Attribute Extraction and Sentiment Analysis in NLP

Product reviews, support tickets, and survey comments rarely express sentiment as a single document-level label. They mention concrete attributes, such as battery life, delivery, staff, screen, price, or policy, and assign different polarities to each of them. This paper presents Boundary-Aware Attribute Sentiment Extraction (BASE), a joint framework for extracting attribute spans, opinion spans, and attribute-level sentiment labels from review text. BASE combines a pre-trained language encoder with a compact sequence-labeling layer, a span-pair compatibility scorer, and a calibration step that suppresses unsupported sentiment assignments. The framework is designed for settings in which annotated aspect-based sentiment data are limited, domain vocabulary shifts across products, and downstream users need inspectable attribute--opinion evidence rather than only a polarity label. A controlled evaluation on laptop, restaurant, and multi-aspect review benchmarks compares rule-based extraction, CRF tagging, neural sequence labeling, BERT-style sentence-pair classification, and the proposed joint model. Results show that joint span evidence improves attribute sentiment F1 under multi-aspect sentences and cross-domain transfer, while calibrated abstention reduces unsupported sentiment assignments. The study argues that attribute extraction and sentiment analysis should be evaluated as linked evidence recovery rather than as independent extraction and classification subtasks.