Entity-Level Sentiment Analysis with Sentence Relevance Detection

Abstract

The task of entity-level sentiment analysis (Elsa) is to extract sentiment scores for a given entity (such as person names or organization names) from a text. Elsa is a challenging task and involves processing of longer documents, where several entities may be mentioned with varying importance for the final score aggregation. Fine-tuning encoder-based Transformers (such as BERT) constitutes the state of the art for sentiment predictions, however, these models are still limited by their restricted input lengths. Decoder-only models so far still underperform on the task. We approach the context limitation by learning to extract segments that are relevant for the sentiment prediction for a given entity, without preprocessing by chunking and aggregation. For decoder models, we explore fine-tuning these through supervised fine-tuning and pairwise comparison, a method borrowed from reward modeling for preference optimization. Both methods perform well and set a new standard for the Elsa task. We further show that pairwise classification is faster, simpler, and shows less variance than the more common direct supervision for this task.

Resources

Details