A Dictionary-Based Approach to Identifying Aspects Implied by Adjectives for Opinion Mining

Geli Fei, Bing Liu, Meichun Hsu, Malú Castellanos, Riddhiman GhoshCOLING 2012

Introduction

Opinion mining often focuses on extracting "aspects" or features of an entity. While explicit aspects—nouns directly mentioned in text (e.g., "The battery is long-lasting")—are relatively straightforward to identify, implicit aspects pose a significant challenge. For instance, in the sentence "This camera is heavy," the user is critiquing the aspect weight, though the word itself is absent.

Traditional corpus-based approaches to implicit aspect extraction rely on co-occurrence statistics between adjectives and nouns. However, these methods frequently fail due to unnatural phrasing (people rarely say "the price is expensive"), data sparsity for rare adjectives, and the limited scope of any single corpus. This paper presents a novel dictionary-based framework that leverages the structured definitions found in online dictionaries to map adjectives to their implied attributes.

Methodology

The proposed framework operates in three stages:

  1. Extraction: Glosses (definitions) for opinion adjectives are collected from five major online dictionaries, including Merriam-Webster and Dictionary.com.
  2. Candidate Generation: Nouns are extracted from these definitions using Part-of-Speech (POS) tagging to serve as potential attribute candidates.
  3. Collective Classification: Rather than treating each adjective-noun pair independently, the system employs an Iterative Classification Algorithm (ICA). This approach exploits the "social" connections between words—such as synonyms and antonyms—to propagate labels across a relational graph.

The model incorporates three feature types:

  • Local Features: Standard metrics like n-gram frequency within definitions.
  • Static Relational Features: Lexical relationships (e.g., synonyms of a noun appearing in an adjective's definition).
  • Dynamic Relational Features: Features that evolve during the iterative process based on the current labels of neighboring words in the graph.

Experimental Setup

The researchers utilized a subset of 310 adjectives from the Hu and Liu (2004) sentiment lexicon. A gold standard dataset of 4,410 adjective-noun pairs was manually annotated for evaluation. The performance was compared against a baseline using WordNet and various supervised learning configurations.

Results and Discussion

The dictionary-based approach significantly outperformed traditional methods. A key finding was the effectiveness of dictionaries as "gold mines" for this task; lexicographers naturally define adjectives using their core attributes (e.g., defining expensive via the word "price").

Algorithm Performance

The evaluation demonstrated that Logistic Regression (LR) generally provided better results than Support Vector Machines (SVM) for this classification task. Furthermore, the transition from traditional supervised learning (Strategy 1) to collective classification using ICA (Strategy 2) yielded notable improvements, particularly when incorporating dynamic features.

Performance comparison of different feature sets and classification strategies

Comparison with WordNet

While WordNet provides high-precision mappings, its coverage is extremely limited for this specific task. The proposed method identified nearly ten times as many correct attributes as the WordNet baseline, achieving a vastly superior F-score of 0.768 compared to WordNet's 0.146.

Comparison of the proposed method against the WordNet baseline

Conclusion

This research highlights the power of dictionary-based collective classification for identifying implicit aspects. By moving beyond simple corpus statistics and leveraging the inherent relational structure of language, the method provides a robust and universal mapping for opinion mining. Future work may extend this approach to other parts of speech, such as verbs and adverbs, or explore hybrid models that combine dictionary breadth with corpus-specific context.

Summary updated

BibTeX
@inproceedings{fei2012dictionary,
  author = {Geli Fei and Bing Liu and Meichun Hsu and Malú Castellanos and Riddhiman Ghosh},
  title = {{A Dictionary-Based Approach to Identifying Aspects Implied by Adjectives for Opinion Mining}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2012},
}
Figure