Automatic Extraction of Polar Adjectives for the Creation of Polarity Lexicons

Silvia Vázquez, Muntsa Padró, Núria Bel, Julio GonzaloCOLING 2012

TL;DR

This paper introduces a novel bootstrapping algorithm to automatically discover, extract, and label subjective adjectives for building sentiment analysis lexicons. By leveraging linguistic cues and a new category of "highly subjective" adjectives, the method significantly improves the precision of polarity lexicon creation, especially for context-dependent terms.

What Problem Does This Paper Address?

Sentiment Analysis, a crucial area in Natural Language Processing, heavily relies on understanding the polarity (positive, negative, or neutral) of words. Manually creating comprehensive polarity lexicons is a time-consuming and labor-intensive process. Existing methods, while useful, often struggle with domain-specific sentiment (where a word's polarity can change depending on the context, e.g., "entertaining" in a movie review vs. a car review) and the nuances of "highly subjective" adjectives that can flip polarity even within the same domain (e.g., "big" car can be good or bad). This paper aims to automate the creation of these lexicons, addressing the need for efficient, scalable, and context-aware sentiment analysis tools.

Methods

The authors propose a bootstrapping algorithm that iteratively expands a seed list of polar adjectives. The core idea is based on the hypothesis that:

  • Adjectives joined by "and" (or "y" in Spanish) tend to share the same semantic orientation.
  • Adjectives joined by "but" (or "pero" in Spanish) tend to have opposite semantic orientations.

The algorithm starts with a small set of domain-independent seed adjectives. It then searches a corpus for adjectives conjoined with these seeds. Based on the conjunction type, it infers the polarity of new adjectives and adds them to the lexicon. This process is repeated, using the newly identified adjectives as seeds, to iteratively discover more polar terms.

A key innovation is the introduction of a "highly subjective adjective" category. Adjectives that appear in contexts implying both positive and negative orientations are flagged as highly subjective. This allows for a more nuanced understanding of sentiment, acknowledging that some words' polarities are not fixed.

Datasets

  • Training/Seed Data: A small set of domain-independent seed adjectives, manually labeled by five annotators as reliable polar words. These were taken from a previous work by Bel and Vázquez (2012).
  • Corpus for Experimentation:
    • A primary corpus of 250,000 words from car reviews (extracted from a larger 8-million-word corpus spanning various domains like movies, mobile phones, video games, and sports teams).
  • A Gold Standard dataset, consisting of 12% of the car corpus (200 documents), manually annotated for polar adjectives and their context-specific semantic orientation. This Gold Standard was used for evaluation.

Key Insights and Ideas

  • Context is King: The paper strongly emphasizes that adjective polarity is often context-dependent, a crucial insight for building robust sentiment analysis systems.
  • "Highly Subjective" Adjectives Matter: Identifying adjectives that can swing between positive and negative polarity is vital for achieving higher precision and avoiding misclassifications. This category captures a significant source of ambiguity in sentiment.
  • Bootstrapping for Scalability: The iterative bootstrapping approach offers a scalable way to expand lexicons beyond initial seed sets, reducing manual effort.
  • Linguistic Cues as Powerful Indicators: The reliance on simple conjunctions ("and," "but") as linguistic cues proves effective for inferring semantic orientation, demonstrating the power of structural patterns in language.
  • Slang and Informal Language: The method's ability to extract slang terms (like "cojonudo" and "fardón") highlights its potential for analyzing less formal text, as it doesn't depend on pre-existing external lexicons.

Key Results

The bootstrapping process significantly expanded the initial seed lexicon:

  • Positive adjectives grew from 28 to 173.
  • Negative adjectives grew from 7 to 37.
  • 13 highly subjective adjectives were identified.

In the evaluation using the Gold Standard:

  • Recall: The method achieved a recall of 67% of all adjectives appearing in the conjunction pairs.
  • Precision:
    • 97.6% for positive adjectives.
  • 71.5% for negative adjectives.

These precision figures are reported to be better than other published works while maintaining recall.

Limitations and Future Work

  • Limitations:
    • The method's effectiveness relies on the presence of conjunctive constructions.
  • The precision for negative adjectives (71.5%) is notably lower than for positive ones (97.6%), indicating a potential area for improvement.
  • The study was conducted primarily on Spanish text, though the authors suggest the method is applicable to other languages where similar conjunctive constructions exist.
  • The reduction in corpus size for the Gold Standard evaluation meant that no highly subjective adjectives were identified in that specific subset.
  • Future Work:
    • Extend to other word types: Adapt the bootstrapping method to extract and annotate polar nouns, verbs, and adverbs.
  • Cross-lingual and cross-domain application: Further explore the adaptability of the method to different languages and domains.
  • Improve negative adjective precision: Investigate strategies to enhance the accuracy of identifying negative polarities.

Summary updated

BibTeX
@inproceedings{vazquez2012automatic,
  author = {Silvia Vázquez and Muntsa Padró and Núria Bel and Julio Gonzalo},
  title = {{Automatic Extraction of Polar Adjectives for the Creation of Polarity Lexicons}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2012},
}
Figure