Aspect-based sentiment analysis with multi-granularity information mining and sentiment hint

Abstract

Most aspect-based sentiment analysis approaches at this stage primarily focus on extracting information from the comment sentence and analyzing the interaction between the comment sentence and aspect words while ignoring the importance of key phrase information. Sentences are composed of words and phrases following specific grammatical rules, and mining phrase information can provide valuable insights into short-distance collocations between words. Thus, integrating multi-granularities (sentences, phrases, and aspect words) in the learning process can lead to better vector representations. This paper proposes a multi-granularity information mining and sentiment hint model (MGSH). Specifically, we first use two phrase extraction techniques to separately extract a set of key phrases from comment sentences and use heuristic algorithms to merge and eliminate the redundancy of the two sets. In addition, we design a constrained multi-granularity fusion module to dynamically integrate information from different granularities, encouraging the model to capture semantic associations among these granularities. Finally, we conducted a statistical analysis on the number of aspect words in the comment sentences. We found a potential correlation between the overall emotion of the sentence and the aspect words. Therefore, we pre-assign sentiment pseudo-labels to the commented sentences to identify sentiment cues specific to aspects and potential sentiment relationships between sentences and aspects. Our experimental evaluations, conducted on three benchmark datasets, showcase the importance of multi-granularity information mining and interaction as well as the effectiveness of utilizing sentiment cue information.