📄 OpinionFinder: A System for Subjectivity Analysis

Summary

OpinionFinder is a demonstration system that automatically detects subjective language in English text, meaning opinions, emotions, speculations and other private states. Its purpose is to act as a support component for downstream NLP tasks. Examples include opinion-oriented question answering, summarizing different viewpoints, mining product reviews, and filtering opinionated sentences out of information extraction. The system is a single pipeline. A general preprocessing stage runs first: Sundance partial parsing, OpenNLP tokenization, sentence splitting and POS tagging, stemming, optional dependency parsing, and lexicon clue matching. A subjectivity stage follows, built from four components:

• sentence-level subjective/objective classification,
• detection of speech events and direct subjective expressions,
• identification of opinion sources (holders),
• marking of words inside positive or negative sentiment phrases.

Output is delivered as SGML/XML annotations layered over the original documents. The system works in a batch mode or an interactive mode that queries online news. The contribution is an integrated, reusable tool rather than a new benchmark result. The paper itself reports no evaluation numbers and points readers to the component papers.

Key points

• The subjective sentence classifier is Naive Bayes. Its training sentences come from unannotated text that was automatically labeled by two high-precision rule-based classifiers.
• Opinion source identification combines a Conditional Random Field tagger with extraction pattern learning. It is trained on the MPQA Opinion Corpus.
• Sentiment expression detection uses two BoosTexter classifiers trained on the MPQA Corpus. One finds sentiment expressions, and the other separates positive from negative ones.
• Preprocessing tools:
  • Sundance for semantic classes, named entities and subjective extraction patterns.
  • OpenNLP 1.1.0 for tokenization, sentence splitting and POS tagging.
  • The Abney stemmer.
  • Collins constituency parses converted to dependency trees (Xia and Palmer).
• Dependency parsing is slow, so it runs only in batch mode. The source identifier depends on it, so source identification is also batch-only.

Interesting ideas

• I1. Bootstrapping training data with precise rules. High-precision rule-based classifiers can label raw text to create training data for a statistical classifier. This could help bootstrap subjectivity or other classifiers for low-resource languages such as Ukrainian, where annotated corpora are scarce.
• I2. Annotations over the original text. Results are returned as markup on the source documents rather than in a separate format. This keeps the analysis traceable and easy for other tools to consume.
• I3. Speed-dependent feature tiers. Expensive components (full parsing, and anything that depends on it) are enabled only in offline batch runs, while the interactive mode uses a lighter pipeline. This is a practical way to trade accuracy for latency.
• I4. Sequence labeling plus pattern learning. Opinion holders are found by combining a CRF tagger with learned extraction patterns. Mixing statistical tagging with pattern knowledge could transfer to other span-finding tasks, such as identifying who made a statement in news or trade reports.
• I5. Subjectivity filtering as a preprocessing step. Removing opinionated sentences can also benefit fact-oriented pipelines like information extraction, which makes subjectivity detection useful beyond sentiment applications.