Aspect Term Extraction via Dynamic Attention and a Densely Connected Graph Convolutional Network

Xin Sun*, Yongqing Mi, Jia Liu, Hongao Li

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Aspect term extraction is a crucial step in aspect-level sentiment analysis, significantly affecting the accuracy of sentiment classification. Therefore, improving the precision of aspect term extraction is vital for enhancing the performance of sentiment analysis. The limitations of existing methods include inadequate consideration of syntactic information and inter-word dependencies, as well as the challenge of mitigating weight noise during dependency tree conversion. To address these issues, we propose an aspect term extraction approach that leverages dynamic attention and graph convolutional network. Our method utilizes a densely connected graph convolutional network to capture dependency information between distant terms, thereby enriching vector semantics. Furthermore, it integrates a dynamic attention mechanism informed by dependency parsing to highlight critical dependencies and mitigate noise interference. We benchmark our model against state-of-the-art approaches on four widely used public datasets. The results indicate that our proposed method significantly enhances the performance of aspect term extraction. Specifically, our model improves upon baseline models on the Lap14 and Rest15 datasets, with increases in macro-F1 scores of 0.45, and 0.04, respectively.

Original languageEnglish
Title of host publicationPRICAI 2024
Subtitle of host publicationTrends in Artificial Intelligence - 21st Pacific Rim International Conference on Artificial Intelligence, PRICAI 2024, Proceedings
EditorsRafik Hadfi, Takayuki Ito, Patricia Anthony, Alok Sharma, Quan Bai
PublisherSpringer Science and Business Media Deutschland GmbH
Pages383-395
Number of pages13
ISBN (Print)9789819601158
DOIs
Publication statusPublished - 2025
Event21st Pacific Rim International Conference on Artificial Intelligence, PRICAI 2024 - Kyoto, Japan
Duration: 18 Nov 202424 Nov 2024

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume15281 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference21st Pacific Rim International Conference on Artificial Intelligence, PRICAI 2024
Country/TerritoryJapan
CityKyoto
Period18/11/2424/11/24

Keywords

  • Aspect term extraction
  • Dynamic attention mechanism
  • Graph convolutional network

Fingerprint

Dive into the research topics of 'Aspect Term Extraction via Dynamic Attention and a Densely Connected Graph Convolutional Network'. Together they form a unique fingerprint.

Cite this

Sun, X., Mi, Y., Liu, J., & Li, H. (2025). Aspect Term Extraction via Dynamic Attention and a Densely Connected Graph Convolutional Network. In R. Hadfi, T. Ito, P. Anthony, A. Sharma, & Q. Bai (Eds.), PRICAI 2024: Trends in Artificial Intelligence - 21st Pacific Rim International Conference on Artificial Intelligence, PRICAI 2024, Proceedings (pp. 383-395). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 15281 LNAI). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-981-96-0116-5_32