[論文レビュー] Dual Encoder: Exploiting the Potential of Syntactic and Semantic for Aspect Sentiment Triplet Extraction

Xiaowei Zhao, Yong Zhou+1 more|arXiv (Cornell University)|Feb 23, 2024

Sentiment Analysis and Opinion Mining被引用数 5

ひとこと要約

D2E2Sはデュアルエンコーダを導入し（意味論のためのBERT、拡張LSTMによる構文）と、異種特徴相互作用モジュールを備えたASTEの性能を向上させ、4つのベンチマークでSOTAを超える。

ABSTRACT

Aspect Sentiment Triple Extraction (ASTE) is an emerging task in fine-grained sentiment analysis. Recent studies have employed Graph Neural Networks (GNN) to model the syntax-semantic relationships inherent in triplet elements. However, they have yet to fully tap into the vast potential of syntactic and semantic information within the ASTE task. In this work, we propose a \emph{Dual Encoder: Exploiting the potential of Syntactic and Semantic} model (D2E2S), which maximizes the syntactic and semantic relationships among words. Specifically, our model utilizes a dual-channel encoder with a BERT channel to capture semantic information, and an enhanced LSTM channel for comprehensive syntactic information capture. Subsequently, we introduce the heterogeneous feature interaction module to capture intricate interactions between dependency syntax and attention semantics, and to dynamically select vital nodes. We leverage the synergy of these modules to harness the significant potential of syntactic and semantic information in ASTE tasks. Testing on public benchmarks, our D2E2S model surpasses the current state-of-the-art(SOTA), demonstrating its effectiveness.

研究の動機と目的

• 構文情報と意味情報を統合することで恩恵を受ける微細な感情抽出タスクとしてASTEを動機付ける。
• 意味論（BERT）と統語論（enhanced LSTM）特徴を別々に捉えるデュアルエンコーダアーキテクチャを提案する。
• 異種特徴相互作用モジュール（HFIM）を開発し、SADPool、GCNConv、GatedGraphConvを組み合わせて複雑な相互作用をモデル化する。
• 特徴干渉を減らすために構文的および意味的類似性を分離する方法を導入する。
• 複数のASTEベンチマークで最先端の性能を示し、オープンソースのコードを提供する。

提案手法

• Dual encoders: a BERT-based channel for semantic relations and an enhanced LSTM channel for syntactic/ local dependencies.
• Syntactic graph (SynGCN) built from dependency parses; semantic graph (SemGCN) built from multi-head attention induced scores.
• SADPool-based heterogeneous feature interaction with multi-layer GCNConv and GatedGraphConv to propagate and fuse features.
• KL-divergence-based loss terms to separate syntactic and semantic adjacency distributions.
• Span representation and filtering to generate candidate aspect and opinion terms, followed by a sentiment classifier over span pairs.
• Joint optimization with L_sp (span prediction), L_tri (triplet sentiment), and L_kl (syntax-semantics separation) as the total loss.

実験結果

リサーチクエスチョン

• RQ1 Can a dual-encoder architecture effectively exploit both syntactic and semantic information for ASTE?
• RQ2 Does the heterogeneous feature interaction module improve the integration of syntactic and semantic signals compared to prior methods?
• RQ3 How does explicit separation of syntactic and semantic similarities impact ASTE performance?
• RQ4 What gains are achievable on standard ASTE benchmarks using the proposed D2E2S framework?

主な発見

• D2E2S achieves state-of-the-art results on four ASTE benchmarks (14lap, 14res, 15res, 16res).
• Ablation studies show that removing syntactic or semantic features degrades performance, confirming their complementary roles.
• Removing HFIM or replacing it with Mutual BiAffine worsens results, underscoring the benefit of the proposed interaction module.
• Using dual encoders (BERT for semantic, enhanced LSTM for syntax) and the separation strategy yields better performance than single-encoder variants.
• Case study indicates D2E2S can handle ultra-long distance modeling and complex relations better than some baselines.