Post-processing method with aspect term error correction for enhancing aspect term extraction

Ruyan Wang

doi:10.1007/s10489-022-03380-z

Return

Post-processing method with aspect term error correction for enhancing aspect term extraction

2022-03-18

PRE

DOI：10.1007/s10489-022-03380-z

Original

Original request for help

Save

Abstract

En 中文

Aspect Term Extraction (ATE), which aims to extract aspect terms from review sentences, is an important subtask of sentiment analysis. Existing studies have proposed many sequence taggers, which have achieved impressive progress. However, previous work ignores the errors in aspect terms extracted by sequence taggers. This paper points out that there are aspect number error and aspect boundary error in the extracted aspect terms, which have a severe negative impact on the performance of the model on ATE task. We design post-processing method with aspect term error correction, which contain aspect number determining module and aspect boundary modifying module, to address the errors. To solve the inconsistency between the number of extracted aspect terms and the number of ground-truth aspect terms, we use the aspect number determining module to control the number of extracted aspect terms to match the number of ground-truth aspect terms. For the problem that the boundary of the extracted aspect term is not completely matched with the boundary of the ground-truth aspect term, we utilize the aspect boundary modifying module to correct the boundary of the extracted aspect term to make it the same as the boundary of the ground-truth aspect term. Experiments on four SemEval datasets show that the post-processing method can address the errors in extracted aspect terms, and our model outperforms other SOTA models. Our post-processing method can be coupled with various sequence taggers to boost their performance and have well transferability.

Keywords:

Sentiment analysis

Aspect term extraction

Aspect number error

Aspect boundary error

Aspect term error correction