Rule-based methods use predefined linguistic rules or patterns to identify aspects in text. Linguistic features often include syntactic patterns, part-of-speech (POS) tags and dependency relations within sentences [17]. Rules may be crafted to detect implicit expressions or contextual cues that indicate the presence of certain aspects [18]. By defining rules based on linguistic features, aspect extraction systems can effectively parse and extract relevant aspects from text data. Rule-based aspect extraction is often implemented in conjunction with other methods, such as machine learning techniques, to improve accuracy and coverage. Alqaryouti et al. [19] used a combined method implementing both a lexicon and rule-based approach to extract both explicit and implicit aspects, along with performing sentiment classification for these aspects. Venugopalan and Gupta [20] utilized a hierarchical rule-based method for aspect term extraction by prioritizing high recall and minimizing false positives. Although rule-based methods can be highly accurate for the datasets they are fine-tuned on, their adaptability is limited because extensive manual efforts are needed to create and update rules for different domains. It is labor-intensive to develop enough rules to categorize the extracted aspect terms due to the variability and ambiguity of natural language.

The syntactic relationships of words in a sentence can be determined automatically with dependency parsing, a procedure which examines the grammatical structure of the sentence. Wang et al. [21] constructed an RNN-based dependency tree, called the Recursive Neural Conditional Random Fields (RNCRF) model, to learn sentence feature representations and then to mark aspect labels by introducing a Conditional Random Field (CRF). Luo et al. [22] presented a model (BiDTressCRF) which is based on a bidirectional dependency tree network and a Bi-LSTM to extract syntactic features as well as contextual features, and to feed the features into a CRF to predict aspect labels. Dai et al. [23] developed the Ruled Incorporated Neural Aspect and Opinion Term Extraction (RINANTE) model, which automatically mines extraction rules based on syntactic information and applies these rules to label aspects. Following this, an LSTM and CFM neural network are trained to predict aspects based on the automatically labeled aspects and the ground truth labels. Liang et al. [24] introduced the Dependency Relation Embedded Graph Convolutional Network (DREGCN), to model the dependency relation of words in sentences using a graph convolutional network and applies an effective message-passing mechanism to enhance model learning for aspect extraction.

Unsupervised techniques do not require labeled data. Instead, they might use clustering algorithms to group similar terms or phrases, assuming each cluster represents a distinct aspect. Topic modeling techniques like Latent Dirichlet Allocation (LDA) are also popular for identifying topics within text, which can correspond to aspects [25, 26]. Probabilistic Latent Semantic Analysis (PLAS) was one of the topic modeling methods, in which two probability distributions, the document-topic and topic-word distributions [27, 28]. The former describes the probability of topics occurring in a document, and the latter represents the probability of words appearing in a given topic [29]. Topic-based approaches may not be able to capture the contextual information of documents and long-term dependency of long sentences [30]. Another challenge with unsupervised methods is that the aspects they extract might not always align with the semantic categories of interest, requiring post-processing or human validation [31].

Supervised methods use labeled data to train models that can recognize and extract aspects. Techniques range from traditional machine learning models like Support Vector Machines (SVM) and Random Forests to neural network architectures, including Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) [32]. Recently, many deep learning models have been developed to capture complex language patterns for aspect extraction. For instance, Alslaity and Orji [33] integrated a set of rule-based techniques with the deep net to enhance the performance of both the aspect extraction and sentiment scoring methods. Wang et al. [34] employed coupled multi-layer attentions (CMLA) to learn the relation tokens in a sentence, which could exploit indirect relations between terms for more precise information extraction. Li et al.[35] introduced the Truncated History-Attention (THA) model, which employs LSTM to generate initial word representations and integrates aspect detection history into the current aspect representation through truncated history attention, creating a history-aware aspect representation for aspect extraction. Xu et al. [36] developed the Double Embeddings CNN (DE-CNN), which concatenates general-purpose embeddings with domain-specific embeddings in a CNN model for aspect extraction. Tran et al. [37] presented the Bi-GRU-CRF model, combining a bidirectional Gated Recurrent Unit (Bi-GRU) with word embedding and CRF into a unified framework for extracting aspects. Luo et al. [38] proposed a dual cross shared RNN framework (DOER) to determine the polarity of aspect terms, incorporating ReGu to improve feature extraction capability. Chen et al. [39] applied BERT layers and direction graph convolutional networks (D-GCN) to implement aspect extraction while encoding syntactic information.

Hybrid methods combine the strengths of rule-based, supervised, and unsupervised approaches to improve aspect extraction performance. A hybrid system might use rule-based filtering to pre-process data, apply supervised learning for initial aspect extraction, and then refine the results using unsupervised clustering. For example, Chauhan et al., [40] used a two-step mixed unsupervised model by combining linguistic patterns with deep learning techniques to improve aspect extraction from unstructured data on social networks. Pereg et al., [41] combined linguistic information with a self-attention mechanism and the BERT model to extract aspects. Xu et al. [42] developed a deep learning method called DomBERT in combination with in-domain corpus and relevant domain corpora for ABSA. Kotagiri et al. [43] applied reptile search optimization, extreme gradian boosting and deep learning methods to the aspect extraction and sentiment analysis across multiple datasets. Venugopalan & Gupta [44] enhanced the LDA model using BERT based semantic filtering, particle swarm optimization strategy and linguistic rules. Liu et al., [8] used a hybrid model graph-based multi-grained convolution (CMGC) with attention mechanism to perform the sentiment analysis on automotive website consumers. Hybrid approaches aim to balance the precision of rule-based and supervised methods with the broad applicability of unsupervised techniques [32].

Figure 2 illustrates the architecture of the proposed Seq2Seq model, designed to extract aspect phrases or words from the input text sequence by generating the corresponding label sequence . Given that aspect phrases often span multiple tokens, the model must effectively capture both the contextual relationships and dependencies between these tokens. Properly labeling multi-token aspects also necessitates maintaining label consistency across dependent tokens, which requires an in-depth understanding of their syntactic structure. To achieve this, the model incorporates a global position-aware attention mechanism, which dynamically weighs the importance of each token based on its semantic and positional relevance [45]. The position-aware alignment score is enhanced by incorporating dependency parsing, which encodes the syntactic relationships between tokens as positional biases [46]. This approach ensures that the model attends to both the most relevant tokens and their syntactic dependencies, addressing inconsistencies that arise in multi-token aspect labeling [47] [48]. Furthermore, aspects often exhibit syntactic dependencies that should be captured by the model. The integration of position and dependency-aware attention enables the framework to generalize across domains, capturing cross-domain relationships while maintaining accuracy in domain-specific contexts [47] [46].

The proposed Seq2Seq model mainly comprises embedding, and multiple bidirectional GRU (Bi-GRU) and attention mechanism layers. In the embedding layer, Google's pre-trained word2vec word embedding transforms a word token, , in a vector that represents a review sentence, , into an embedding vector , where d represents the dimension of the embedding feature. The embedding output that corresponds to is denoted as e =. The Bi-GRU layer connects multiple recurrent units in a series to generate subsequent sequences that enhance the acquisition of a hidden state representation, h =, for all N tokens. The inclusion of the Bi-GRU architecture ensures the model captures both forward and backward dependencies, enhancing the contextual understanding of the sequence [45].

Incorporating both positional and syntactic information, attention mechanisms are applied to generate the global and local context vectors by dynamically focusing on the most relevant parts of the input sequence. The global attention mechanism first computes alignment scores between the encoder and decoder states. These scores are then normalized using the SoftMax function to produce attention weights ​, which are probabilities that reflect the importance of each encoder hidden state ​ at the current time step t. Each is subsequently weighted by its corresponding , generating the global context vector. This vector captures the most relevant semantic information from the input, aligned with the decoder's focus at time step t.

In addition to the global context vector, the model computes a local context vector ​, which emphasizes tokens that are syntactically connected to the current token based on dependency parsing. These syntactic relationships such as modifiers or dependents are derived from a dependency tree of the input sequence. Unlike the global attention mechanism, which considers all encoder hidden state ​, the local attention mechanism focuses only on a subset corresponding to the syntactic children of the current token. Local alignment scores and attention weights are then computed similarly to the global attention mechanism.

Next,​ and ​ are integrated through a context fusion gate. The gate, regulated by a sigmoid function, controls the contribution of each context vector, ensuring that the fused context vector ​ captures both global semantics and local syntactics. is then concatenated with the embedding of the previous output token and fed into the decoder’s GRU, along with the previous decoder hidden state . This concatenated input results in a more contextually aware and coherent sequence labeling. If incorrect token labels are generated, the model can dynamically adjust its focus and predictions during subsequent decoding steps through the feedback. This mechanism enhances the model’s ability to capture long-range dependencies and maintain label consistency across related tokens by leveraging both the syntactically and semantically informed context in conjunction with the decoder’s evolving internal state.

The decoder’s attention mechanism is implemented with the linear and SoftMax layer [49] to determine the predicted label, for the token , as follows:,

where and are weight parameters. The iteration continues until the decoder reaches the predefined maximum sequence length and generates an end-of-sequence (< EOS>) token. Consequently, the model predicts a complete label sequence

To train the model, a set of labeled review sentences, , was used for minimizing the loss function, , which is calculated with cross entropy and L2 regularization as follows [50]:

where w​ is a vector of the trainable weight parameters in the proposed Seq2Seq model, and λ is the hyperparameter of L2 regularization. The Adam optimizer was used to minimize the loss function. Further details regarding this framework are provided in the following sections.

In the proposed model, a serial Bi-GRU structure is employed to capture information from both preceding and subsequent time steps to generate the current state. The serial Bi-GRU layer consists of a forward and backward GRU. The forward GRU, denoted as , processes the input sequence in its original order, producing a sequence of forward hidden states, . Here, represents a non-linear transformation function such as tanh or ReLU. On the other hand, the backward GRU, denoted as , reads the sequence in the reverse order, yielding a sequence of backward hidden states, . For token , its embedding vector is and the hidden state at the preceding time step is . The forward and backward hidden states, and , are computed as follows:,,

where is a weight matrix for the input embeddings, and is a recurrent weight matrix for the hidden state, with m representing the dimension of the hidden state. The final hidden state ​ of token can be obtained by concatenating and , as described by using Eq. (5) [51]:

It is assumed that context words with different relative positions have varying effects on the representation learning of aspect terms [52], and context words located closer to the aspect term may have a greater influence. Inspired by the attention mechanism and position awareness, a global position aware attentional model was proposed by different researchers such as Zhao et al., [47] and Xu et al., [42]. This concept involves incorporating all hidden states of the encoders and integrating their positional information when deriving the context vector.

The designed global attention mechanism begins by calculating alignment scores that measure the relevance of each encoder hidden state to the decoder’s hidden state at the previous step . This relevance is determined by combining ​ and ​ through a learned function, typically involving a concatenation of the two vectors resulting in a raw alignment score. At each decoding step, ​ is concatenated with ​, forming a joint representation that encapsulates both the input sequence's contextual features and the decoder's current focus. Without , the attention mechanism would only rely on , which could result in less accurate focus. The decoder state ​ provides a summary of what has been generated so far and what the model might need to focus on next [47].

The framework of the global position-aware attention model is presented in Fig. 3.

Based on their position index, the position distance is defined as the sub-linear function in Eq. (6), where is the relative distance between encoder token and the decoder’s current focus token . We adopt this logarithmic form because it provides a smooth attenuation of syntactic influence, nearby tokens receive stronger emphasis, while more distant tokens are gradually down-weighted without completely suppressing long-range dependencies that often carry meaningful aspect–opinion relations [53]. Then, the concatenation function is derived from the previous decoder's hidden state and each of the encoder’s hidden states is used to calculate alignment scores. The word position distance is considered as the weight decay rate for each word to optimize the alignment scores, as seen in Eq. (7). Subsequently, the attention weights are computed using the SoftMax function, as shown in Eq. (8). The encoder’s hidden states and their respective alignment scores are multiplied to form the global context vector, , which is defined in Eq. (9).,,,,

where is a trainable weight matrix and is a trainable weight vector.

Combining ​​ with the encoder hidden states enables the attention mechanism to dynamically assign weights to input tokens based on their relevance to the decoder’s hidden state at t-1. As explained in the neural machine translation context, this mechanism dynamically focuses on important parts of the input sequence, effectively aligning encoder and decoder interactions. Computed as a weighted sum of ​, ​ bridges this alignment by incorporating attention scores because capture the interplay between past decoding outputs and the input sequence.

Dependency parsing plays a critical role in syntactic analysis by identifying the relationships between words in a sentence [48]. In this study, we integrate dependency parsing with the encoder’s output vectors contextual representations of tokens to enhance the model’s ability to identify aspects across different domains. Reviews from various domains exhibit recurring syntactic patterns that facilitate information propagation. For example, syntactic relationships such as subject-predicate and subject-adjective connections between aspect terms and descriptive words are consistently observed across domains, as seen in Fig. 4. In the sentences of “The screen is great” and “The pizza is nice,” the parser identifies “screen” and “pizza” as subjects linked to their descriptive words “great” and “nice.” Although the vocabularies are different, the syntactic pattern remains the same, helping the model learn generalizable aspect relationships across domains. These patterns form the basis for transferring syntactic knowledge, enabling models to extract aspects more effectively [21] [41] [54].

(det: determiner; nsubj: nominal subject; acomp: adjectival complement).

Unlike the global attention mechanism, which attends to all encoder hidden states, the dependency based local attention mechanism selectively focuses on the syntactic children of each token. These relationships are determined using a dependency parse tree, which starts with a root token that usually is the main verb of a sentence. In this tree, every word is treated as a node, and its direct edges represent syntactic dependencies such as “nsubj,” “acomp,” “conj,” etc. As illustrated in Fig. 5, the token “staff” has children, “the,” “and” and “food”; both the token and the children are considered when calculating the local context vector. In this study, we used the spaCy library to generate a dependency parsing tree for each sentence.

The parse tree allows the model to identify the linguistic roles of words and perform the attention mechanism more precisely [55]. The encoder’s hidden states are structured according to the dependency tree of the sentence. Each word is considered as a parent node (or a parent head) connecting with its child nodes, and a child node can be a parent node for another set of child nodes. Table 1 displays the dependency parsing information of the sentence “The staff and food are exceptional” where token “are” was identified as Root. At time step t = 1 the token "staff" has three children in the dependency tree: “the” (i = 0), “and” (i = 2), and “food” (i = 4). At time step t, if a parent node (head) has m child nodes and their hidden states are (parent), a tanh transformation function can be established to compute the attention score as follows:,

where and are the weight matrices for the parent node and child nodes, and is a trainable weight vector. Similar to the global attention mechanism, the decoder’s hidden state ​ from the previous time step is concatenated with the encoder’s hidden states before applying the tanh activation function. The function allows the attention mechanism to focus on grammatically important words rather than merely adjacent ones. The attention weight for at time step t can be obtained by normalizing with a SoftMax function in the same way as Eq. (8). The local context vector is computed as the weighted sum of the relevant hidden state:.

Here, represents the relevance of the i-th token in contributing to the local context vector at time step t, effectively capturing the local syntactic context for the current token.

To integrate the global and local context vectors, a context fusion gate is built as shown in Fig. 6. The gate uses the sigmoid activation function (σ) to concatenate the global context vector and local context vector in a way shown in Eq. (13), which automatically controls how much information of the final context vector should be taken from the global and local context vectors. The final context vector can be obtained by multiplying the global and local vectors with their corresponding importance weight ( [56]:,

+ , (14)

where σ is sigmoid activation function which outputs a gating coefficient . is a trainable weight matrix used in the fusion gate. The fusion gate dynamically balances information between the global context vector and the local context vector by computing a gating coefficient with the sigmoid function that controls the importance of each input. This fusion keeps only the most relevant information for the decoding.

The input-feeding strategy has been implemented to enhance decoder performance in Seq2Seq models. Cha et al. [57] incorporate context vectors into subsequent hidden states, implicitly achieving a "coverage" effect that maintains alignment information across decoding steps. Similarly, Xu et al. [36] introduce a doubly attentional approach, which enforces training constraints to distribute attention evenly across all input parts, ensuring global coverage but limiting the model's flexibility. Kumar et al. [58] used attentional vectors as the feedback to maintain past alignment information and guide the decoder's focus. Our approach was built on these techniques by explicitly concatenating the previous prediction with the current context vector as the input to the decoding Bi-GRU. Let ​ denote the embedding of the previously generated token label . The vector , formed by concatenating and , is equal to . The decoder’s Bi-GRU then updates its hidden state at time step t as follows:

This explicit concatenation enables the decoder to adapt its focus based on both past outputs and the current input context and enhances task-specific contextual awareness through alignment-focused translation.

In this study, we compared our model against eight baseline models for aspect extraction. These baselines consist of four deep learning-based models and four models augmented with dependency parsing. The four deep learning-based models are:

The four-dependency parsing augmented models, which incorporate syntactic knowledge into deep neural networks, are:

The baseline models and our sequence-to-sequence with dual context (Seq2Seq-DC) model were trained using three annotated datasets: SE14-Laptop, SE14-Restaurant, and SE15-Restaurant [59] [60]. These datasets include aspect term annotations (e.g., battery, screen, food) with sentiment polarity labels (positive, negative, neutral). SE14-Laptop includes 3,841 review sentences focused on laptops, while SE14-Restaurant comprises 3,845 restaurant-related review sentences. SE15-Restaurant expands the dataset with an additional 2,000 sentences and extends sentiment classification beyond the sentence level [60]. Each dataset was partitioned into training, validation, and test sets using a split ratio of 63:16:21.

To evaluate the generalizability of our model, we applied it to unannotated target datasets derived from the Amazon Review Data collected by McAuley and Leskovec, [61]. The data include metadata, review texts and product ratings across 24 domains. For this study, we selected five domains: Musical Instruments (Music), Office Products (Office), Sports and Outdoors (Sport), Grocery and Gourmet Food (Food), and Video Games (Game). Table 2 summarizes key statistics for these five target datasets.

Our Seq2Seq-DC model and ten baseline models were trained and evaluated on three annotated datasets, with their performance measured by F1 scores as shown in Table 3. The F1 scores for the baselines were sourced from the respective cited references. Among the vanilla deep learning baselines, DE-CNN achieved the highest F1 score of 81.59% on the SE14-Laptop dataset, while Seq2Seq excelled on the SE14-Restaurant and SE15-Restaurant datasets with F1 scores of 87.42% and 72.34%, respectively. Overall, Seq2Seq appears to be the most effective model within this group. In the dependency parsing-augmented baselines, RINANTE had the highest F1 score of 80.61% on SE14-Laptop, while DREGCN achieved the highest F1 scores of 87.00% and 73.30% on the two restaurant datasets. The average F1 scores for the two groups of baselines across the three datasets (bolded numbers in the table) indicate that the dependency parsing-augmented baselines slightly outperformed the vanilla deep learning baselines overall. This suggests that incorporating syntactic information improves aspect extraction performance beyond what is achieved with contextual information alone. Our Seq2Seq-DC model, which integrates Seq2Seq with dependency parsing, significantly outperformed both groups of baselines. It achieved F1 scores of 82.66% for SE14-Laptop, 88.66% for SE14-Restaurant, and 74.04% for SE15-Restaurant. These scores represent increases of 3.55%, 3.26%, and 5.45%, respectively, for the three datasets, as opposed to the average F1 scores of the eight baselines. The Δ values represent the relative differences in F1 scores between the baselines and Seq2Seq-DC, highlighting the performance gain achieved by Seq2Seq-DC.

* F1 scores of the baselines were sourced from the respective cited references. The bolded numbers are the average F1 scores of the deep learning-based and dependency parsing-augmented baselines across the three datasets. Δ values represent the relative differences in F1 scores of the baselines compared to those of Seq2Seq-DC.

As previously described, the Seq2Seq-DC model primarily consists of a series of Bi-GRUs and an attention mechanism layer. The attention mechanism layer includes two modules designed to capture both global and local contextual information from review sentences. To evaluate the contributions of the global contextual (GC) and local contextual (LC) modules to the model’s performance, we conducted an ablation study by selectively removing each module and evaluating the resulting models using the F1 score. The LC module includes dependency parsing (DP), thereby reflecting its impact on performance as well. Table 4 presents a comparison of F1 scores between the ablated models and the complete Seq2Seq-DC model (Bi-GRU + GC + LC). Specifically, Bi-GRU + GC refers to the model without the LC module, while Bi-GRU + LC denotes the model without the GC module. As shown in Table 4, both Bi-GRU + GC and Bi-GRU + LC show negative Δ values, reflecting a decrease in F1 scores compared to Seq2Seq-DC. However, Bi-GRU + LC consistently achieves higher F1 scores across all three datasets than Bi-GRU + GC, suggesting that the LC module has a greater impact on the model’s performance in aspect extraction. This can be attributed to the LC module’s ability to capture syntactically adjacent words, which are particularly important for accurately identifying aspect terms.

Aspect summarization is a process to distill a wide range of aspect terms into a concise set of representative concepts, which are more focused and interpretable for downstream tasks like recommendation or visualization [62]. To validate the applicability of our model, we performed aspect summarization using the five Amazon review datasets. In the Music dataset, which contains 3105 unique tokens (excluding repetitions), the model extracted 232 aspect terms. For the other datasets, the number of extracted aspect terms was as follows: 506 for Office (8638 tokens), 341 for the Food dataset (4997 tokens), 583 for Game (7513 tokens), 232 for Sport (3134 tokens). The number of the aspect terms in each dataset reflects a high-dimensional semantic space. To reduce the dimensionality, PCA was applied to project the aspect-term space onto a smaller set of principal components (PCs). PCA acted as a semantic filter by emphasizing components that capture the most meaningful variance while discarding less informative ones. We implemented PCA with Python’s Scikit-learn library, and plotted the cumulative explained variance against the number of PCs for each Amazon review dataset in Fig. 7. The dashed line in the figure marks the number of PCs required to explain 95% of the total variance: 19 for Office, 23 for Game, 30 for Food and Music, and 34 for Sport.

Table 5 presents the contribution of the PCs for each dataset. The first PC captures the largest amount of variance, representing the direction of greatest variability in the data. Each subsequent PC accounts for progressively smaller portions of variance along orthogonal directions, ensuring no redundancy in variance explanation. The last PC for each dataset was determined by setting a cumulative variance threshold of 95%, meaning the remaining unexplained variance is less than 5%.

PCs are orthogonal linear combinations of the original aspect terms. To filter out less informative terms, those with variance primarily located in the bottom 5% of components can be excluded from further analysis [63]. As a result, we focus only on the PCs that collectively explain 95% of the variance. Table 6 shows the number of aspect terms before and after applying PCA, with reductions of 36.64% for Music, 62.64% for Office, 53.37% for Food, 34.05% for Sport, and 68.10% for Game.

Having reduced the dimensionality with PCA, we applied K-means clustering to group aspect terms with similar semantic meanings in the principal component space. The optimal number of clusters for each dataset was determined using the elbow method [64], which indicated that five clusters would provide a balance between model complexity and variance explained across all datasets. K-means clustering was implemented using Scikit-learn library. Since the first three PCs capture the majority of the variance in the original aspect term space, they were used to visualize the spatial distribution of clustered terms across the five datasets, as shown in Fig. 8. Each point represents an aspect term, with colors indicating the clusters assigned by K-means clustering. Notably, the Music and Food datasets exhibit relatively well-separated clusters, implying stronger semantic distinctions among their aspects, while the Office dataset shows more overlap between clusters, suggesting thematic interdependence or ambiguity among its terms. Each cluster in a dataset groups aspect terms with similar semantic meanings, allowing for the derivation of a more abstract label, referred to as the primal term. However, the significance of individual aspect terms can vary within a cluster, making it necessary to narrow down the terms used when identifying an appropriate primal term. The frequency of a term mentioned in reviews of the dataset may be used as an indicator of the term’s significance.

As an example, Fig. 9 presents the sorted relative frequencies of the top 15 terms in each cluster from the Office dataset. In each cluster, the frequencies are concentrated among a few leading terms, indicating their dominance within the cluster. To automatically derive a concise primal term for each semantic cluster, we used the GPT-3.5 API with prompts to for cluster-level abstraction that generate primal terms that represent the core semantic content of the cluster [65]. For each cluster, we selected several top-frequency aspect terms as prompts since their frequencies are dominant the distributions (Fig. 9). Through trial tests using different numbers of top terms, we found that the top five aspect terms can achieve a good balance between semantic richness and interpretability, consistently producing meaningful and representative primal terms.

Table 7 lists the top five aspect terms and their relative frequencies from user reviews for each cluster across the five Amazon datasets. For instance, in the Game dataset, Cluster 2 includes the top five terms: gameplay (1.6%), fun (1.3%), adventure (1.3%), player (1.0%), and story (1.0%). The GPT-3.5 API generated “gameplay experience” as the primal term. In the Music dataset, the top five terms for one cluster are strings (8.4%), guitar (7.9%), pick (3.5%), pedals (2.8%), and price (2.0%), which led to the primal term “guitar gear.” The results suggest that the GPT-3.5-generated primal terms are both semantically appropriate and scalable for representing the core content of each cluster. These terms serve as meaningful labels that highlight the key issues that users care about most, and thus provide useful perspectives for the subsequent aspect-based sentiment analysis.

*The numbers in the parentheses are the relative frequencies of the top five aspect terms appearing in the user reviews.