Attribute‑Level Sentiment Analysis for E‑commerce: Tasks, Challenges, and System Design

Sentiment analysis, a core problem in natural language processing, has grown rapidly with the rise of user‑generated content on social media and review platforms. It can be divided into three granularity levels: document‑level, sentence‑level, and aspect (attribute)‑level, each addressing increasingly fine‑grained opinion extraction.

In the e‑commerce scenario, massive product reviews on platforms like Taobao contain rich opinion information. To help users make informed purchase decisions, the task is formulated as Aspect Sentiment Triplet Extraction, which aims to extract <aspect‑term, opinion‑word, polarity> triples from unstructured text.

The proposed pipeline decomposes the problem into three sub‑tasks: (1) joint extraction of aspect terms and opinion words, (2) pairing of aspect‑opinion, and (3) fine‑grained polarity classification. The system leverages a BERT + BiLSTM + CRF baseline, enhanced by continue‑pretraining on review data, multi‑domain representation learning (shared and private spaces with attention), and attribute normalization using Sentence‑BERT embeddings clustered by DBSCAN.

Key technical challenges include extreme domain variation (different product categories use distinct vocabularies and sentiment cues), severe class imbalance (positive to negative ratios up to 20:1), and scarce annotated data due to high labeling cost. The pipeline addresses these by domain‑aware pre‑training, knowledge‑enhanced masking of aspect‑opinion pairs, and semi‑supervised consistency training on unlabeled data.

Experimental results show notable business impact: impression‑word CTR increased by 400%, personalized ranking CTR by 25% and UV‑CTR by 20%, and negative‑sentiment F1 improved by 4 points. The system has been deployed across multiple downstream scenarios such as impression‑word ranking, mini‑detail page recommendation, short‑video tag generation, and negative‑sentiment filtering for UI safety.

Finally, the article outlines future work, including multi‑entity aspect analysis, handling comparative and contrastive sentences, reducing error propagation in the pipeline, and exploring end‑to‑end aspect‑sentiment extraction models.