Aspect-Level Sentiment Classification

• Aspect-level sentiment classification is defined as determining sentiment polarity for specific aspects within texts rather than overall sentiment.
• It leverages varied input representations—from BoW and TF-IDF to transformer embeddings—enabling nuanced, context-sensitive analysis.
• Research trends highlight a shift from knowledge-based methods to deep neural, attention, and graph-based models, with transfer learning addressing data scarcity.

Aspect-level sentiment classification, also termed aspect-based sentiment classification, predicts the polarity expressed toward a specified aspect rather than assigning a single label to an entire sentence or document. In a standard formulation, a sentence $s=(w_1,w_2,\dots,w_n)$ and an aspect term $a$ at token positions $p=\{p_1,p_2,\ldots,p_m\}$ are mapped by a classifier $f(a,s)=y$, where $$y\in\{\text{negative},\text{neutral},\text{positive}\}$$; equivalently, one may predict $\hat y=\mathrm{softmax}(Wx+b)$ and choose $y=\arg\max_c \hat y_c$ (Bhoi et al., 2018). The task differs from sentence-level sentiment analysis because a single context can contain conflicting opinions about different aspects, as in examples where one attribute is praised and another is criticized (Brauwers et al., 2022).

1. Problem variants and input representation

Two closely related problem settings are commonly distinguished. In aspect-category (AC) classification, the input contains a pre-defined aspect category such as “food” or “service,” and the category never appears explicitly in the sentence. In aspect-term (AT) classification, the input contains an aspect span $a\subset x$ whose exact positions are known because the aspect is a subsequence of the sentence (Li et al., 2018). This distinction is consequential because AC tasks lack term-level position information, whereas AT tasks permit explicit position-aware modeling.

Input construction has evolved from manual feature engineering to end-to-end neural representations, but several representational primitives recur. A survey of ABSC notes that explicit aspects may be handled through a window around the target or a dependency-based subtree, while implicit aspects often require the entire sentence or proxy-word clustering; inputs may be represented either as a vector $x\in\mathbb{R}^{d_x}$, for example with BoW or TF-IDF, or as a matrix $X\in\mathbb{R}^{d_x\times n_x}$ in which each column is a word embedding (Brauwers et al., 2022). The 2018 survey on “Various Approaches to Aspect-based Sentiment Analysis” enumerates standard preprocessing operations: tokenization, padding or truncation to a fixed length $a$0, and three aspect-marking schemes—ID-Encoding, Bit-Masking, and Location-Encoding. It also lists Bag-of-Words, POS-tag features, dependency-path features, and position embeddings, with traditional classifiers receiving the concatenated representation

$a$1

(Bhoi et al., 2018).

This representational layer already exposes a central modeling question: how to encode the aspect so that sentiment-bearing tokens relevant to one target are amplified while sentiment-bearing tokens associated with other targets are suppressed. Much of the subsequent literature can be read as progressively more structured answers to that question.

2. From knowledge-based systems to multi-hop neural models

A broad taxonomy groups ABSC models into three major categories: knowledge-based, machine learning, and hybrid models. Knowledge-based systems include dictionary-based, ontology-based, and discourse-based methods; machine-learning systems include SVMs, tree-based models, deep-learning models, attention-based architectures, and transformer-based models; hybrid systems inject dictionary, ontology, or discourse signals into learned classifiers (Brauwers et al., 2022). This taxonomy also tracks the field’s reported trend: a shift from purely knowledge-based methods to SVMs, then to deep nets, attention mechanisms, and transformers.

In feature-based settings, the decision rule is typically linear. The 2018 survey presents logistic regression with $a$2 and cross-entropy loss, alongside multiclass SVM formulations, over concatenated handcrafted features (Bhoi et al., 2018). On modified SemEval-2016 Task 5 datasets, the same survey reports the following accuracies for representative classical systems: Naive Bayes + Oh, 0.540 on Tech and 0.520 on Food; SVM + Oh, 0.611 and 0.663; XGBoost + Loc, 0.651 and 0.633; ExtraTrees + Loc, 0.702 and 0.636. In the same comparison, MemNet $a$3 reaches 0.713 on Tech and 0.787 on Food, outperforming all classical models (Bhoi et al., 2018).

The deep memory network is an early neural architecture explicitly designed around aspect-conditioned relevance estimation. It treats all non-aspect word embeddings as an external memory matrix, applies a content-based attention mechanism at each hop, and updates the query vector by combining attended context with the previous hop representation. Tang et al. report that multiple computational layers improve performance, that single-hop performance is approximately LSTM quality, and that accuracy rises up to 6 hops before saturating around 7–9 hops; on SemEval-2014, the 9-hop model achieves 72.21% on Laptop and 80.95% on Restaurant, while being 15 times faster than LSTM with a CPU implementation (Tang et al., 2016). In this phase of the literature, the key departure from standard recurrent encoders is not recurrence per se but repeated aspect-specific re-reading of context.

Transformer-based models later became the dominant benchmark setters. The survey records that BERT-ABSC reaches accuracy up to 0.846 on SemEval-14 restaurants, and LCF-BERT reaches 0.871 accuracy and 0.817 F$a$4 on the same benchmark (Brauwers et al., 2022). This suggests that contextualized encoders improved both token semantics and the quality of downstream target conditioning, even before explicit syntactic or knowledge-graph structure was added.

3. Attention, mutual interaction, and aspect-conditioned convolution

A major strand of ABSC research focuses on modeling the target and its context as two interacting sequences rather than collapsing the aspect into a single static vector. Interactive Attention Networks (IAN) instantiate this view with two parallel LSTM-based subnetworks: one for the context and one for the target. The averaged target summary attends to context hidden states to produce $a$5, while the averaged context summary attends to target hidden states to produce $a$6; the concatenation $a$7 is then classified by a feed-forward layer and softmax (Ma et al., 2017). On SemEval 2014 Task 4, IAN reports 0.786 accuracy on Restaurant and 0.721 on Laptop, outperforming LSTM, TD-LSTM, AE-LSTM, and ATAE-LSTM; ablations show that both directions of interaction contribute, since “No-Target,” “No-Interaction,” and “Target2Content” all underperform the full model (Ma et al., 2017).

Attention-over-Attention (AOA) sharpens this interaction by constructing a sentence–aspect interaction matrix

$a$8

applying column-wise and row-wise softmax to obtain aspect-to-sentence attention $a$9 and sentence-to-aspect attention $p=\{p_1,p_2,\ldots,p_m\}$0, averaging $p=\{p_1,p_2,\ldots,p_m\}$1 into $p=\{p_1,p_2,\ldots,p_m\}$2, and computing the final sentence attention as $p=\{p_1,p_2,\ldots,p_m\}$3 (Huang et al., 2018). The attended sentence representation is then classified by a linear layer and softmax. On SemEval 2014 Task 4, AOA-LSTM reports 0.812 accuracy on Restaurant and 0.745 on Laptop, exceeding earlier LSTM-based baselines including IAN (Huang et al., 2018). The qualitative analyses in that paper emphasize two recurring ABSC difficulties: multiple aspects in the same sentence and multi-word aspect terms such as “boot time.”

Convolutional architectures pursue the same aspect-conditioning objective with different inductive biases. Xing et al. introduce an attention-based input layer in which cosine similarity between each context embedding $p=\{p_1,p_2,\ldots,p_m\}$4 and the aspect embedding $p=\{p_1,p_2,\ldots,p_m\}$5 yields weights $p=\{p_1,p_2,\ldots,p_m\}$6, after which the transformed input is passed through CNN filters of sizes $p=\{p_1,p_2,\ldots,p_m\}$7, max-over-time pooling, and a final softmax (Xing et al., 2018). On the Twitter benchmark of Dong et al. (2014), the best variant, CNN+attenEmb$p=\{p_1,p_2,\ldots,p_m\}$8, achieves Acc=0.725 and F1=0.702, outperforming TD-LSTM and TC-LSTM without using a syntactic parser or external sentiment lexicons (Xing et al., 2018). Huang and Carley further parameterize the convolution itself: PF-CNN makes the filters depend on the aspect representation, whereas PG-CNN uses aspect-conditioned gates to modulate general convolutional responses. On SemEval-2014, PF-CNN reaches 70.06% on Laptops and 79.20% on Restaurants for the 3-way task, while PG-CNN reaches 69.12% and 78.93%, both improving over a vanilla CNN baseline (Huang et al., 2019).

Across these models, the central design move is the same: sentiment classification becomes target-conditioned representation learning. The differences lie in whether target conditioning is expressed through mutual attention, interaction matrices, input transformation, parameterized filters, or gates.

4. Syntax, graph structure, and external knowledge

Sequence models with token-distance attention often struggle when the decisive opinion word is not linearly adjacent to the aspect. Syntax-aware models address this by replacing or augmenting sequence proximity with graph structure derived from dependencies. TD-GAT constructs an undirected word-level dependency graph, collapses a multi-word aspect into a single special token “target,” applies multi-layer graph attention, and updates the target node through an LSTM-based target-dependency mechanism that propagates layer-wise context while suppressing noise (Huang et al., 2019). With GloVe embeddings, TD-GAT-GloVe reaches 74.0 on Laptop and 81.2 on Restaurant; with fixed BERT representations projected to 300 dimensions, TD-GAT-BERT reaches 80.1 on Laptop and 82.8 on Restaurant (Huang et al., 2019). The paper’s analysis argues that syntax graphs place sentiment modifiers directly adjacent to their targets in the graph even when they are distant in token order.

A different graph construction appears in SDGCN, which treats aspects in the same sentence as nodes in an aspect graph and uses GCN layers to capture sentiment dependencies between multiple aspects (Zhaoa et al., 2019). The model first derives aspect-specific representations by bidirectional attention with position encoding, then applies GCN over either an adjacent-relation graph or a fully connected global-relation graph. On SemEval-2014, SDGCN-G reaches 82.95 / 75.79 on Restaurant and 75.55 / 71.35 on Laptop in Accuracy / Macro-F1 for GloVe-based models, and SDGCN-BERT reaches 83.57 / 76.47 on Restaurant and 81.35 / 78.34 on Laptop (Zhaoa et al., 2019). This line of work treats co-occurring aspects not as independent instances but as coupled sentiment variables.

External knowledge has also been incorporated directly into the representation. Knowledge Graph Enhanced Aspect-Level Sentiment Analysis combines a two-layer Transformer (“BERT”) encoder, knowledge-graph synonym retrieval, a dynamic attention mechanism with a sentinel vector, a location-weighted memory bank, and a DCGRU sequence model (Sharma et al., 2023). Its reported results are 79.64 / 75.43 on Laptop, 86.25 / 77.74 on Restaurant, and 76.13 / 75.27 on Twitter in ACC / Mac-F1 (Sharma et al., 2023). The model targets context-specific word meanings by allowing synonym knowledge to be fused or ignored dynamically.

Interpretability and bias have become additional concerns in this structurally enriched phase. SARL proposes span-based anti-bias aspect representation learning: an adversarial module removes an aspect’s prior sentiment, while a distilled aspect-opinion dependency aligner scores candidate opinion spans and can return a dummy “no opinion” option for neutral cases (Wang et al., 2021). The paper reports average improvements of +1.0 ppt Accu and +1.3 ppt Ma-F1 over the next best RoBERTa-based baseline, and unsupervised opinion extraction top-3 F1 of 77.1% on Laptop14 and 86.1% on Rest14 (Wang et al., 2021). A plausible implication is that syntax and knowledge are not only accuracy-enhancing priors but also mechanisms for making aspect-specific evidence more explicit.

5. Transfer learning, low-resource regimes, and zero-shot classification

Data scarcity is a recurring structural constraint in ABSC because aspect-level annotation is more expensive than document-level labeling. One early remedy is transfer from document sentiment. “Exploiting Document Knowledge for Aspect-level Sentiment Classification” trains an attention-based LSTM aspect classifier and transfers document-level knowledge through pretraining (PRET), multi-task learning (MULT), or their combination (PRET+MULT), using 30K balanced reviews each from Yelp2014 and Amazon Electronics (He et al., 2018). PRET+MULT reports 79.11 / 69.73 on Restaurant2014, 71.15 / 67.46 on Laptop2014, 81.30 / 68.74 on Restaurant2015, and 85.58 / 69.76 on Restaurant2016 in Accuracy / Macro-F1, setting a new state-of-the-art on all four benchmarks in that comparison (He et al., 2018). The transfer is operationalized by sharing embeddings and LSTM parameters, not merely by domain adaptation at the classifier head.

A more fine-grained transfer problem arises between coarse aspect categories and fine-grained aspect terms. MGAN addresses this “coarse-to-fine task transfer” with a source AC branch and a target AT branch linked by shared BiLSTM, Context2Aspect, and Position-aware Sentiment modules; a source-specific Coarse2Fine attention identifies likely term-level realizations for aspect categories, while contrastive feature alignment pulls same-class cross-domain pairs together and pushes different-class pairs apart (Li et al., 2018). Averaged across Laptop, Restaurant, and Twitter target domains, full MGAN reaches 77.44% accuracy and 72.14% Macro-F1, outperforming both non-transfer baselines and transfer baselines such as FT, M-DAN, and M-MMD (Li et al., 2018).

Attention Transfer Network (ATN) transfers not just generic sentiment features but attention knowledge itself. A pre-trained document-level BiLSTM attention module supplies global attention $p=\{p_1,p_2,\ldots,p_m\}$9, which is then used either as a soft target for aspect-level attention through Attention Guidance or fused with target-dependent attention through Attention Fusion (Zhao et al., 2020). On SemEval 2014 Task 4, ATN-AG reaches 81.39 / 72.44 on Restaurant and 76.41 / 72.59 on Laptop, while ATN-AF reaches 82.36 / 74.00 and 76.48 / 72.60 in Accuracy / Macro-F1, with improvements significant $f(a,s)=y$0 over all baselines (Zhao et al., 2020). The reported effect of increasing document-level data size is monotonic improvement from 0% to 100% DSC data.

Zero-shot ASC pushes this logic further by removing aspect-level supervision entirely. AF-DSC trains only on document-level reviews with ratings and assumes that document sentiment is composed of the sentiments of all aspects mentioned in the review; it extracts potential aspects by POS rules, uses BERT to build token representations, masks attention to nouns, and trains with word sentiment prediction, mask word prediction, and document-rating prediction (Deng et al., 2022). On zero-shot evaluation, AF-DSC reports 69.6 / 61.7 on Res14 and 65.4 / 57.8 on Lap14 in accuracy / macro-F1, and on ARTS it reaches 39.1 / 38.3 on Restaurant and 39.3 / 36.7 on Laptop (Deng et al., 2022). This suggests that explicit aspect-to-document composition can outperform prior zero-shot systems even when trained with only 30K document-level reviews.

6. Evaluation practice, failure modes, and expanding benchmarks

Reported evaluation in ABSC usually centers on accuracy, precision, recall, F$f(a,s)=y$1, and Macro-F$f(a,s)=y$2, with Macro-F$f(a,s)=y$3 emphasized as robust under class imbalance; ranking loss, MSE, and AUC also appear in the survey taxonomy, especially outside the standard three-way SemEval setting (Brauwers et al., 2022). Benchmarking has historically been dominated by SemEval 2014 Task 4 and later SemEval restaurant tasks, but the field’s evaluation culture has increasingly focused on harder distributions, multilingual settings, and domain-specific corpora.

One important corrective comes from “A Failure of Aspect Sentiment Classifiers and an Adaptive Re-weighting Solution,” which shows that many systems trained on standard benchmarks behave more like sentence-level sentiment classifiers than genuine aspect-sensitive models (Xu et al., 2019). In the SemEval 2014 data examined there, only 11.3% of Laptop training sentences and 16.1% of Restaurant training sentences are contrastive. On the full test set, BERT-DK reports 76.9 / 73.7 on Laptop and 84.2 / 76.2 on Restaurant in Accuracy / Macro-F1, but on the contrastive test set it drops to 51.1 / 50.0 and 65.5 / 66.9; with ARW, the contrastive scores rise to 57.3 / 56.5 on Laptop and 74.0 / 74.6 on Restaurant (Xu et al., 2019). The paper’s diagnosis is that rare contrastive sentences let models minimize loss by ignoring the aspect input.

Recent datasets make this contextualization problem explicit. CARBD-Ko introduces Korean aspect-opinion pairs with dual labels: aspect-agnostic polarity $f(a,s)=y$4, aspect-specific polarity $f(a,s)=y$5, and intensity $f(a,s)=y$6 (Jang et al., 2024). On this benchmark, kr-electra reaches 79.8% on $f(a,s)=y$7, 85.5% on $f(a,s)=y$8, and 74.1% when both are required to be correct, while all evaluated PLMs exceed 90% on NSMC but perform markedly worse on the dual-polarity task (Jang et al., 2024). The benchmark formalizes the difference between global and local sentiment that earlier failure analyses had identified implicitly.

ABSC has also moved into specialized scientific-review domains. A 2026 study on multi-round peer review comments from Nature Communications constructs a corpus of 5,063 aspect-sentiment-labeled sentences and trains multiple deep learning classifiers, with LCF-BERT-CDM achieving Accuracy 82.72 and Macro-F1 82.65 (Hana et al., 23 Jun 2026). The downstream analysis reports that as the number of review rounds increases, positive aspect-level sentiments rise and negative sentiments decline, and that aspect sentiment scores are negatively associated with the total number of review rounds, especially for “experiments,” “research significance,” and “result analysis” (Hana et al., 23 Jun 2026). This extension indicates that ABSC now functions not only as a benchmark task but also as an analytic instrument for domain-specific discourse.

Open problems identified in the survey remain active: data scarcity in niche domains and non-English languages, implicit aspect detection, sarcasm and thwarting, interpretability of deep and transformer models, dynamic sentiment shifts, and fine-grained multi-polar, multi-label, and multi-task settings (Brauwers et al., 2022). Proposed future directions include semi-supervised and cross-lingual methods, automatic construction of domain ontologies or lexicons, deeper integration of symbolic and subsymbolic AI, modeling temporal evolution of aspect sentiments, and incorporation of multimodal signals (Brauwers et al., 2022). In that sense, aspect-level sentiment classification has developed from a narrowly defined three-way sentence classification problem into a broader family of target-conditioned inference problems spanning structure, transfer, explanation, and robustness.