Open-Domain Targeted Sentiment Analysis via Span-Based Extraction and Classification

Abstract: Open-domain targeted sentiment analysis aims to detect opinion targets along with their sentiment polarities from a sentence. Prior work typically formulates this task as a sequence tagging problem. However, such formulation suffers from problems such as huge search space and sentiment inconsistency. To address these problems, we propose a span-based extract-then-classify framework, where multiple opinion targets are directly extracted from the sentence under the supervision of target span boundaries, and corresponding polarities are then classified using their span representations. We further investigate three approaches under this framework, namely the pipeline, joint, and collapsed models. Experiments on three benchmark datasets show that our approach consistently outperforms the sequence tagging baseline. Moreover, we find that the pipeline model achieves the best performance compared with the other two models.

• The paper presents a span-based extract-then-classify framework that narrows the search space for opinion targets compared to traditional sequence tagging.
• It employs a multi-target extractor and a polarity classifier using span representations, leveraging BERT's contextual encoding for more consistent sentiment determination.
• Experiments on LAPTOP, REST, and TWITTER datasets demonstrate superior performance and guide future research in scalable sentiment analysis.

Open-Domain Targeted Sentiment Analysis via Span-Based Extraction and Classification

The paper "Open-Domain Targeted Sentiment Analysis via Span-Based Extraction and Classification" addresses the challenge of extracting opinion targets and determining their sentiment polarities within sentences. Traditionally, this task has been approached through sequence tagging methods which have limitations like a large search space and the risk of sentiment inconsistency across target words. The work introduces a novel span-based framework, proposing that opinion targets be extracted and classified based on their span boundaries.

Span-Based Extract-Then-Classify Framework

The authors propose a span-based extract-then-classify approach, which contrasts with the prevalent sequence tagging methods. This framework consists of two primary components: a multi-target extractor and a polarity classifier.

1. Multi-Target Extractor: Instead of tagging sequences, this method identifies the start and end positions of opinion targets within sentences. This change significantly narrows the computational search space, making it linearly proportional to sentence length rather than exponentially increasing with target length. An algorithm named heuristic multi-span decoding assists in effectively extracting multiple targets.
2. Polarity Classifier: For sentiment classification, the framework employs span-based representations rather than isolated word tags. This design decision ensures that polarity determination considers all target words, thereby improving sentiment consistency.

The BERT transformer serves as the backbone architecture for this framework, leveraging its contextual encoding capabilities. The study conducts experiments using BERT exclusively for both span extraction and sentiment classification tasks.

Experimental Evaluation

The framework was evaluated on three benchmark datasets: LAPTOP, REST, and TWITTER. Across each dataset, the authors compared the novel span-based approach to traditional sequence tagging baselines—both utilizing BERT—and a state-of-the-art unified model. The span-based method consistently outperformed others. Notably, the span-based pipeline model provided superior performance over the joint and collapsed model variants, suggesting limited interaction between the extraction and classification tasks that necessitates independent modeling.

Target Extraction and Polarity Classification Insights

Target extraction results exhibited the span-based approach's robustness, especially on longer sentences where traditional methods faltered due to increased search space. Although tagging methods showed higher accuracy on shorter sequences, they struggled with sentiment consistency. The span-based method's strength lies in its ability to construct comprehensive target representations before performing classification, duly noted in polarity classification performance.

Implications and Future Research

The efficacy of the span-based extract-then-classify method offers impactful implications for sentiment analysis. It reduces modeling complexity for long sentences while ensuring sentiment consistency, which is critical for tasks where opinions on multiple entities are parsed and analyzed. This provides a guideline for further research in model architectures, including potential expansions into multilingual sentiment tasks and diverse opinion mining contexts.

This paper propels the sentiment analysis landscape by asserting that the transition from token-centric to span-centric sentiment frameworks addresses several core challenges in the domain, providing a scalable and adaptable solution moving forward. Future developments could explore optimization techniques within the decoding algorithms, the effectiveness of different pre-trained models beyond BERT, and extensions to broader natural language processing tasks involving complex sentiment structures.