrm{dataset} D_{train}$   $\mathrm{Test} \mathrm{dataset} D_{test}$ Output:   $\mathrm{Predicted} \mathrm{named} \mathrm{entities} E=\left\{E_1, E_2, \dots , E_n\right\}$

Step 1: Initialize feature layers $L_{word}\leftarrow W$$, L_{pos}\leftarrow P$$, L_{chunk}\leftarrow C$ Step 2:$\mathrm{Define} \mathrm{model} \mathrm{variants} \mathrm{with} \mathrm{feature} \left(F\right)$$\mathrm{and} \mathrm{output} \left(O\right)$ layers $F_v=\left\{\begin{array}{cc}{L}_{word},\hfill & v=1, 2\hfill \\ \left[L_{word}, L_{pos}\right],\hfill & v=3, 4\hfill \\ \left[L_{word}, L_{pos}, L_{chunk}\right],\hfill & v=5, 6\hfill \end{array}\right.$ $O_v= \left\{\begin{array}{cc}\mathrm{Softmax},& v\in \left\{1, 3, 5\right\}\hspace{1em} \hfill \\ \mathrm{CRF},& v\in \left\{2, 4, 6\right\}\hfill \end{array}\right.$ Step 3: Model Training and Prediction $\mathrm{For} \mathrm{each} \mathrm{variant} v\in \left\{1, 2, \dots , 6\right\}$ Build model ${Model}_v=BuildModel (F_v, LSTM/BiLSTM,O_v)$ Train model ${Model}_v\leftarrow Train({Model}_v, D_{train})$ Predict labels ${Pred}_v\leftarrow {Model}_v. predict\left(D_{test}\right)$ $\mathrm{Extract} \mathrm{named} \mathrm{entities} \mathrm{from} \mathrm{predicted} \mathrm{sequences} {Pred}_i$ Step 4: Output $\mathrm{Return} \mathrm{identified} \mathrm{entities} E=\left\{E_1, E_2, \dots , E_n\right\}$

First, the single-layer variant (V₁) establishes baseline performance using only word-level embeddings with a SoftMax output layer. This allows the model to capture semantic information from the text while providing a reference point for measuring the effectiveness of additional linguistic features.

Second, multi-layer variants (V_3–5) progressively incorporate additional contextual features such as POS tags and chunking information. These linguistic features enrich contextual representation, enabling the model to better capture syntactic structure, boundary information, and contextual dependencies that are particularly important for distinguishing entities from surrounding tokens in domain-specific review texts.

Third, the CRF-based variants (V_{2, 4, 6}) are included to model inter-label dependencies at the sequence level. Unlike independent token classification using softmax, the CRF layer ensures that predicted label sequences follow valid structural patterns (e.g., IOB tagging constraints), thereby improving consistency and reducing boundary errors in entity extraction.

Finally, V₆ combines multiple feature layers with CRF decoding, representing the most feature-enriched and structurally aware configuration. Comparing all six variants enables a comprehensive analysis to determine the impact of adding linguistic features, the effectiveness of CRF-based decoding, and the optimal architecture for accurate and robust entity identification across domains.

Figure 2 illustrates the six-layer architecture of the LSTM-based model. For the LSTM variants, V₁ serves as the baseline, using only word embeddings, a unidirectional LSTM layer, and a Softmax output layer, while V₂ extends this configuration by adding a CRF output layer. Variant V₃ incorporates POS embeddings into the input, and V₄ further integrates the CRF layer for sequence-level decoding. V₅ expands the input representation by concatenating word, POS, and chunking features without CRF, whereas V₆ represents the most comprehensive configuration, combining all input features with CRF-based sequence decoding.

Similarly, Figure 3 depicts the seven-layer architecture of the BiLSTM model. The BiLSTM variants follow the same feature-inclusion strategy as the LSTM models but introduce a bidirectional LSTM layer to capture both forward and backward contextual dependencies. Consequently, each BiLSTM variant contains one additional layer compared to its corresponding LSTM variant, with V₆ incorporating all input features along with CRF decoding for optimal sequence labeling performance.

Variants V₁–V₄ are evaluated across Movie, Restaurant, and Laptop review datasets, whereas the more feature-intensive configurations (V₅–V₆) are applied to the Restaurant and Laptop datasets, where additional linguistic annotations (POS and chunking) are available to enhance model performance.

3.2. Automated Dataset Annotation Procedure

To apply the proposed NER model to the selected dataset, the first step is the annotation process, which involves labeling entities to create a training set. For the movie review dataset, we developed an automated annotation pipeline using rule-based entity labeling from our previous work, Mir and Mahmood [27], which demonstrated higher entity extraction accuracy compared to baseline methods.

During the automated annotation process, unstructured text is first tokenized into a sequence of tokens. Each token is then assigned its corresponding POS tag to capture syntactic structure. Using these tokens and POS tags, word chunks are defined to capture syntactic structures that help in disambiguating entities. To further improve the identification of named entities, orthographic features, such as word casing, are also added.

Algorithm 2 illustrates an annotation pipeline with explicit rules conditioned on POS tags, chunk labels, and orthographic features. The $\mathrm{makeNLP}\left(r_i\right)$ method performs tokenization, POS tagging, chunking, and extraction of word-case features. Its output consists of four feature lists: ($featur{e}_1, \dots , featur{e}_4$), representing tokens, POS tags, chunks, and word-case information.

Once the features are extracted, named entities (NEs) are labeled, including Person NE, Movie Title NE, Feature NE, and Opinion NE, following the IOB tagging scheme widely used in NER for handling multi-token entities. This results in nine distinct labels: B-Person, I-Person, B-Movie, I-Movie, B-Feature, I-Feature, B-Opinion, I-Opinion, and O (other). The annotate () method at line 9 of Algorithm 2 performs this labeling, marking candidate aspects that are identified as relevant entities through rule-based validation, ensuring alignment with ABSA objectives.

Algorithm 2. Automated Annotation Procedure for Movie Review Corpus

Input:   $\mathrm{Movie} \mathrm{review} \mathrm{corpus} R=\left\{R_1, R_2, \dots , R_n\right\}$ Output:   Annotated aspects for each review

1: $\mathrm{For} \mathrm{each} \mathrm{review} r_i\in R$ Do 2:  Extract NLP features: 3:  $\left[{feature}_1, {feature}_2, {feature}_3, {feature}_4\right]\leftarrow Make NLP( r_i)$ 4:  $\mathrm{For} \mathrm{each} \mathrm{token} j$$\mathrm{in} r_i$ do 5:   $\mathrm{if} {feature}_1^j, \dots , {feature}_4^j$ is True 6:   $\mathrm{if} {feature}_1^{j+1}$$, \dots , {feature}_4^{j+1}$ is True then 7:     $candidate\text{\_}aspect$ $\leftarrow$ ${concat(feature}_1^j, {feature}_1^{j+1})$ 8:     $\mathrm{if} candidate\text{\_}aspect$ is relevant then 9:      $Annotate \left(candidate\text{\_}aspect\right)$ 10:      End if 11:     End if 12:    End if 13:  End For

To label these named entities, conditional linguistic rules leveraging POS tags, chunk labels, and orthographic features are employed, as detailed in our prior annotation scheme Mir and Mahmood [27]. For instance, if a token’s POS tag is NN, its chunk label is B-NP, and the word is uppercase, it represents a movie aspect. Similar patterns of POS tags, chunking, and orthographic features are used to identify other entities such as Person, Movie Title, movie-specific aspects, and opinion words. While these rules effectively label entities in the movie corpus, they may also capture irrelevant tokens. To filter out such noise, two pruning methods are applied: imdbpy (https://pypi.org/project/IMDbPY/ accessed on 5 March 2025) and spaCy (https://spacy.io/ accessed on 5 March 2025). IMDbPy removes irrelevant person names and movie titles, while spaCy filters out unrelated movie-specific aspects, such as “actor,” “director,” “script,” and “scene.” Together, these methods enhance the precision of the movie corpus annotation. An example of an unstructured movie review is shown in Figure 4, with its corresponding annotation presented in

Table 1. An excerpt of a movie review after annotation.

Word	POS	Chunk	Case	IOB
Story	NNP	B-NP	TC	O
of	IN	B-PP	LC	O
a	DT	B-NP	LC	O
man	NN	I-NP	LC	O
who	WP	B-NP	LC	O
has	VBZ	B-VP	LC	B-Opinion
unnatural	JJ	B-NP	LC	B-Opinion
feelings	NNS	I-NP	LC	O
for	IN	B-PP	LC	O
a	DT	B-NP	LC	O
pig	NN	I-NP	LC	O
.	.	O	LC	O
Starts	VB	B-VP	TC	O
out	RP	B-PRT	LC	O
with	IN	B-PP	LC	O
a	DT	B-NP	LC	O
opening	NN	I-NP	LC	O
scene	NN	I-NP	LC	B-Feature
that	WDT	B-NP	LC	O
is	VBZ	B-VP	LC	B-Opinion
a	DT	B-NP	LC	O
terrific	JJ	I-NP	LC	B-Opinion
example	NN	I-NP	LC	I-Opinion
of	IN	B-PP	LC	O
absurd	JJ	B-NP	LC	B-Opinion
comedy	NN	I-NP	LC	I-Opinion
.	.	O	LC	O

.

A non-trivial ambiguity arises for movie titles composed of common words and pronouns (e.g., Up, Her, It, Us). The proposed BiLSTM-CRF model addresses this challenge through the following mechanism. First, the BiLSTM captures both preceding and following context for each token. For a sentence such as ‘Her was visually striking and emotionally resonant’, the subject position of ‘Her’ followed by a verb (‘was’) and predicate adjectives provides strong contextual evidence that the token functions as a movie title rather than a possessive pronoun. Second, the CRF layer learns that B-Movie labels in subject position are more probable when followed by tokens that receive opinion or feature labels, forming the entity–aspect–opinion chains characteristic of review sentences. This global sequence constraint reduces the likelihood of labeling ‘Her’ as O when the surrounding tokens carry opinion polarity. Third, the titles used as grammatical subjects in review sentences typically receive a POS tag of NN or NNP, which differs from the PRP (possessive pronoun) or RB (adverb) tags assigned to the same in a non-title context. While POS tagging of such tokens is itself ambiguous, the feature still provides a probabilistic signal that the CRF can weigh.

3.3. Human Annotation and Quality Validation

The annotation of unstructured text is a challenging task, requiring careful attention and consistency. Ambiguities can arise because the same entity may be interpreted differently by different annotators. To quantify annotation quality, a stratified random sample of 300 reviews was independently annotated by 15 annotators. By combining automated annotation using linguistic rules with human validation, we ensure that the movie corpus is accurately labeled. This dual approach strengthens the authenticity of the dataset and provides a reliable ground truth for evaluating the performance of the proposed model.

Previous studies have demonstrated the importance of human validation. For instance, Pontiki et al. [28] annotated datasets for Laptops, Restaurants, and Hotels using three annotators, resolving disagreements through a third annotator, and employed the BRAT web-based annotation tool. Similarly, Al-Smadi et al. [29] conducted human annotation for Arabic book reviews, organizing annotators into seven groups of three postgraduate students each. Groups reviewed peer annotations, and final checks were performed by the course instructor using BRAT. In contrast, Zahra et al. [30] supplemented self-annotation with crowdsourcing via the Figure Eight platform to annotate additional samples and validate the self-labeled data.

Following previous approaches, we adopted a human annotation approach to validate the automated annotation process. Fifteen undergraduate students, proficient in English and familiar with English movies, were recruited, with each annotating 20 reviews, resulting in a total of 300 annotated reviews for validation. Unlike shorter Twitter or product reviews, movie reviews are longer and contain more complex sentence structures, requiring more time per review. Therefore, 300 reviews were randomly selected for the validation process to maintain annotation quality while ensuring a manageable workload for the annotators. To ensure reliability, the annotation process was supervised by an instructor who resolved conflicts when necessary. Fleiss’ kappa was used to measure the level of agreement among annotators, calculated as:

$$k={\displaystyle \frac{P_1-P_0}{1-P_0}}$$

(1)

where $P_1$ is the observed agreement and $P_0$ is the expected agreement by chance.

Table 2. Calculation of Fleiss’ kappa for 10 randomly selected reviews annotated by 15 annotators.

Reviews	Feature	Person	Movie Title	Opinion		Sum	Sum of Square	Sum of Square/15 × 14
1	15	0	0	0		15	210	1
2	0	15	0	0		15	210	1
3	1	5	5	4		15	52	0.25
4	3	2	10	0		15	98	0.47
5	0	0	0	15		15	210	1
6	1	1	1	12		15	132	0.63
7	0	0	15	0		15	210	1
8	3	0	10	2		15	98	0.47
9	1	1	12	1		15	132	0.63
10	2	2	1	10		15	94	0.45
							$Sum$	6.89
							${\displaystyle \frac{Sum}{10}}$	0.689
$Sum$	26	26	54	44
${\displaystyle \frac{sum}{150}}$	0.173	0.173	0.36	0.29
${\left({\displaystyle \frac{sum}{150}}\right)}^2$	0.03	0.03	0.13	0.086	0.276

presents the inter-annotator agreement calculated from 10 randomly selected reviews annotated by 15 annotators for four categories: Feature (movie-specific aspects), Person (named entities such as actors, directors, etc.), Movie Title, and Opinion. Each row represents one review, and four category columns indicate how many of the 15 annotators assigned that label to the review. The observed agreement $P_1$ = 0.689 indicates that annotators agreed 68.9% of the time on average across the sampled reviews. Considering the expected agreement by chance $P_0$ = 0.27, the resulting Fleiss’ kappa is 0.57, which falls within the moderate agreement range (0.41–0.60).

A kappa value of 0.57 was obtained by substituting $P_1$ and $P_0$ into Equation (1) is sufficiently higher than the agreement expected by chance (0.276). This indicates that the annotator agreement is systematic rather than random, providing a credible validation basis for the automated annotation process used to train the model. Furthermore, six reviews in the sample (Reviews 1, 2, 5, and 7, among others) show perfect agreement, where all 15 annotators unanimously assigned the same category. This represents approximately 60% of the sampled reviews and further supports the reliability of the rule-based annotation approach for identifying dominant entity types.

4. Experimental Settings and Results

4.1. Datasets Description

In this study, we rigorously evaluated the performance of the proposed NER model using three datasets: movie reviews, restaurant reviews, and laptop reviews. The Movie Review Dataset (https://ai.stanford.edu/~amaas/data/sentiment/ accessed on 20 January 2025) contains 50k reviews, divided equally into training and testing, with document-level sentiment annotations. Unlike the previous study [27] trained on a subset of 12,000 reviews, the proposed model uses 5000 reviews augmented with POS tags and chunking features. This enriched representation provides superior contextual information, enabling better performance despite the smaller sample size, as validated by experimental results.

The movie review corpus is preprocessed in several steps. First, tokenization is applied to split sentences into words. Then, label encoding is done using the LabelEncoder method to convert unique tokens, POS tags, chunk tags, and class labels into numerical indices. These sequences are further one-hot encoded for class labels. Since LSTM networks require fixed-length inputs, sequence padding is applied to tokens, POS, chunks, and label sequences. Finally, the encoded corpus is split into 80% for training and 20% for testing.

Table 3. Movie dataset distribution.

Movie Dataset	Total Number of Tokens	Total Entities
Training set	1,192,359	67,175
Test set	238,472	13,435

shows the training and test sets for the movie dataset, with a total of 1,192,359 tokens and 67,175 entities in the training set, and 238,472 tokens and 13,435 entities in the test set.

For restaurant and laptop reviews, we adopted the dataset format from SemEval-2014 Task 4 (https://alt.qcri.org/semeval2014/task4/# accessed on 11 April 2025). The original annotation for restaurant and laptop reviews was accomplished using an XML tagging scheme, as shown in Figure 5. However, in order to use it, XML-based annotation was converted into the IOB tagging scheme. In Figure 6, the transformed form of the reviews can be observed, which previously appeared in Figure 5. In this transformed version, all the tags have been removed, and the sentences have been tokenized and labeled with POS and chunker tags. The appropriate tokens are further labeled with the Feature Named Entity, represented as “B-Feature” (beginning of a feature). The labels in the transformed data consist of three categories: “B-Feature”, “I-Feature” (inside a feature), and “O” (others). Figure 6 displays five columns. The first column represents the token number, followed by a list of words in the second column. The third and fourth columns show the POS and chunker tags, respectively. The last column displays the IOB tags, indicating the named entities’ boundaries and their relationships within the sentences. An annotation scheme for a laptop review before and after annotation is presented in Figure 7 and Figure 8.

Table 4. Restaurant and Laptop dataset distribution.

Dataset Partition	Restaurant	Laptop
Test Set	800	800
Training Set	3041	3045

provides the distribution for the Restaurant and Laptop datasets, used for comparative experiments.

4.2. LSTM and BiLSTM Architectures and Training Configurations

To rigorously evaluate the proposed NER model, we implemented and compared two state-of-the-art sequence labeling architectures: LSTM–CRF and BiLSTM–CRF, as discussed in Section 3.

The LSTM–CRF architecture consists of six layers: three embedding layers (for word, POS, and chunk features), a unidirectional LSTM layer, a dense layer, and a CRF output layer. The embeddings, each of dimension 20, are concatenated to form rich token representations that incorporate multi-level linguistic features, enabling the model to effectively capture contextual information within sequences. The unidirectional LSTM (50 units, return_sequences = True, recurrent dropout = 0.1) processes the input sequence in a forward direction, modeling dependencies from past tokens. The CRF layer on top enforces valid label sequences by penalizing invalid transitions (e.g., an I-Feature tag without a preceding B-Feature tag), ensuring accurate entity boundary detection and improving sequence labeling consistency.

The BiLSTM–CRF variant extends this architecture by replacing the unidirectional LSTM with a bidirectional LSTM, using 50 units per direction. This allows the model to simultaneously capture forward and backward contextual information, which is particularly effective for resolving ambiguities in token labeling (e.g., distinguishing between features and opinion expressions within the same sentence). This bidirectional processing substantially improves NER performance, as demonstrated in prior sequence labeling benchmarks [28].

To ensure a fair comparison, both the LSTM–CRF and BiLSTM–CRF models were trained using identical architectural settings and hyperparameters. The input sequences were padded to a fixed maximum length of 75 tokens, with mask_zero = True applied in the embedding layer to ignore padded positions during training. The vocabulary size used for the embedding layer was 117,337, with an embedding dimension of 20. The LSTM layer contained 50 hidden units, while the BiLSTM configuration used 50 units in each direction, enabling the model to capture contextual information from both past and future tokens.

The models were trained using the RMSprop optimizer with a learning rate of 0.001 and optimized using the CRF negative log-likelihood loss function, which is suitable for sequence labeling tasks. A batch size of 32 was used, and the models were trained for 5 epochs with a validation split of 0.10. To mitigate overfitting and improve generalization, dropout with a rate of 0.1 was applied to the recurrent layers, and early stopping was employed during training.

Performance was evaluated using precision, recall, and F1-score, which are standard metrics for sequence labeling and NER tasks. The detailed layer-wise architecture and training parameters are shown in

Table 5. Layer-wise configuration of LSTM-CRF and BiLSTM-CRF models.

Parameter	Input Layer	Embedding Layer	LSTM Layer	BiLSTM Layer	Dense Layer	CRF Layer
Units of length	75	–	50	50 (per direction)	50	–
Input Dimension	–	117,337	–	–	–	–
Output Dimension	–	20	–	–	–	–
Sequence Length	–	75	–	–	–	–
Mask Zero	–	True	–	–	–	–
Return Sequences	–	–	True	True	–	–
Recurrent Dropout	–	–	0.1	0.1	–	–
Activation Function	–	–	–	–	ReLU	–
Output Units	–	–	–	–	–	8

and

Table 6. Training parameters for LSTM-CRF and BiLSTM-CRF models.

Sr. No.	Parameter	Value
1	Epochs	5
2	Batch size	32
3	Optimizer	RMSprop
4	Loss function	CRF loss function
5	Learning rate	0.001
6	Validation split	0.10
7	Metric	CRF accuracy
8	Verbose	1

.

5. Ablation Study on Movie Reviews Dataset

In this section, we present the experimental results obtained from applying different variants of LSTM and BiLSTM on the movie review dataset.

5.1. Performance of LSTM/BiLSTM Variants

The performance of LSTM variants is presented in

Table 7. Performance of the variants of LSTM for movie reviews (best-performing model shown in bold).

LSTM/Features	Precision	Recall	F1-Score
LSTM (baseline)	0.78	0.80	0.79
LSTM + POS	0.72	0.70	0.71
LSTM + CRF	0.78	0.75	0.76
LSTM + POS + CRF	0.88	0.87	0.87

. Among the LSTM variants, the best performing one incorporates the CRF layer and the POS feature. The second-best-performing variant includes LSTM without CRF and the POS feature. On the other hand, the least performing LSTM variant does not incorporate any input or output layer.

Table 8. Performance of the variants of BiLSTM for movie reviews (best-performing model shown in bold).

BiLSTM/Features	Precision	Recall	F1-Score
BiLSTM (baseline)	0.81	0.83	0.82
BiLSTM + POS	0.86	0.86	0.86
BiLSTM + CRF	0.80	0.79	0.79
BiLSTM + POS + CRF (full)	0.91	0.90	0.91

shows the overall performance of the various variants of BiLSTM. The highest performance is achieved by BiLSTM with both the CRF layer and the POS feature, surpassing the variant of LSTM with the same features by 0.03 precision, 0.03 recall, and 0.04 F1-score. The second-best-performing BiLSTM variant incorporates only the POS feature, outperforming the LSTM variant with only the POS feature by 0.14 precision, 0.16 recall, and 0.1 F1-score. The BiLSTM variant without additional features still outperforms the LSTM variant without features, showing 0.03 higher precision, 0.03 higher recall, and 0.03 higher F1-score. It is evident that LSTM and BiLSTM without the CRF layer and POS feature do not perform as effectively, while these deep learning techniques show higher performance with the inclusion of the CRF layer and POS feature for movie reviews.

5.2. Entity-Wise Performance of BiLSTM and Error Analysis

Table 9. Named entity wise performance of movie reviews using BiLSTM + POS + CRF.

Movie Aspects	Precision	Recall	F1-Score
Feature	0.99	0.98	0.99
Movie	0.49	0.55	0.55
Opinion	0.92	0.90	0.91
Person	0.81	0.83	0.82

shows the entity-wise performance of the BiLSTM + POS + CRF (the best variant of BiLSTM among others). Among the NEs, Feature entities achieved the best performance (F1 = 0.99), followed by Opinion entities (F1 = 0.91) and Person entities (F1 = 0.82). Movie title entities exhibit the lowest performance (F1 = 0.55).

The reason for the lower performance of Movie title NEs is due to the varying lengths of titles, ranging from short two-word titles to long titles exceeding 10 words, which introduces inconsistency in token sequences. Many titles also contain special characters, punctuation, or unconventional formatting, complicating tokenization and sequence labeling. Additionally, movie titles consist of multiple words forming n-grams, making it difficult for the CRF layer to accurately detect entity boundaries, especially in longer or complex titles. In contrast, Feature, Opinion and Person entities are usually shorter and linguistically more predictable, which explains the consistently high precision and recall for these classes.

To further understand the performance of the proposed BiLSTM + POS + CRF model, we conducted an entity-wise error analysis, as illustrated in Figure 9. The figure highlights both the proportion of correct predictions and the distribution of error types, including missed entities, boundary errors, and misclassifications for each named entity type. Feature entities were detected with the highest accuracy, indicating the model’s strength in recognizing well-defined movie aspects. In contrast, Movie title entities showed the lowest correct prediction rate, primarily due to the challenges of multi-word titles, variable n-gram lengths, and special characters, which complicate tokenization and boundary detection for the CRF layer. Opinion and Person entities achieved relatively high accuracy, reflecting reliable extraction of opinion and actor or director names.

The error analysis confirms the effectiveness of incorporating POS features and CRF decoding while highlighting directions for enhancing entity-aware aspect identification in complex textual datasets.

Table 10. Examples of correct and incorrect predictions of the proposed BiLSTM–CRF NER model on the movie review dataset.

Example Movie Review	True Labels	Predicted Labels	Type of Error	Observations
“The visual effects of The Final Battle were stunning.”	Feature: visual effects, Movie: The Final Battle Opinion: stunning	Feature: visual effects, Movie: The Final Battle Opinion: stunning	Correct	Accurate detection of multi-word Feature and single-word Opinion; Movie title correctly labeled.
“Spider-Man: No Way Home’s storyline was weak, but Tom Holland acted brilliantly.”	Movie: Spider-Man: No Way Home’s Feature: storyline Opinion: weak Person: Tom Holland Opinion: brilliant	Movie: Spider-Man: No Way Home’s Feature: storyline Person: Tom Holland Opinion: brilliant	Partially Correct	Long-distance dependency caused [Opinion: weak] missed for the storyline feature; Feature, Movie title, and Person were detected correctly.
“Although the plot was light, the cinematography was impressive.”	Feature: plot, cinematography Opinion: light, impressive	Feature: plot Opinion: light Feature: cinematography Opinion: impressive	Correct	Correct detection of multiple aspects and opinions despite sentence complexity.
“The soundtrack of the movie was light and forgettable.”	Feature: soundtrack, Opinion: light, forgettable	Feature: soundtrack, Opinion: forgettable	Partially Correct	Ambiguous Opinion light misclassified; Feature detected correctly.
“The actor’s performance and dialog delivery were excellent.”	Person: actor Feature: performance, dialog delivery Opinion: excellent	Person: actor Feature: performance Opinion: excellent	Partially Correct	Multi-word Feature dialog delivery missed; Person correctly detected.
“Avengers: Endgame exceeded all expectations with breathtaking action sequences.”	Movie: Avengers: Endgame Feature: action sequences Opinion: breathtaking	Feature: action sequences Opinion: breathtaking	Partially Correct	Movie title missed due to n-gram complexity and special characters; Feature and Opinion were correctly detected

presents selected qualitative examples of correct and partially correct predictions. The analysis shows that the model accurately identifies multi-word aspects and their associated opinion expressions for syntactically predictable entities such as Feature and Opinion. However, errors occur in sentences containing overlapping entities, contextually ambiguous words, or long-distance dependencies between entities and their associated features or opinions. For instance, Movie titles with varying lengths and punctuation are often partially missed, while multi-word Person or Feature entities may be incompletely labeled when appearing in conjunction with multiple sentiment expressions.

5.3. Comparison with the Existing Techniques

Table 11. Comparison with the state-of-the-art technique [29].

	Precision	Recall	F1-Score
MAIM [29]	0.899	0.889	0.894
Cross-domain NER [20]	-	-	0.7312
NER Model (ours)	0.913	0.902	0.908

presents a comparison between the proposed NER model and our previous work, MAIM [27] using the movie review dataset. The results clearly demonstrate that the proposed model outperforms MAIM. Notably, MAIM was trained on a dataset of 50,000 movie reviews, while the proposed model was trained on a smaller dataset of only 5000 movie reviews. Furthermore, MAIM did not utilize any feature engineering, whereas the proposed model incorporates feature engineering layers in BiLSTM. The integration of feature engineering layers significantly enhances the performance of movie named entity recognition in the proposed model. In addition, the proposed model is also compared with the state-of-the-art technique, Cross-domain NER [20], and the proposed model achieved 0.17 more F1-score than the Cross-domain NER.

6. Performance Across Datasets

To comprehensively evaluate the robustness and generalization capability of the proposed NER model, different variants of LSTM and BiLSTM were tested not only on movie reviews but also on restaurant and laptop review datasets. This cross-domain evaluation helps demonstrate how the model performs when applied to heterogeneous text corpora with varying entity types and linguistic characteristics.

6.1. Performance on Restaurant Reviews

To further validate the proposed NER model, variants of LSTM are applied to restaurant reviews. The LSTM variant with POS feature demonstrates 0.01 higher precision, 0.05 higher recall, and 0.03 higher F1-score compared to the LSTM variant without POS feature. This indicates that incorporating the POS feature improves the performance of LSTM.

Table 12. The performance of variants of LSTM on restaurant reviews (best-performing model shown in bold).

LSTM/Features	Precision	Recall	F1-Score
LSTM (baseline)	0.63	0.67	0.65
LSTM + POS	0.64	0.72	0.68
LSTM + CRF	0.49	0.61	0.54
LSTM + CRF + POS	0.31	0.68	0.43
LSTM + CRF+ POS + Chunker (full)	0.61	0.68	0.64
LSTM + POS + Chunker (no CRF)	0.85	0.86	0.85

shows that LSTM with a CRF layer alone performs poorly compared to LSTM without CRF and LSTM with POS features. Likewise, LSTM with a CRF layer and POS feature achieves a recall of 0.68, but its efficiency is compromised by very low precision at 0.31. However, LSTM with CRF layer, POS, and Chunker (full) shows some improvement in accuracy, with 0.30 higher precision and 0.21 higher F1-score than LSTM with CRF and POS features.

The highest performing variant of LSTM incorporates POS and chunker features (no CRF), showing significant improvement with 0.24 higher precision, 0.18 higher recall, and 0.21 higher F1-score compared to the variant of LSTM with CRF layer, POS feature, and Chunker.

The variants of BiLSTM are also applied to restaurant reviews, with similar performance observed between BiLSTM with and without POS features, as shown in

Table 13. The performance of variants of BiLSTM on restaurant reviews (best-performing model shown in bold).

BiLSTM and Features	Precision	Recall	F1-Score
BiLSTM (baseline)	0.70	0.72	0.71
BiLSTM + POS	0.69	0.73	0.71
BILSTM + CRF	0.57	0.68	0.62
BILSTM + CRF + POS	0.68	0.76	0.72
BILSTM + CRF + POS + Chunker	0.83	0.90	0.86
BILSTM + POS + Chunker (no CRF)	0.90	0.89	0.89

. BiLSTM without any feature layer and CRF demonstrates 0.07 higher precision, 0.05 higher recall, and 0.06 higher F1-score than LSTM with the same configuration. Furthermore, BiLSTM with POS feature achieved 0.05 higher precision, 0.01 higher recall, and 0.03 higher F1-score compared to LSTM with POS feature. In contrast, BiLSTM with a CRF layer yields more promising results than LSTM with CRF, achieving 0.08 higher precision, 0.07 higher recall, and 0.08 higher F1-score. Notably, BiLSTM with a CRF layer and POS feature significantly outperforms LSTM with the same features, showing 0.37 higher precision and 0.29 higher F1-score.

Likewise, BiLSTM with CRF, POS, and chunker layers showed even higher performance with 0.22 higher precision, recall, and F1-score compared to LSTM with the same configurations. Conversely, BiLSTM with POS and chunker features produces comparable results, with 0.05 higher precision, 0.03 higher recall, and 0.04 higher F1-score as compared to LSTM with POS and chunker features. Thus, the variant of BiLSTM with POS and chunker features emerges as the highest performing model among all LSTM and BiLSTM variants for the restaurant domain.

6.2. Performance on Laptop Reviews

The proposed NER model is applied to a multi-application domain, including laptop reviews. As indicated in

Table 14. The performance of variants of LSTM on laptop reviews (best-performing model shown in bold).

LSTM and Features	Precision	Recall	F1-score
LSTM (baseline)	0.50	0.53	0.52
LSTM + POS	0.54	0.58	0.56
LSTM + CRF	0.45	0.65	0.53
LSTM + CRF + POS	0.48	0.68	0.56
LSTM + CRF + POS + Chunker (full)	0.43	0.60	0.50
LSTM + POS + Chunker (no CRF)	0.76	0.79	0.77

, LSTM without any CRF and feature layers shows 0.04 lower precision, 0.05 lower recall, and 0.04 lower F1-score compared to LSTM with CRF and POS features. On the other hand, LSTM with CRF and POS layers outperforms LSTM without these features by 0.03 in precision, recall, and F1-score. Interestingly, LSTM without a CRF layer and with POS and chunker performs even better, with 0.33 higher precision, 0.19 higher recall, and 0.27 higher F1-score than LSTM with full features and a CRF output layer.

Table 15. The performance of variants of BiLSTM on laptop reviews.

BiLSTM and Features	Precision	Recall	F1-Score
BiLSTM (baseline)	0.74	0.78	0.76
BiLSTM + POS	0.62	0.67	0.64
BiLSTM + CRF	0.74	0.72	0.73
BiLSTM + CRF + POS	0.55	0.65	0.60
BiLSTM + CRF + POS + Chunker (full)	0.67	0.96	0.79
BiLSTM + POS + Chunker (no CRF)	0.75	0.71	0.73

demonstrates the varying performances of the BiLSTM variants on laptop reviews, revealing that overall, the BiLSTM variants outperform the LSTM variants. Specifically, BiLSTM without any CRF and additional features demonstrates 0.25 higher precision, 0.26 higher recall, and 0.24 higher F1-score compared to LSTM without any CRF layer and additional features. Although LSTM with POS feature showed some improvement, BiLSTM with POS feature still outperforms it with 0.08 higher precision, 0.09 higher recall, and 0.08 higher F1-score. Additionally, BiLSTM with only the CRF layer achieves 0.29 higher precision, 0.07 higher recall, and 0.2 higher F1-score compared to LSTM with the CRF layer. This indicates that BiLSTM with CRF yields stable results. Conversely, LSTM with CRF and POS features demonstrates comparable results but falls behind 0.07 in precision, 0.03 in recall, and 0.04 in F1-score compared to BiLSTM with CRF and POS features. Moreover, BiLSTM with CRF layer, POS, and chunker features significantly outperforms LSTM with the same features, with 0.24 higher precision, 0.36 higher recall, and 0.29 higher F1-score.

However, LSTM without CRF but with POS and chunker layers showed slightly better performance, with 0.01 higher precision, 0.08 higher recall, and 0.04 higher F1-score compared to BiLSTM without CRF layer, but with POS and chunker features. In this case, the variant of LSTM performs better than the variant of BiLSTM.

6.3. Comparison with Existing Techniques on Laptop and Restaurant Reviews

The proposed NER model was compared with several state-of-the-art techniques evaluated on the SemEval-2014 Task 4 (https://alt.qcri.org/semeval2014/task4/ accessed on 11 April 2025), which includes laptop and restaurant reviews. It should be noted that the reported results for the baseline methods are adopted from previously published studies addressing the SemEval ABSA challenge, which typically consists of two subtasks: aspect extraction and aspect sentiment classification. Recent approaches in ABSA often incorporate syntactic and contextual information, such as part-of-speech embeddings, dependency-based embeddings, contextualized embeddings (e.g., BERT or RoBERTa), and self-attention mechanisms to capture relationships between aspect terms and surrounding context. These models are primarily designed to identify aspect terms and determine their associated sentiment polarity.

In contrast, the proposed model focuses on NER built on a BiLSTM–CRF sequence labeling architecture, where the CRF layer explicitly models label dependencies across tokens to enforce valid entity sequences (e.g., B–I tag consistency). This structured decoding mechanism is particularly effective for accurate entity boundary detection, which is a key requirement in NER tasks. Transformer models such as DeBERTa do not inherently incorporate such structured sequence constraints unless an additional sequence-labeling layer (e.g., CRF) is integrated on top of the Transformer encoder.

Therefore, although the baseline results are derived from previously published studies using the same SemEval-2014 Task 4 datasets, the underlying tasks are related but not identical. The comparison is thus intended to provide a performance benchmark and contextual reference, illustrating how the proposed model performs relative to established approaches on widely used datasets rather than representing a strictly task-equivalent evaluation.

Table 16. Comparison of the proposed NER model with state-of-the-art techniques (best-performing model shown in bold).

	Restaurants			Laptops
Accuracy	Precision	Recall	F1-Score	Precision	Recall	F1-Score
MC&SF-ABSA [31]	-	-	0.87	-	-	0.78
EliXa [28]	-	-	0.70	-	-	-
AUEB-ABSA [32]	0.72	0.69	0.77	-	-	-
UATE BiLSTM-CRF [33]	0.83	-	0.63	0.75	-	0.44
CMLA [34]	-	-	0.85	-	-	0.78
DeBERTa-RCL [17]	-	-	0.85			0.80
IDGNN + BERT [16]	-	-	0.81	-	-	0.78
NER Model (Ours)	0.90	0.89	0.89	0.67	0.96	0.79

provides a comprehensive comparison between the proposed model and state-of-the-art techniques using restaurant and laptop reviews.

For restaurant reviews, the proposed model achieves a 0.02 higher F1-score than MC&SF-ABSA [31]. Furthermore, it significantly outperforms EliXa [28] with a 0.19 higher F1-score. Similarly, AUEB-ABSA [32] exhibits 0.18 lower precision, 0.2 lower recall, and 0.12 lower F1-score compared to the proposed model. Additionally, UATE BiLSTM-CRF [33] demonstrates 0.07 lower precision and 0.26 lower F1-score than the proposed model. Similarly, CMLA [34] a 0.04 lower F1-score in comparison to the proposed model. The second last study, DeBERTa-RCL [29], has 0.04 less F1-score and finally, IDGNN + BERT [16] has 0.08 less F1-score than the proposed model.

In the laptop domain, the proposed model exhibits a 0.01 higher F1 score than MC&SF-ABSA [31]. However, UATE BiLSTM-CRF [33] shows 0.08 higher precision but 0.35 lower F1-score compared to the proposed model. The proposed model demonstrates a 0.01 higher F1-score than CMLA [34]. Additionally, DeBERTa-RCL [17] has 0.01 more F1-score than the proposed model. Last but not least, IDGNN + BERT [16] has 0.01 less F1-score. The proposed model, utilizing BiLSTM with feature engineering, demonstrates higher efficiency compared to other variants of BiLSTM and LSTM.

7. Conclusions

This study proposed a feature-enhanced NER-based aspect identification model that jointly detects movie-specific aspects, named entities (e.g., person names and movie titles), and associated opinion expressions using LSTM and BiLSTM variants augmented with POS, chunking features and a CRF decoding layer. Experimental results demonstrate that incorporating linguistic features significantly improves the performance of both LSTM and BiLSTM models, with BiLSTM-based variants achieving the best results. Cross-domain validation on restaurant and laptop review datasets further confirms the robustness and generalization capability of the proposed framework. Comparative evaluations with recent state-of-the-art techniques indicate that the proposed model consistently achieves superior F1-scores across domains, demonstrating its effectiveness for entity-aware aspect identification.

However, several limitations remain. First, the recognition of long n-gram entity expressions and complex multi-word movie titles remains challenging, particularly for entities with unconventional punctuation or highly variable lengths. Second, the current approach relies on linguistic features (POS and chunking), which may not fully capture deeper semantic dependencies present in complex sentences. Third, while cross-domain evaluation demonstrates generalization, the model has not been extensively tested on large-scale datasets beyond the movie, restaurant, and laptop domains.

Future work will focus on two main directions. First, the model can be enhanced to recognize longer n-gram entity expressions and more complex entity structures by integrating contextual embeddings from transformer-based architectures, creating a hybrid BiLSTM-CRF–Transformer framework. This hybrid approach can leverage the rich contextual representations of transformers while retaining the structured sequence modeling benefits of CRF layers. Second, the framework could be extended using multi-task learning approaches to jointly model entities, aspects, and sentiment polarity, enabling a fully integrated end-to-end ABSA system applicable across diverse domains. These extensions are expected to further improve the model’s accuracy, generalization, and practical relevance in real-world applications.