A system for fine-grained aspect-based sentiment analysis of Chinese

Janna LipenkovaACL 2015

TL;DR

This paper introduces a novel pipeline for fine-grained aspect-based sentiment analysis (ABSA) of Chinese automotive reviews. By integrating linguistic knowledge with a rule-based system, it achieves higher precision and granularity in identifying opinions about specific product features compared to purely probabilistic approaches.

What Problem Does This Paper Address?

Traditional sentiment analysis often focuses on the overall sentiment of a text, which can be too coarse-grained. Aspect-Based Sentiment Analysis (ABSA) aims to pinpoint opinions about specific features of a product or entity. However, applying ABSA to languages like Chinese presents challenges due to its unique linguistic structure and the scarcity of specialized resources. Existing Chinese ABSA approaches often rely heavily on probabilistic models and general sentiment lexicons, potentially missing nuanced relationships between opinions and their targets. This paper addresses this gap by proposing a system that leverages linguistic rules and a structured approach to capture these fine-grained relationships more effectively.

Methods

The proposed system employs a two-step approach:

  1. Unit Identification: This step identifies linguistic units that express features, evaluations, and emotions. It uses a customized word segmentation and POS tagging tool (jieba) that incorporates domain-specific lexicons. This allows for more accurate identification of features (e.g., "rear space") and evaluations (e.g., "not bad"). The system then uses a phrase rule grammar to identify phrasal units corresponding to these identified units.
  2. Relation Extraction: This step identifies the relationships between the identified units, specifically focusing on "causer" (emotions caused by features) and "theme" (features being evaluated or described by emotions) relations. Instead of relying on full syntactic parsing, the system uses a sequence classifier trained on common patterns observed in the data to determine these relations, especially when words separate the opinion and its target.

A key innovation is the use of a "bottom-up" approach that progresses through lexical, phrasal, and sentence levels, integrating linguistic knowledge to build opinion representations. Evaluations are represented as pairs of a feature and a numerical value on a scale, allowing for a unified treatment of implicit and explicit features.

Datasets

The system was trained on a corpus of approximately 1 million automotive reviews, with an average length of 135 Chinese characters. The evaluation was performed on a test set of 800 annotated sentences. The annotation process involved two independent annotators and a third member for resolving disagreements, ensuring high inter-annotator agreement for both unit identification and relation extraction.

Key Insights and Ideas

  • Linguistic knowledge enhances ABSA: Integrating general linguistic rules and knowledge about Chinese sentence structure significantly improves the precision and granularity of sentiment analysis, going beyond purely probabilistic methods.
  • Hierarchical feature representation: The system's ability to represent features hierarchically (e.g., "Indicator" as a sub-feature of "Steering Wheel") allows for a more nuanced understanding of product aspects.
  • Unified treatment of evaluations and emotions: By representing both evaluations and emotions with a numerical value and a concept, the system can handle them more uniformly, simplifying the analysis.
  • Rule-based approach for relation extraction: Using a sequence classifier for relation extraction, rather than full parsing, offers a balance between accuracy and computational efficiency, especially for complex sentence structures.

Key Results

The paper reports strong performance on both unit identification and relation extraction:

  • Unit Identification: Achieved an F-score of 87.43% for feature phrases, 89.21% for evaluation phrases, and 88.56% for emotion phrases.
  • Relation Extraction: Demonstrated high precision and recall for identifying relationships between features and evaluations/emotions, with F-EV relations achieving 89.2% precision and 87.33% recall, and F-EM relations (theme and causer) also showing strong results.

Limitations and Future Work

The authors acknowledge that while the system achieves considerable accuracy, it is currently extended with more complex sentence-level relations, specifically comparative structures and modal operators. Future work will focus on incorporating these more advanced linguistic phenomena into the system.

Summary updated

BibTeX
@inproceedings{lipenkova2015system,
  author = {Janna Lipenkova},
  title = {{A system for fine-grained aspect-based sentiment analysis of Chinese}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2015},
}
Figure