Fine‑Grained Sentiment Analysis and Opinion Quadruple Extraction: Methods, Tasks, and Applications

The article begins with an overview of text sentiment analysis, distinguishing between sentiment (positive/negative) and emotion (fine‑grained affect), and outlines its wide applications such as product reviews, political monitoring, and financial analysis.

It then details the main tasks in sentiment analysis: sentence/document‑level classification, aspect‑level (attribute) analysis, and target‑level analysis, introducing the five‑element schema (entity, aspect, opinion, holder, time) and the concept of a target as a combination of entity and aspect.

For attribute‑level sentiment analysis (TG‑ABSA), the paper reviews three families of methods: unsupervised rule‑based approaches using syntactic parsing, reading‑comprehension‑style models that treat each attribute as a question answered by BERT, and a proposed multi‑task multi‑label classification framework that predicts sentiment for each predefined attribute while handling missing labels via label‑masking and active learning.

Experimental results on automotive and mobile‑phone datasets show that the multi‑task model achieves over 90% accuracy for each attribute and Fuzzy F1 scores around 0.79, with human evaluation confirming roughly 96% perceived correctness.

The second part introduces opinion‑quadruple extraction, which not only predicts the sentiment polarity of an attribute but also locates the attribute word and the opinion word in the text, forming a four‑element tuple (attribute, attribute‑type, opinion, polarity). A joint extract‑and‑classify architecture is presented, combining a CRF‑based sequence labeler for span detection with a BERT/ RoBERTa encoder for attribute‑level classification.

Data annotation procedures, including a custom labeling platform and strategies for handling imbalanced attribute distributions, are described, followed by evaluation using a fuzzy F1 metric and a manual sanity‑check that yields high agreement.

Finally, the article discusses deployment considerations on Huawei Cloud, future trends such as low‑cost domain adaptation, self‑supervised large‑scale pre‑training, multimodal sentiment analysis, and knowledge‑graph‑enhanced models.