INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
Enhancing E-Commerce Recommender System Inputs Using  
Transformer-Based Aspect-Based Sentiment Analysis  
Dr. Nazima Khanam, Dr Karen Robinson  
Westcliff University, California, USA  
DOI : https://doi.org/10.51583/IJLTEMAS.2025.1412000098  
Received: 24 December 2025; Accepted: 31 December 2025; Published: 09 January 2026  
ABSTRACT  
Aspect-based sentiment analysis (ABSA) has become an important analytical technique in e-commerce research  
for understanding how customers perceive specific product features. Conventional sentiment analysis  
approaches typically treat a review as a single unit and assign an overall sentiment label, even though customers  
frequently express differing opinions about multiple product attributes within the same review. Such aggregation  
often obscures feature-level preferences and limits the usefulness of sentiment outputs for personalization and  
decision support. To address this limitation, this study proposes a transformer-based ABSA pipeline that  
integrates KeyBERT for unsupervised aspect extraction with Bidirectional Encoder Representations from  
Transformers (BERT) for contextual sentiment classification at the aspect level. The proposed approach is  
evaluated using a dataset of 10,000 Amazon product reviews obtained from publicly available open-source  
review data and is benchmarked against a widely used lexicon-based sentiment analysis method, Valence Aware  
Dictionary and Sentiment Reasoner (VADER). Model performance is assessed using precision, recall, and F1-  
score to capture both classification accuracy and balance across sentiment classes. Experimental results  
demonstrate that the transformer-based pipeline consistently outperforms the lexicon-based baseline, particularly  
in reviews containing mixed or contrasting sentiments across different product attributes. The findings show that  
contextual embeddings enable more accurate identification of sentiment polarity shifts and nuanced opinion  
expressions that are frequently missed by rule-based methods. Overall, the results indicate that transformer-  
based ABSA provides more reliable and interpretable sentiment representations, making it well suited for  
supporting personalized recommendations, feature-level analysis, and improved customer insight generation in  
e-commerce systems.  
Keywords: Aspect-based sentiment analysis, transformer models, BERT, KeyBERT, e-commerce  
personalization, sentiment analysis  
INTRODUCTION  
Customers of e-commerce platforms encounter numerous products in the catalogs of e- commerce platforms.  
Recommender systems are integrated into e-commerce platforms to streamline this process. Current models  
incorporate structured data to provide recommendations through user ratings, purchase history, browsing history,  
among others. However, these methods do not seem to understand the user’s sentiment regarding the product  
and its specific attributes, as they only capture user’s general preferences [1], [2]. Similarly, these systems often  
recommend the same products to users if they have the same ratings, despite the products differing in attributes  
that users may find important.  
Another data source that can be helpful in addressing the shortcomings of the structured data are customer  
reviews. Reviews are more likely to capture the user’s sentiment through their detailed description of a user’s  
experience. Reviews can also portray user sentiment through the varied attributes discussed in the review, such  
as quality, price, and durability, to name a few. Negative and positive sentiments can also be included in the  
same review. However, analyzing customer reviews not only through sentiment data, but through sentiment  
analysis methods that treat reviews as a singular unit by assigning distinct sentiment labels to the reviews, causes  
www.ijltemas.in  
Page 1095  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
important details to be lost. This could also result in losing mixed sentiments regarding the product, which could  
ultimately correlate with a user’s preference. [3].  
Aspect-based sentiment analysis (ABSA) was conceived with the goal of overcoming the limitations of  
traditional sentiment analysis, which is grossly inadequate as it treats a review in its entirety as a unit of analysis.  
By ascribing a sentiment to a review’s different attributes, it becomes easier to unlock the complexity of customer  
feedback. Past work shows that aspect-level sentiment information is useful for personalization and for  
explaining the transparency of the recommendation system in a number of different application areas that employ  
multi-criteria evaluative techniques [4], [5].  
More recently, the application of transformer-based models to sentiment analysis has been enabled by their state-  
of-the-art performance in a broad range of natural language processing tasks due to their ability to capture and  
represent the contextual meaning of a text at a high level. Bidirectional Encoder Representations from  
Transformers (BERT) is particularly illustrated as a strong candidate for sentiment classification because of its  
bidirectional left and right context representation in building its word vectors [6]. In sentiment analysis, as in  
many text processing tasks, keyword and phrase location is a quintessential means of aspect extraction.  
Unsupervised approaches such as KeyBERT are practical for large review datasets because they employ  
transformer embeddings to identify important keywords and phrases without prescriptive labeled training data  
[7].  
Even though there are more advanced techniques, lexicon methods, such as Valence Aware Dictionary and  
Sentiment Reasoner (VADER), are still the most common due to their relative simplicity and low computational  
cost, and are thus appealing to practitioners developing large scale solutions [3]. VADER and other lexicon  
methods, on the other hand, handle the context, negation and contradictory views that are common in real-world  
e-commerce reviews, even with the simplicity of the methods. This brings to question whether other more  
complex techniques such as transformer-based ABSA are justifiably more complex and whether they are able to  
improve on the foundational sentiment understanding of the other methods on large datasets.  
This are the issues in sentiment ABSA methodologies that the current study seeks to understand through the  
empirical comparison of ABSA methodologies. To this end, a transformer-based aspect sentiment modeling  
pipeline using 10,000 product reviews on Amazon obtained from the publicly available Amazon open-source  
review dataset is analyzed. This review explains the use of KeyBERT for aspect extraction and BERT for  
sentiment classification. This pipeline's VADER benchmark and benchmarked against VADER on the basis of  
precision, recall and F1 score. The aim of this work is not to build a complete recommender system, but to assess  
the underlying aspect-level sentiment data to provide personalization and interpretability to e-commerce  
systems.  
This work has a few notable contributions. It first evaluates the performance of a transformer-based ABSA model  
on a massive data set of e-commerce reviews. It places second on the comparison of contextual transformer and  
lexicon-based baseline models on aspect-level sentiment. It finally helps determine how suitable transformer-  
based sentiment models would be for supporting personalization in e-commerce.  
RELATED WORK  
The present work is concerned with sentiment analysis in e-commerce applications; aspect-based sentiment  
analysis; and sentiment analysis using transformer models. These three branches of sentiment analysis research  
comprise the interdisciplinary field of sentiment analysis in e-commerce. Therefore, it is necessary to consider  
the work done in these branches before clearly describing the research gap that this work intends to fill.  
Sentiment Analysis in E-Commerce  
In e-commerce, the sentiment expressed by customers in reviews and feedback comments is analyzed to  
understand customer views. Early work in this field analyzed user sentiment at the document or sentence level,  
where an entire review or a single sentence was labeled as positive, negative, or neutral [9]. These techniques  
www.ijltemas.in  
Page 1096  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
have been relevant to tasks such as product ranking, reputation analysis, and customer satisfaction assessment.  
However, document-level sentiment analysis, while effective in revealing overall sentiment trends, provides  
limited insight into user preferences related to specific product features.  
In recent years, some research has focused on combining sentiment scores with collaborative filtering and  
content-based recommendation approaches [10], [11]. These systems perform better than models built on ratings  
alone; however, they still operate using aggregated sentiment scores. As a result, they have difficulty capturing  
the nuanced and complex sentiments expressed in textual reviews. Such approaches often overlook positive or  
negative opinions related to specific product attributes when products share similar overall sentiment scores.  
Aspect-Based Sentiment Analysis  
Aspect-based sentiment analysis extends traditional sentiment analysis methods by evaluating individual  
attributes mentioned in text and associating sentiment with each attribute. This approach has been applied in  
domains such as hospitality, consumer electronics, and online retailing [12], [10]. Due to the finer-grained nature  
of ABSA systems, sentiment associated with each attribute is explicitly separated, making ABSA a strong  
candidate for supporting personalization and decision-making processes.  
Prior research shows that aspect-level sentiment analysis helps identify which specific attributes trigger positive  
or negative sentiment and improves the interpretability of recommendation outputs [11]. In e-commerce settings,  
ABSA techniques have also been used to summarize customer reviews and determine overall positive and  
negative opinions about products, supporting feature-level comparison between competing items [10]. However,  
many early ABSA approaches rely on lexicon-based rules or supervised learning models that require large  
labeled datasets, which limits their scalability.  
Transformer-Based Models for Sentiment Analysis  
Recent developments in deep learning have encouraged the use of transformer-based models for sentiment  
analysis. Bidirectional Encoder Representations from Transformers (BERT)–based models have demonstrated  
strong performance across a range of sentiment analysis tasks due to their ability to capture contextual and  
semantic relationships within text [6]. Transformer-based approaches have been shown to outperform traditional  
machine learning and lexicon-based methods in multiple sentiment classification benchmarks.  
To reduce the dependence on annotated datasets, unsupervised and weakly supervised techniques for aspect  
extraction have also gained attention. Methods such as KeyBERT use transformer embeddings to identify  
important keywords and phrases that represent potential aspects in text [7]. These techniques are particularly  
suitable for large review datasets, as they avoid the need for labor-intensive manual annotation. Despite their  
potential, gaps remain in the application of end-to-end transformer-based ABSA pipelines within large-scale e-  
commerce contexts.  
Research Gap  
Previous studies demonstrate the potential benefits of sentiment analysis and ABSA for e-commerce  
applications. However, several gaps remain in the existing literature. First, many studies rely on limited datasets,  
which restricts their ability to represent real-world scenarios. Second, while individual transformer models have  
shown strong performance, fewer studies evaluate complete ABSA pipelines in comparison with widely used  
lexicon-based methods. Finally, there is limited empirical evidence demonstrating whether the increased  
complexity of transformer-based ABSA systems leads to meaningful improvements over simpler, lexicon-based  
approaches when applied to large-scale datasets.  
www.ijltemas.in  
Page 1097  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
RESEARCH METHODOLOGY  
This study employs an experimental research methodology to examine whether transformer-based aspect-level  
sentiment modeling offers measurable advantages over traditional lexicon-based sentiment analysis in the  
context of e-commerce reviews. The methodological framework is designed to ensure reproducibility,  
scalability, and fair comparison between models with fundamentally different underlying assumptions. By  
combining unsupervised aspect extraction with contextual sentiment classification, the proposed approach seeks  
to address limitations commonly observed in document-level sentiment analysis and to assess whether these  
improvements justify the additional computational complexity.  
The methodology consists of dataset construction and preprocessing, aspect extraction, sentiment classification,  
baseline comparison, and evaluation using both categorical and probabilistic performance measures. Each stage  
of the pipeline is described in detail in the following subsections.  
Research Question and Hypotheses  
The research methodology is guided by a clearly defined research question that reflects the central objective of  
this study:  
RQ:  
How effectively does transformer-based aspect-level sentiment analysis improve sentiment classification  
outcomes compared to lexicon-based methods in the analysis of e-commerce product reviews?  
This research question is motivated by the growing adoption of deep learning–based language models in  
sentiment analysis and the need to evaluate whether such models provide tangible benefits beyond established,  
low-cost lexicon-based techniques. While prior studies have demonstrated the effectiveness of transformer  
architectures in general sentiment classification tasks, fewer studies have rigorously compared complete ABSA  
pipelines against traditional baselines in realistic e-commerce settings.  
Based on this research question, the following hypotheses are formulated:  
Null Hypothesis (H₀):  
There is no statistically significant difference in sentiment classification performance, measured by accuracy,  
precision, recall, and F1-score, between transformer-based aspect-level sentiment analysis and lexicon-based  
sentiment analysis methods when applied to e-commerce product reviews.  
Alternative Hypothesis (H₁):  
Transformer-based aspect-level sentiment analysis achieves significantly higher sentiment classification  
performance, measured by accuracy, precision, recall, and F1-score, than lexicon-based sentiment analysis  
methods when applied to e-commerce product reviews.  
These hypotheses are evaluated through quantitative performance metrics, error pattern analysis, and qualitative  
examination of aspect-level sentiment outputs.  
Dataset Description  
The empirical evaluation was conducted using a dataset of 10,000 Amazon product reviews, sourced from a  
publicly available open-source review corpus. Amazon reviews are widely used in sentiment analysis research  
due to their diversity, scale, and detailed user feedback. Reviews in the dataset span multiple product categories,  
including consumer electronics, apparel, and household goods, thereby providing a heterogeneous testing  
environment.  
www.ijltemas.in  
Page 1098  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
Each review consists of free-text feedback and an associated star rating. Consistent with prior sentiment analysis  
literature, sentiment labels were derived from the rating values to create a three-class sentiment framework [1],  
[2]. Reviews with ratings of 1 or 2 were labeled as negative, reviews with a rating of 3 were labeled as neutral,  
and reviews with ratings of 4 or 5 were labeled as positive. This labeling scheme reflects commonly adopted  
practices in e-commerce sentiment modeling and supports comparison with prior work.  
The selected dataset size balances computational feasibility with sufficient sample diversity to observe  
meaningful performance differences between models.  
Aspect Extraction Using KeyBERT  
Aspect extraction constitutes a critical stage of the proposed methodology, as it determines which product  
attributes are subsequently evaluated for sentiment. In this study, aspect extraction was performed using  
KeyBERT, an unsupervised keyword extraction technique that leverages transformer-based embeddings to  
identify semantically salient terms and phrases [7].  
KeyBERT operates by generating document-level embeddings and comparing them with candidate n-gram  
embeddings to identify terms that best represent the content of a review. This approach enables the extraction of  
meaningful aspects without requiring manually annotated training data. The use of an unsupervised method was  
intentionally selected to improve scalability and to avoid domain-specific labeling constraints that often limit  
supervised ABSA approaches.  
Extracted aspects typically correspond to product attributes frequently discussed by customers, such as quality,  
price, shipping, durability, and customer service. These aspects serve as the basis for subsequent sentiment  
classification.  
Sentiment Classification Using BERT  
Sentiment classification was performed using Bidirectional Encoder Representations from Transformers  
(BERT), a transformer-based language model that has demonstrated state-of-the-art performance across a wide  
range of natural language processing tasks [6]. BERT’s bidirectional attention mechanism allows it to capture  
contextual relationships between words, enabling more accurate interpretation of sentiment expressions that  
depend on surrounding context.  
For each extracted aspect, contextual embeddings were generated by feeding the corresponding review text into  
the BERT model. Sentiment polarity was then assigned at the aspect level, allowing the model to capture  
opposing sentiments associated with different attributes within the same review. This aspect-level classification  
framework addresses a key limitation of document-level sentiment analysis, which assigns a single sentiment  
label to an entire review.  
The ability of BERT to handle negation, concessive clauses, and sentiment shifts is particularly important in e-  
commerce reviews, where users often express both satisfaction and dissatisfaction in a single piece of feedback.  
Baseline Method: VADER  
To evaluate the effectiveness of the transformer-based pipeline, results were benchmarked against Valence  
Aware Dictionary and Sentiment Reasoner (VADER), a widely used lexicon-based sentiment analysis tool.  
VADER assigns sentiment scores based on a predefined dictionary of sentiment-laden words and a set of  
heuristic rules designed to handle punctuation, capitalization, degree modifiers, and negation [8].  
VADER was selected as the baseline due to its popularity in applied sentiment analysis and its low computational  
requirements. Its rule-based design makes it particularly attractive for large-scale applications where  
computational resources are limited. However, the absence of contextual modeling limits VADER’s ability to  
www.ijltemas.in  
Page 1099  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
capture nuanced sentiment expressions, making it an appropriate comparator for evaluating the benefits of  
transformer-based approaches.  
Evaluation Metrics  
Model performance was evaluated using precision, recall, and F1-score under both macro-averaged and  
weighted configurations. Macro-averaged metrics treat each sentiment class equally and are useful for assessing  
performance on minority classes, while weighted metrics account for class distribution and reflect overall  
classification effectiveness.  
In addition to these aggregate metrics, confusion matrices were examined to analyze class-specific error patterns  
and misclassification tendencies. For the transformer-based model, probability-based evaluation metrics were  
also computed, including Kullback–Leibler (KL) divergence and cosine similarity. These measures provide  
insight into how closely predicted probability distributions align with ground truth labels and offer a more  
nuanced understanding of model confidence and calibration.  
Algorithm 1: Transformer-Based Aspect Sentiment Analysis Pipeline  
Input:  
● Review dataset R={r1,r2,…,rn}  
● Pre-trained BERT model  
● KeyBERT model  
● VADER sentiment lexicon  
Output:  
● Aspect–sentiment pairs for each review  
● Performance metrics (Precision, Recall, F1-score)  
Steps:  
1. Data Preparation  
Load the Amazon open-source review dataset and select review text fields.  
Remove empty or extremely short reviews.  
2. Text Preprocessing  
For each review:  
○ Convert text to lowercase  
○ Remove punctuation and non-alphanumeric characters  
○ Remove stop words  
○ Tokenize the cleaned text  
www.ijltemas.in  
Page 1100  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
3. Aspect Extraction (KeyBERT)  
For each preprocessed review:  
○ Apply KeyBERT to extract a set of candidate aspects  
○ Store extracted keywords and key phrases as aspect candidates  
4. Aspect-Level Sentiment Classification (BERT)  
For each extracted aspect in review :  
○ Identify contextual text surrounding  
○ Input the context–aspect pair into the BERT sentiment classifier  
○ Assign a sentiment polarity label (positive, negative, or neutral)  
5. Baseline Sentiment Analysis (VADER)  
For each review:  
○ Apply VADER to compute a document-level sentiment score  
○ Assign sentiment polarity based on VADER thresholds  
6. Performance Evaluation  
○ Compare transformer-based aspect-level sentiment predictions with VADER outputs  
○ Compute precision, recall, and F1-score for both approaches  
7. Result Aggregation  
○ Aggregate aspect–sentiment pairs across all reviews  
○ Analyze performance differences between transformer-based and lexicon-based methods  
End  
Experimental Results and Analysis  
This section presents the experimental evaluation of the proposed transformer ABSA approach and compares its  
performance with the lexicon VADER. The evaluation was conducted using 10,000 Amazon product reviews,  
and the results are analyzed using both quantitative performance metrics and qualitative aspect-level  
interpretations.  
Dataset Characteristics  
The distribution of sentiment classes in the dataset is summarized in Table 1. Among the 10,000 reviews, 4,076  
reviews (40.8%) are labeled as negative, 3,972 reviews (39.7%) as positive, and 1,952 reviews (19.5%) as  
neutral. This distribution indicates that the dataset is relatively balanced between positive and negative sentiment,  
with a smaller but meaningful proportion of neutral reviews.  
www.ijltemas.in  
Page 1101  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
Table 1. Distribution of Reviews by Sentiment Class  
Sentiment Class  
Negative  
Frequency  
4,076  
Percentage  
40.8%  
Positive  
3,972  
39.7%  
Neutral  
1,952  
19.5%  
The same distribution is visually illustrated in Figure 1, which confirms that no single sentiment class dominates  
the dataset. This balance supports the use of macro-averaged evaluation metrics and reduces bias toward any  
one class.  
Figure 1. Sentiment class distribution of the Amazon review dataset (N = 10,000)  
In addition to sentiment balance, the dataset was analyzed for review length variability. Table 2 reports  
descriptive statistics of review length in terms of word count. Reviews range from very short comments to longer,  
detailed narratives, with a mean length of approximately 75 words and a standard deviation of 43 words. This  
variability increases the likelihood that individual reviews contain multiple aspects and mixed sentiments.  
Table 2. Descriptive statistics of review length  
Statistic  
Mean  
Value  
75 words  
43 words  
8 words  
Std. Dev  
Minimum  
www.ijltemas.in  
Page 1102  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
25th %ile  
Median  
39 words  
67 words  
105 words  
198 words  
75th %ile  
Maximum  
Aspect Extraction Results  
Aspect extraction was performed using KeyBERT, and representative examples are presented in Table 3. The  
extracted aspects correspond to semantically meaningful product attributes such as battery, screen, delivery,  
price, and quality. These results demonstrate that the unsupervised extraction approach successfully identifies  
relevant attributes without requiring annotated training data.  
Table 3. Examples of aspect extraction using KeyBERT  
Original Review  
Cleaned Review  
Extracted Aspects  
“The  
screen  
and “screen camera bright battery life [screen,  
battery]  
camera,  
camera is bright but the short”  
battery life  
is  
too  
short”  
“Fair looking dress “fair dress design, long shipping”  
[design, shipping]  
[price, quality]  
design,  
shipping  
long”  
but  
took  
the  
too  
“Price is reasonable “price reasonable, quality excellent”  
and quality is  
excellent”  
The extracted aspects form the basis for aspect-level sentiment classification and enable the model to analyze  
opinions associated with individual product features rather than treating the review as a single sentiment unit.  
Aspect-Level Sentiment Classification  
Aspect-level sentiment outputs generated using the BERT model are illustrated in Table 4. These examples show  
how sentiment polarity is assigned independently to each extracted aspect. In reviews where users express both  
positive and negative opinions, the transformer-based model correctly identifies opposing sentiments associated  
with different attributes.  
www.ijltemas.in  
Page 1103  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
Table 4. Examples of aspect-level sentiment classification using BERT  
Cleaned Review  
Extracted Aspects  
Aspect Sentiments  
Why BERT is Superior  
“screen camera bright [screen,  
battery life short” battery]  
camera, {screen: Positive,  
camera: Positive  
VADER only gives Neutral;  
BERT  
distinguishes  
both  
positive and negative aspects.  
battery: Negative}  
“fair dress design, long [design, shipping]  
{design:  
Positive, Lexicon-based VADER misses  
shipping”  
shipping: Negative}  
the contrast introduced by “but”;  
BERT captures opposing  
sentiments.  
“price reasonable, quality [price, quality]  
excellent”  
{price:  
quality: Positive}  
Positive, VADER interprets “reasonable  
price” weakly; BERT assigns  
positive  
polarity  
to  
both  
attributes.  
“dress beautiful though [dress, stitching]  
{dress:  
Positive, VADER  
is  
swayed  
by  
stitching loose”  
stitching: Negative}  
“beautiful” and ignores stitching  
flaw; BERT captures mixed  
sentiment.  
“necklace shiny but clasp [necklace, clasp]  
{necklace:  
Positive, Both return Neutral overall, but  
broke”  
clasp: Negative}  
only BERT exposes the flaw in  
clasp at aspect-level.  
{service:  
Negative,  
delivery: Positive}  
“customer service rude [service, delivery]  
yet delivery fast”  
VADER focuses on “rude” and  
misses concession in “yet”;  
BERT separates  
These results highlight a key advantage of ABSA: the ability to preserve sentiment granularity. In contrast to  
document-level sentiment analysis, which collapses multiple opinions into a single label, aspect-level  
classification provides a more faithful representation of customer feedback.  
Quantitative Performance Comparison  
The quantitative performance of the transformer-based ABSA model and the VADER baseline is summarized  
in Table 5, which reports precision, recall, and F1-score using both macro-averaged and weighted metrics.  
www.ijltemas.in  
Page 1104  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
Table 5. Performance comparison of VADER and BERT-based ABSA  
Model  
Precision(Macro)  
Recall  
F1  
Precision  
Recall  
F1  
(Macro)  
(Macro)  
(Weighted)  
(Weighted)  
(Weighted)  
VADER  
BERT  
0.455  
0.570  
0.437  
0.502  
0.394  
0.491  
0.515  
0.638  
0.532  
0.512  
0.471  
0.531  
Across all reported metrics, the transformer-based model outperforms the VADER baseline. The improvement  
is particularly notable in F1-score, indicating a better balance between precision and recall. While VADER  
performs reasonably well in detecting strongly polarized sentiment, its overall performance is lower due to its  
limited ability to handle contextual and mixed sentiment expressions.  
Confusion Matrix Analysis  
To further examine classification behavior, confusion matrices were generated for both models.  
Figure 2 presents the confusion matrix for VADER. The matrix reveals a strong tendency to classify reviews as  
positive, including a substantial number of negative and neutral reviews incorrectly predicted as positive. This  
bias reflects the limitations of lexicon-based approaches when applied to complex, real-world review text.  
Figure 2. Confusion matrix for VADER sentiment classification  
In contrast, Figure 3 shows the confusion matrix for the BERT-based sentiment classifier. The transformer-based  
model demonstrates improved discrimination across all three sentiment classes, particularly for neutral reviews,  
and exhibits a more balanced distribution of classification errors.  
www.ijltemas.in  
Page 1105  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
Figure 3. Confusion matrix for BERT-based sentiment classification  
Distributional Evaluation of Transformer Predictions  
Beyond categorical accuracy metrics, distributional evaluation metrics were computed to assess the quality and  
stability of the BERT model’s probabilistic predictions.  
Figure 4 illustrates the distribution of Kullback–Leibler (KL) divergence values between predicted sentiment  
probability distributions and ground-truth labels. Most values are concentrated at lower divergence levels, with  
a median value of approximately 0.75, indicating reasonable alignment between predictions and true labels.  
Figure 4. KL divergence distribution for BERT predictions  
Figure 5 presents the cosine similarity distribution between predicted and true sentiment vectors. The  
concentration of similarity values close to 1.0 indicates strong directional agreement and suggests that the model  
produces stable and well-calibrated probability outputs.  
www.ijltemas.in  
Page 1106  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
Figure 5. Cosine similarity distribution for BERT predictions  
Summary of Results  
Overall, the experimental results demonstrate that the transformer-based ABSA approach consistently  
outperforms the lexicon-based VADER baseline across quantitative metrics, error analysis, and qualitative  
interpretation. The findings show that contextual modeling and aspect-level sentiment classification provide  
more accurate and informative representations of customer sentiment, particularly in reviews containing multiple  
or conflicting opinions.  
Interpretation and Implications  
This section interprets the experimental findings and discusses their practical relevance, and broader implications  
for sentiment analysis and personalization in e-commerce systems  
Practical Implications  
From a practical standpoint, the results demonstrate that lexicon-based sentiment analysis tools such as VADER  
may oversimplify complex customer feedback. When reviews contain multiple opinions about different product  
features, VADER assigns a single aggregated sentiment label, which can distort representations of customer  
perceptions and obscure actionable insights.  
In contrast, the transformer-based ABSA pipeline disaggregates sentiment at the feature level, enabling more  
precise and actionable interpretations. This capability supports data-driven decision-making in product  
development, targeted marketing, and recommendation personalization. By reducing reliance on coarse-grained  
sentiment scores, organizations can better align product offerings with specific customer preferences and sources  
of dissatisfaction.  
Aspect-Level Utility: Illustrative Use Cases  
The practical utility of aspect-based sentiment analysis is illustrated through representative examples drawn from  
multiple product categories:  
● Electronics (Smartphone):  
Review: “The camera takes stunning photos, but the battery drains within hours.”  
Aspect-Level Output: Camera = Positive; Battery = Negative  
www.ijltemas.in  
Page 1107  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
Implication: The product can be recommended to photography-oriented users, while battery  
performance can be flagged for engineering improvement.  
● Clothing (Dress):  
Review: “The fabric feels premium and elegant, although the stitching came loose after one wash.”  
Aspect-Level Output: Fabric = Positive; Stitching = Negative  
Implication: Retailers can emphasize fabric quality while addressing stitching issues through supplier  
quality control.  
● Jewelry (Watch):  
Review: “The design is gorgeous and looks expensive, but the strap feels cheap.”  
Aspect-Level Output: Design = Positive; Strap = Negative  
Implication: Products can be positioned toward design-conscious customers while improving  
component durability.  
● Appliances (Vacuum Cleaner):  
Review: “It cleans carpets thoroughly, but the noise is unbearable.”  
Aspect-Level Output: Cleaning Performance = Positive; Noise = Negative  
Implication: The product can be marketed to users prioritizing cleaning efficiency over noise sensitivity.  
These examples demonstrate how ABSA enables feature-driven recommendations rather than generalized  
sentiment-based selection.  
Future Implications  
The findings suggest that transformer-based ABSA pipelines can be effectively integrated into recommender  
systems, product feedback loops, and customer support triage processes. Prior studies emphasize that aspect-  
level sentiment analysis improves not only classification accuracy but also interpretability, which is increasingly  
important for the adoption of AI-driven systems in business environments [9].  
The improved performance observed in this study supports broader application of transformer-based sentiment  
models in domains requiring nuanced decision-making and transparent analytical outputs.  
Recommendations for Future Research  
Future research should focus on cross-domain validation of ABSA pipelines, extension to multilingual datasets,  
and exploration of hybrid models that combine lexicon-based features with transformer embeddings.  
Additionally, integrating ABSA-driven insights into live recommender systems and examining human–AI  
collaboration in decision-making contexts represent promising avenues for further investigation.  
CONCLUSION  
This study demonstrates that transformer-based ABSA approaches, specifically the integration of KeyBERT and  
BERT, provide clear performance and interpretability advantages over lexicon-based sentiment analysis  
methods. The findings confirm that the additional model complexity yields meaningful improvements in  
www.ijltemas.in  
Page 1108  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XII, December 2025  
sentiment classification for multi-opinion e-commerce reviews and directly supports feature-level  
personalization strategies.  
REFERENCES  
1.Babaali, M., Fatemi, A., & Nematbakhsh, M. A. (2024). Aspect extraction with enriching word  
representation and post-processing rules. Expert Systems with Applications, 240, Article 120304.  
https://doi.org/10.1016/j.eswa.2024.124174  
2.Bellar, O., Baina, A., & Ballafkih, M. (2024). Sentiment analysis: Predicting product reviews for e-  
commerce recommendations using deep learning and transformers. Mathematics, 12(15), Article 2403.  
https://doi.org/10.3390/math12152403  
3.Cai, H., Xie, Q., Zhao, Q., & Li, K. (2023). Memd-absa: A multi-element multi-domain dataset for  
aspect-based sentiment analysis. Language Resources and Evaluation, 59(3), 2501–2529.  
https://doi.org/10.1007/s10579-025-09820-9  
4.Chauhan, G. S., Nahta, R., & Meena, Y. K. (2023). Aspect-based sentiment analysis using deep learning  
approaches:  
https://doi.org/10.1016/j.cosrev.2023.100576  
2. 5.Cui, Y., Zhou, P., Yu, H., Sun, P., Cao, H., & Yang, P. (2024). ASKAT: Aspect sentiment knowledge  
graph attention network for recommendation. Electronics, 13(1), Article 216.  
https://doi.org/10.3390/electronics13010216  
A
survey.  
Computer  
Science  
Review,  
48,  
Article  
100505.  
3. 6.Darraz, N., Karabila, I., El-Ansari, A., Alami, N., & El Mallahi, M. (2025). Integrated sentiment  
analysis with BERT for enhanced hybrid recommendation systems. Expert Systems with Applications,  
261, Article 125533. https://doi.org/10.1016/j.eswa.2024.125533  
4. 7.Davoodi, L., Mezei, J., & Heikkilä, M. (2025). Aspect-based sentiment classification of user reviews  
to understand customer satisfaction of e-commerce platforms. Electronic Commerce Research, 1–43.  
https://doi.org/10.1007/s10660-025-09948-4  
5. 8.Dogra, V. (2024, July). Aspect-based approaches for measuring customer feedback in the e-commerce  
industry. In Proceedings of the 2024 2nd International Conference on Sustainable Computing and Smart  
Systems (ICSCSS) (pp. 479–484). IEEE. https://doi.org/10.1109/ICSCSS60660.2024.10625381  
6. 9.Elzeheiry, S., Gab-Allah, W. A., & Mekky, N. (2023). Sentiment analysis for e-commerce product  
reviews:  
Current  
trends  
and  
future  
directions.  
Preprints.  
https://doi.org/10.20944/preprints202305.1649.v1  
7. 10.Haq, B., Daudpota, S. M., Imran, A. S., Kastrati, Z., & Noor, W. (2023). A semi-supervised approach  
for aspect category detection and aspect term extraction from opinionated text. Computers, Materials &  
Continua, 77(1), 115–137. https://doi.org/10.32604/cmc.2023.040638  
8. 11.Xu, Y., & Ibrahim, N. F. (2022). Cross-domain aspect-based sentiment analysis for enhancing  
customer experience in electronic commerce. Advances in Artificial Intelligence and Machine Learning,  
2(1), 112–127. https://doi.org/10.54364/aaiml.2024.43151  
9. 12.Zhao, Z., Fan, W., Li, J., Liu, Y., Mei, X., Wang, Y., & Li, Q. (2024). Recommender systems in the  
era of large language models. IEEE Transactions on Data Engineering, 36, 6889-6907.  
https://doi.org/10.1109/tkde.2024.3392335  
www.ijltemas.in  
Page 1109