Opinion Holder and Target Extraction on Opinion Compounds – A Linguistic Approach

Michael Wiegand, Christine Bocionek, Josef RuppenhoferNAACL 2016

TL;DR

This paper introduces a novel approach to extracting opinion holders and targets from opinion compounds (e.g., "user rating"). They propose new linguistic features that go beyond traditional noun compound analysis, achieving state-of-the-art results on this challenging task.

What Problem Does This Paper Address?

Sentiment analysis aims to understand opinions expressed in text. A crucial subtask is opinion role extraction, which involves identifying who is expressing an opinion (opinion holder, OH) and what the opinion is directed towards (opinion target, OT). While this has been studied for opinion verbs and adjectives, opinion compounds (noun phrases like "user rating") present a unique challenge. In these compounds, the relationship between the opinion word (the head noun, e.g., "rating") and its modifier (e.g., "user") is implicit, lacking the explicit syntactic cues found in sentences. This makes it difficult to automatically determine if the modifier refers to the holder or the target of the opinion. Existing methods for noun compound analysis are insufficient, and a dedicated approach is needed.

Methods

The authors tackle this problem by framing it as a supervised classification task, employing Markov Logic Networks (MLNs). MLNs are powerful because they can integrate various features, including global constraints. They propose a multi-stage pipeline:

  1. Initial Role Detection (Dataset I): Distinguish compounds with any opinion role from those without.
  2. Holder/Target Detection (Dataset II): For compounds with opinion roles, further classify the modifier as a holder, target, or neither. This stage is further refined by first checking if the modifier denotes a person, as this simplifies the classification.

Their key contribution lies in the novel features they introduce:

  • Verb Detour Paraphrases: Instead of directly paraphrasing the noun compound, they replace the head noun with its related verb. This allows them to leverage syntactic relationships (subject/object) between the modifier and the verb to infer the role. This is particularly useful for deverbal nouns, which form a significant portion of their data.
  • Joint Paraphrases: These paraphrases consider both the holder and target simultaneously, using dependency relations to disambiguate roles when one relation might be ambiguous.
  • Global Head Constraint: This feature leverages the observation that compounds sharing the same head noun often exhibit similar opinion roles for their modifiers. MLNs are used to enforce this constraint globally across the dataset.
  • Sentiment Views (VIEW): They categorize opinion nouns (heads) based on whether they convey a "speaker view" (implicit holder is the speaker, modifier is likely a target) or an "actor view" (modifier can be a holder or target). This is crucial for distinguishing roles in Dataset II.
  • Subjectivity Disambiguation (SUBJ): A feature to identify compounds that are not subjective, helping to filter out non-opinionated compounds.

Datasets

The authors created a new gold standard dataset by extracting opinion compounds from the deWaC corpus. This corpus is large (1.7 billion words) and provides a rich source of German noun compounds. They then manually annotated two datasets:

  • Dataset I: Used for initial opinion role detection.
  • Dataset II: Used for distinguishing between holders and targets, focusing on compounds where the modifier is a person.

They also mention the STEPS-corpus as a benchmark for German opinion role extraction, noting that their new dataset contains significantly more unique compounds.

Key Insights and Ideas

  • Implicit Relations are Key: The core challenge is the implicit nature of opinion roles within compounds. Traditional syntactic features are less effective, necessitating new linguistic insights.
  • Verb Semantics Unlock Roles: The "verb detour" is a clever way to expose the underlying semantic roles by transforming the noun compound into a verb-argument structure, revealing subject/object relationships.
  • Global Constraints Matter: Enforcing consistency across compounds with the same head noun (the HEAD constraint) significantly improves performance, highlighting the interconnectedness of linguistic structures.
  • Head Noun Type is Informative: The "sentiment view" of the head noun (speaker vs. actor) provides strong cues for determining the role of the modifier, especially in distinguishing holders from targets.
  • Complementarity of Features: Combining various feature types (paraphrases, semantic knowledge, global constraints) leads to substantial improvements, indicating that different features capture complementary aspects of the problem.

Key Results

The proposed approach significantly outperforms baseline methods. For instance, their best model achieved an F-score of 66.42% for role detection, 64.71% for holder detection, and 66.50% for target detection, substantially outperforming majority and distant supervision baselines. The introduction of novel features like verb detour paraphrases and the head constraint led to notable improvements over simpler paraphrase-based methods.

Limitations and Future Work

  • Ambiguity in Reciprocal Relationships: The paper acknowledges that in 3% of compounds, the modifier can represent both a holder and a target (e.g., "sibling jealousy"). This inherent ambiguity remains a challenge.
  • Language Specificity: While the authors claim their features are not language-specific, their work is primarily demonstrated on German. Further validation on other languages would be beneficial.
  • Future Work: The authors suggest exploring more sophisticated methods for paraphrase acquisition and further investigating the role of semantic knowledge. They also hint at the potential for cross-lingual applications.

Summary updated

BibTeX
@inproceedings{wiegand2016opinion,
  author = {Michael Wiegand and Christine Bocionek and Josef Ruppenhofer},
  title = {{Opinion Holder and Target Extraction on Opinion Compounds – A Linguistic Approach}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2016},
}
Figure