The utility of linguistic rules in opinion mining

Xiaowen Ding, Bing LiuSIGIR 2007 · DOI

TL;DR

This paper introduces a novel approach to opinion mining that goes beyond simple keyword matching by incorporating linguistic rules to understand the context-dependent sentiment of opinions expressed about product features. Their system, "Opinion Observer," significantly improves accuracy by considering how words like "small" or "long" can be positive or negative depending on the product feature they describe, outperforming existing methods.

What Problem Does This Paper Address?

Online product reviews are a rich source of information, but extracting meaningful insights about user opinions on specific product features is challenging. Traditional opinion mining methods often rely on pre-defined lists of positive and negative words. However, the sentiment of many words is highly context-dependent. For instance, "small" might be a positive attribute for a phone's portability but a negative one for a TV's screen size. This paper tackles the problem of accurately determining the semantic orientation (positive or negative) of opinions expressed on product features, acknowledging that simple keyword matching is insufficient. They assume product features are already identified and focus solely on sentiment analysis.

Methods

The core of the proposed technique lies in two main components: an improved opinion aggregation function and a set of linguistic rules.

  • Opinion Aggregation Function: Instead of a simple summation of opinion word scores, their function uses a weighted sum where the weight is inversely proportional to the distance between the opinion word and the product feature. This acknowledges that opinion words further away from a feature might have less influence. They also handle cases where the product feature itself acts as an opinion word (e.g., "reliable camera").
  • Linguistic Rules: The paper proposes several rules to infer sentiment based on linguistic context:
  • Intra-sentence Conjunction Rule: Assumes a sentence expresses a single opinion orientation unless a conjunction like "but" signals a change.
  • Pseudo Intra-sentence Conjunction Rule: Infers sentiment by looking for positive or negative cues in the same sentence, even without explicit conjunctions.
  • Inter-sentence Conjunction Rule: Extends the conjunction rule to neighboring sentences, assuming consistent sentiment unless explicitly contradicted.
  • Synonym and Antonym Rule: Leverages known sentiment of a word to infer the sentiment of its synonyms and antonyms.
  • Special Word Handling: Rules for words like "not," "too," etc., are also incorporated.
  • Majority View: For conflicting opinions on the same feature, the algorithm defaults to the majority sentiment.

The system operates iteratively, allowing inferred opinions to help infer others.

Datasets

The experiments were conducted using 742 customer reviews from 14 different products across categories like digital cameras, MP3 players, cellular phones, and routers, all extracted from amazon.com. For comparison, they also used a subset of reviews from a benchmark dataset provided in reference [2].

Key Insights and Ideas

  • Context is King: The most significant insight is that linguistic context and the specific product feature are crucial for accurate sentiment analysis, far more so than just relying on a lexicon of opinion words.
  • Beyond Simple Keywords: The paper demonstrates that sophisticated linguistic rules can effectively capture nuances that simple keyword spotting misses, leading to substantial performance gains.
  • Feature-Opinion Interaction: The approach highlights the importance of understanding the relationship between an opinion word and the specific feature it modifies, rather than treating opinion words in isolation.
  • Iterative Inference: The iterative nature of their algorithm allows for a cascading effect, where initial sentiment inferences can bootstrap further, more complex deductions.

Key Results

  • Significant Performance Improvement: The proposed linguistic rules dramatically improved recall without a significant loss in precision, leading to substantial gains in F-scores compared to baselines that only used opinion words. For example, with 200 opinion words, recall increased from 0.65 to 0.73, and F-score from 0.72 to 0.76.
  • Outperforming Existing Systems: The "Opinion Observer" system outperformed two established systems, FBS and OPINE, on a benchmark dataset, achieving a higher F-score (0.91 vs. 0.83 for FBS and 0.87 for OPINE).
  • Balanced Contribution: Both the new opinion aggregation function and the linguistic rules contributed roughly equally to the overall performance improvement.

Limitations and Future Work

  • Feature Identification: The paper assumes that product features are already identified by another system. The process of automatically discovering these features is not addressed.
  • Ambiguity Resolution: While the majority view is used for conflicting opinions, more sophisticated methods for resolving such ambiguities could be explored.
  • Domain Specificity: While the rules are designed to be general, their effectiveness might vary across different domains, and further domain-specific tuning or rule discovery could be beneficial.
  • Implicit Opinions: The paper focuses on explicit opinions. Future work could explore how to detect more subtle or implicit opinions.

Summary updated

BibTeX
@inproceedings{ding2007utility,
  author = {Xiaowen Ding and Bing Liu},
  title = {{The utility of linguistic rules in opinion mining}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2007},
  doi = {10.1145/1277741.1277921},
}
Figure