Fully Automatic Lexicon Expansion for Domain-oriented Sentiment Analysis

Hiroshi Kanayama, Tetsuya NasukawaEMNLP 2006

Abstract

This paper introduces a fully unsupervised method for expanding sentiment lexicons to handle domain-specific, factual expressions (e.g., "the contrast is sharp"). By leveraging context coherency—the tendency for sentiment to remain consistent across adjacent sentences—the system automatically learns new "polar atoms" from unannotated text with over 94% precision. This approach eliminates the need for manual threshold tuning and adapts robustly to diverse datasets.

Introduction

Standard Sentiment Analysis (SA) often focuses on subjective adjectives like "excellent" or "terrible." However, in specific domains like electronics or automotive reviews, sentiment is frequently conveyed through objective, factual statements. For example, "This camera has a zoom lens" is a positive attribute in context, but "zoom lens" is rarely found in a general-purpose sentiment dictionary. These "polar clauses" are highly domain-dependent and too numerous to catalog manually for every possible industry. Previous unsupervised methods often required researchers to manually set frequency thresholds to filter out noise, a process that does not scale well across diverse datasets.

Methodology

The proposed system learns polar atoms—the smallest syntactic units (usually a predicate and its argument) that carry sentiment. The process follows three main stages:

1. Proposition Detection

Using deep syntactic parsing, the system identifies propositions (clauses) and applies "modality patterns" to handle negation (e.g., "is not beautiful") and auxiliary verbs. This ensures that the core sentiment-bearing unit is correctly isolated before polarity assignment.

2. Context Coherency

The core heuristic is that writers generally maintain the same sentiment polarity in a sequence of ideas unless they use an "adversative" connector like "but" or "however." The authors define two types of context to exploit this:

  • Intra-sentential: Between clauses in the same sentence (joined by "and" or "because").
  • Inter-sentential: Between the main clauses of adjacent sentences.
Figure 3: Visualization of intra-sentential and inter-sentential context coherency used to propagate sentiment polarities.

3. Statistical Expansion

The system starts with a small, domain-independent lexicon. It identifies known polar clauses and then examines unknown clauses adjacent to them. It counts how often an unknown "candidate atom" appears in a positive vs. negative context. Rather than using a fixed threshold, the system employs a binomial distribution to statistically estimate if an atom is truly polar with 90% confidence, making the model robust to varying noise levels in different corpora.

Experimental Setup

The method was evaluated using four Japanese-language corpora harvested from online discussion boards:

  • Digital Cameras: ~1.7 million sentences.
  • Movies: ~637k sentences.
  • Mobile Phones: ~609k sentences.
  • Cars: ~959k sentences.

The system was initialized with a base lexicon of 3,275 domain-independent polar atoms.

Key Results

The system achieved remarkable performance across all tested domains. The "Token Precision" (the accuracy of sentiment detection in actual use) averaged 94%. In the digital camera domain, the expanded lexicon increased the number of detected polar clauses by 28% (a relative recall of 1.28).

Figure 5: Precision-recall curve showing the performance of the proposed method compared to baseline methods across different domains.

Furthermore, the system proved robust to the size of the initial seed lexicon. When starting with only 20% of the original seed words, the system was able to "bootstrap" itself and recover many missing atoms while maintaining high precision. Human evaluators agreed with the system's learned polarities in 89% of cases (Kappa = 0.83), confirming that the learned atoms were intrinsically meaningful.

Discussion and Insights

  • Factual is Polar: The paper demonstrates that describing a desirable feature is a form of positive sentiment, shifting focus from purely subjective adjectives to descriptive factual statements.
  • The "But" Filter: By explicitly modeling adversative conjunctions, the system accurately identifies sentiment flips, preventing the "pollution" of the learned lexicon.
  • Threshold-Free Learning: Using statistical confidence intervals allows the model to adapt to the noise level of the data automatically, making it truly "fully automatic."
  • Syntactic Scoping: The "polar atom" structure (e.g., "to have [a zoom lens]") captures nuances that simple verb-only models miss.

Conclusion

While the current implementation is specific to Japanese syntax and requires high-quality dependency parsing, the logic of context coherency is universal. Future work may address subtle sentiment like sarcasm or complex rhetorical structures. This method offers a powerful tool for "on-demand" analysis of narrow domains where no pre-existing sentiment lexicon exists.

Summary updated

BibTeX
@inproceedings{kanayama2006fully,
  author = {Hiroshi Kanayama and Tetsuya Nasukawa},
  title = {{Fully Automatic Lexicon Expansion for Domain-oriented Sentiment Analysis}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2006},
}
Figure