Aspect-based sentiment analysis (ABSA) goes beyond document-level or sentence-level sentiment classification by identifying the specific aspects (also called features or attributes) of an entity that are discussed and determining the sentiment toward each aspect independently. A restaurant review such as "The food was excellent but the service was terrible" contains two aspects (food and service) with opposite polarities. Document-level analysis would either average these sentiments or arbitrarily assign a single label, losing critical information. ABSA preserves this fine-grained structure, making it essential for applications where understanding specific strengths and weaknesses matters.
Aspect Term Extraction
Step 2: Sentiment classification — determine polarity per aspect
Joint: P(aspect, sentiment | sentence) via sequence labelling
Tags: {B-POS, I-POS, B-NEG, I-NEG, B-NEU, I-NEU, O}
Aspect extraction can be approached as either explicit aspect term extraction or implicit aspect category detection. Explicit aspect terms are noun phrases mentioned directly in the text (e.g., "battery life," "screen resolution"). These can be extracted using sequence labelling models (CRF, BiLSTM-CRF) or dependency parsing rules. Implicit aspects are not explicitly mentioned but are inferred from context: "This phone is too heavy" implicitly refers to the weight aspect. The SemEval ABSA shared tasks (2014–2016) established standard benchmarks for both subtasks, defining aspect categories such as food, service, ambience, and price for the restaurant domain.
Aspect Sentiment Classification
Once aspects are identified, the sentiment toward each aspect must be determined. This is challenging because the same sentence may express different sentiments toward different aspects, and the sentiment-bearing words may be distant from the aspect term. Attention-based neural models address this by learning to attend to the parts of the sentence most relevant to each aspect. The Attention-Based LSTM (Wang et al., 2016) computes attention weights conditioned on the aspect embedding, focusing on sentiment-relevant context words. More recent approaches use graph neural networks over dependency parse trees to model the syntactic relationship between aspect terms and opinion words.
The SemEval shared tasks on Aspect-Based Sentiment Analysis (2014, 2015, 2016) created standardised benchmarks that catalysed research in the field. The tasks defined subtasks for aspect term extraction, aspect category detection, and aspect-level sentiment classification across domains including restaurants and laptops. These benchmarks revealed that aspect extraction is substantially harder than sentiment classification and that joint models that extract aspects and classify sentiment simultaneously can outperform pipeline approaches by exploiting the mutual dependence between the two subtasks.
End-to-end approaches jointly extract aspect terms and classify their sentiment in a single model, avoiding error propagation from the aspect extraction step to the sentiment classification step. Pretrained language models such as BERT have been adapted for ABSA by formulating the task as a sentence-pair classification problem, where the aspect term is paired with the review sentence, or as a question answering problem, where questions about specific aspects are answered with sentiment labels. These approaches achieve state-of-the-art performance and transfer effectively across domains with limited fine-tuning data.