TL;DR
This paper introduces "You Only Read Once" (YORO), a novel one-to-many framework for Aspect-Based Sentiment Analysis (ABSA) that simultaneously predicts the sentiment of all aspects in a sentence. YORO utilizes a constituency-oriented graph convolutional network (CorrGCN) to capture complex relationships between aspects and their contexts, outperforming previous methods on challenging multi-aspect, multi-sentiment datasets.
What Problem Does This Paper Address?
Traditional Aspect-Based Sentiment Analysis (ABSA) models often struggle with sentences containing multiple aspects, each with its own sentiment. Existing approaches typically focus on predicting the sentiment of one aspect at a time, which can be inefficient and miss crucial inter-aspect dependencies and shared global context. This "one-to-one" paradigm fails to fully leverage the rich relational information present in complex sentences, leading to suboptimal performance, especially when multiple aspects have different sentiments. The paper aims to bridge this gap by developing a model that can process a single sentence and predict the sentiment for all its aspects simultaneously, thereby improving both efficiency and accuracy.
Methods
The core of YORO is its Constituency-Oriented Relational Graph Convolutional Network (CorrGCN). Unlike previous methods that rely on dependency parsing trees, CorrGCN leverages constituency parsing trees. These trees offer a hierarchical structure (sentence-phrase-word) that better reflects aspect-context and aspect-aspect relationships. CorrGCN models four types of relationships:
- Aspect-specific: Captures local contexts highly relevant to a specific aspect.
- Inter-aspect: Models interactions and dependencies between different aspects.
- Globally-shared: Incorporates common contextual information relevant to multiple aspects, preventing information loss.
- Opinion word: Enhances attention to opinion words, which are key sentiment cues.
To handle sentences with only one aspect, YORO employs a "pseudo-aspect" strategy. This involves treating constituents with the same label as pseudo-aspects to construct richer graph structures, enabling better modeling of relationships even in simpler sentences. Additionally, YORO incorporates opinion sentiment classification and supervised contrastive learning as auxiliary tasks to improve the model's understanding of opinion words and sentiment label relationships. The model uses BERT as an encoder and outputs sentiment predictions for all aspects simultaneously.
Datasets
The paper evaluates YORO on three benchmark datasets:
- MAMS (Multi-Aspect Multi-Sentiment): This dataset is highlighted as particularly challenging because all sentences contain multiple aspects with different sentiment polarities.
- SemEval 2014 Restaurants (Rest14): Sentences can have one or multiple aspects, with sentiments that may be the same or different.
- SemEval 2014 Laptops (Lap14): Similar to Rest14, it includes sentences with single or multiple aspects and varying sentiment polarities.
Key Insights and Ideas
- Constituency trees are superior to dependency trees for ABSA: The hierarchical structure of constituency trees provides a more natural way to capture aspect-context and inter-aspect relationships compared to dependency trees.
- One-to-many ABSA is feasible and effective: YORO demonstrates that predicting all aspect sentiments simultaneously from a single sentence is not only possible but also leads to improved performance, challenging the traditional one-to-one approach.
- Multi-source relational information is crucial: The combination of aspect-specific, inter-aspect, globally-shared, and opinion word relationships within CorrGCN is key to capturing the complex sentiment landscape of multi-aspect sentences.
- "Pseudo-aspects" effectively handle single-aspect sentences: This strategy allows for richer graph construction and relationship modeling even when only one explicit aspect is present, preventing performance degradation.
- Auxiliary tasks boost performance: Opinion sentiment classification and supervised contrastive learning play a significant role in enhancing the model's ability to understand sentiment nuances and relationships.
Key Results
YORO achieves state-of-the-art or competitive results across the evaluated datasets. Notably:
- On the MAMS dataset, YORO outperforms CHGMAN by 1.03% in accuracy and 1.22% in F1 score, and CapsNet by 2.62% in accuracy and 2.62% in F1 score.
- For Rest14, YORO shows improvements in accuracy and F1 scores over CHGMAN and APSCL.
- On Lap14, YORO achieves the best accuracy and a competitive F1 score, only slightly lower than SDGCN and APSCL.
- Ablation studies confirm the importance of each component, with the removal of CorrGCN leading to a significant performance drop, highlighting the effectiveness of the proposed graph structure.
Limitations and Future Work
The authors acknowledge that while the "pseudo aspect" strategy is effective for single-aspect sentences, there is still room for improvement. Their future work will focus on further enhancing the modeling of aspect interactions, particularly in sentences with only one aspect, to further improve efficiency.
Summary updated