2022,ACL, Enhanced Multi-Channel Graph Convolutional Network for Aspect Sentiment Triplet Extraction

[Sepideh-Ahmadian]

Paper
Enhanced Multi-Channel Graph Convolutional Network for Aspect Sentiment Triplet Extraction

Introduction
The paper introduces a method that uses the Graph Convolutional Network for Aspect Sentiment Triplet Extraction (ASTE). ASTE is a task within ABSA that involves extracting aspect terms, opinion terms, and their associated sentiment polarity in the form of triplets. The proposed model utilizes a graph-based approach to encode relations between words and leverage linguistic features to improve sentiment triplet extraction​.

Main Problem
The paper addresses the challenge of accurately extracting aspect-opinion-sentiment triplets in the ASTE task. The authors aim to utilize the relations between words and linguistic features to improve the performance of the extraction process​.

Illustrative Example
An example given in the paper illustrates the ASTE task:
Sentence: "The gourmet food was delicious."
Extracted Triplet: (Aspect: gourmet food, Opinion: delicious, Sentiment: positive)

Input
A sentence consisting of words and their relations, including aspect terms, opinion terms, and sentiment polarity.

Output
A set of aspect-opinion-sentiment triplets extracted from the input sentence​.

Motivation
The authors were motivated by the limitations of existing methods, particularly the lack of attention to word relations and linguistic features. Many approaches treat words independently, resulting in low performance.

Related works and their gaps
The paper fills the gap in previous works that do not fully consider word relations and linguistic features when extracting aspect sentiment triplets. Many existing methods either treat words in isolation or use simplistic approaches that overlook these critical components​.
The previous works are categorized into following categories:
Pipeline approaches (Peng et al., 2020): they independently extract elements of the triples. This independence ignores the relation between the words.
Relation-based works (Mao et al. (2021) and Chen et al. (2021a))
End-to-end framework (Xu et al., 2020; Wu et al., 2020a; Zhang et al., 2020; Chen et al., 2021b; Yan et al., 2021) by designing new tagging system.

Contribution of this paper
Utilizing various linguistic features (e.g., part-of-speech, syntactic dependency, and distance measures) to enhance the model.

Proposed methods
Not included

Experiments
The model was trained and evaluated on the following datasets:
SemEval 2014 Restaurant: 1,259 training sentences, 493 test sentences.
SemEval 2014 Laptop: 899 training sentences, 332 test sentences.
SemEval 2015 and 2016 Restaurant: Around 600 training sentences each, with around 320 test sentences.

Implementation
https://github. com/CCChenhao997/EMCGCN-ASTE.

Gaps this work
The gaps in this study include the potential dependency on the quality of linguistic features, which may affect performance. Additionally, while the model performs well in the datasets used, further evaluation on more diverse domains and datasets could be useful to assess generalizability​.

[hosseinfani]

@Sepideh-Ahmadian
do you think it's worth adding to LADy as a baseline model? We have done it for other neural models, but only using the aspect result of such models.

Actually, now I remembered a research direction I had, which was studying the effect of backtranslation on opinion or even sentiment detection ...