Clustering for Simultaneous Extraction of Aspects and Features from Reviews

Lu Chen, Justin Martineau, Doreen Cheng, Amit ShethNAACL 2016

TL;DR

This paper introduces CAFE, a novel clustering approach that simultaneously extracts product features and groups them into aspects from online reviews. Unlike prior methods that chain these steps, CAFE integrates them, handling both explicit and implicit features without requiring seed terms, and demonstrating superior performance across multiple domains.

What Problem Does This Paper Address?

When consumers shop online, they often rely on product reviews to make informed decisions. However, sifting through hundreds of reviews for a single product can be overwhelming. To address this, researchers aim to automatically summarize reviews by identifying key product features (e.g., "battery life," "screen size") and the broader aspects they belong to (e.g., "battery," "display"). Existing methods often tackle feature extraction and aspect grouping as separate, sequential steps. Some approaches require predefined "seed terms" or hand-crafted rules, limiting their applicability. Furthermore, many methods struggle to identify "implicit features" – aspects that are not directly named but are implied by opinion words (e.g., "fast" implying "speed"). This paper tackles the challenge of simultaneously extracting both explicit and implicit features and grouping them into coherent aspects in an unsupervised manner.

Methods

The core of the proposed method, CAFE (Clustering for Aspect and Feature Extraction), is a single clustering process that identifies features and aspects concurrently. It begins by identifying candidate terms from reviews, including nouns/noun phrases (potential explicit features) and adjectives/verbs (potential implicit features). To manage computational complexity and noise, it first clusters the most frequent candidates to form "seed clusters."

A key innovation is the development of a domain-specific similarity measure. This measure combines general semantic similarity (using UMBC's Semantic Similarity Service) with corpus-based statistical association. This is crucial because word meanings and their relevance to product aspects can vary significantly across domains (e.g., "ice cream sandwich" is an aspect of a phone, but not a hair dryer). The similarity measure also incorporates a "domain-dependent synonym" component to capture terms that co-occur with similar opinion words.

The clustering process uses an agglomerative approach, merging the closest clusters based on both average candidate distance and the distance between representative terms of clusters. Crucially, CAFE incorporates problem-specific merging constraints to prevent clusters from different aspects from being merged. These constraints include a distance threshold, a requirement for at least one explicit feature in a merged cluster, and a measure of co-occurrence within sentences versus across documents. Finally, the top k clusters are selected as aspects, and nouns/noun phrases within these clusters are identified as explicit features, while adjectives/verbs are identified as implicit features.

Datasets

The experiments were conducted on datasets from three domains:

  • Cell phone reviews: Collected from an online cell phone company's website.
  • GPS reviews: Collected from Amazon.
  • TV reviews: Collected from Amazon.

The paper also mentions that the datasets and annotations are made publicly available.

Key Insights and Ideas

  • Simultaneous Extraction: The most significant insight is that combining feature extraction and aspect grouping into a single clustering process is more effective than sequential approaches. This allows for a more holistic understanding of review content.
  • Implicit Feature Handling: CAFE's ability to identify and group implicit features (like "fast" for "speed") alongside explicit ones is a major advantage, as these are often overlooked by methods focused solely on explicit mentions.
  • Domain-Specific Similarity is Crucial: The paper strongly emphasizes that general semantic similarity is insufficient for aspect extraction. A domain-specific measure that captures how words relate within a particular product context is essential for accurate grouping.
  • Frequency as a Signal: Prioritizing the clustering of frequent terms as "seed clusters" is an effective strategy to generate high-quality initial groupings and improve efficiency.
  • Constraints Improve Coherence: The use of problem-specific merging constraints is vital for ensuring that the resulting clusters represent coherent aspects and not just arbitrary groupings of words.

Key Results

  • Feature Extraction: CAFE significantly improved precision for feature extraction compared to PROP and LRTBOOT, with only a small loss in recall. This resulted in a 21.68% and 9.36% improvement in macro-averaged F-score over PROP and LRTBOOT, respectively. Precision-recall curves showed CAFE consistently outperforming the baselines.
  • Aspect Discovery: CAFE achieved the best macro-averaged Rand Index (a measure of clustering similarity to a gold standard) when compared against methods like MuReinf, L-EM, and L-LDA, especially when using CAFE's own extracted features. This indicates superior performance in grouping features into aspects.
  • Parameter Sensitivity: The study found that the performance of CAFE is sensitive to parameter choices (δ, s, and the similarity weights wg, wt, wgt). Optimal settings were identified through experiments, suggesting that a balance is needed to avoid noise from merging too many clusters.

Limitations and Future Work

  • Parameter Tuning: While CAFE performs well, its performance is influenced by parameter settings, requiring careful tuning. The paper explores this impact but acknowledges that optimal settings might still vary.
  • Aspect Selection: Currently, aspects are selected based on cluster frequency. The authors suggest that future work could leverage domain knowledge to improve aspect selection, potentially identifying less frequent but still important aspects.
  • Implicit Feature Mapping: While implicit features are identified and grouped, the paper notes that their mapping to explicit features is "accomplished naturally during the clustering process." Further research could explore more explicit ways to represent these mappings.

Summary updated

BibTeX
@inproceedings{chen2016clustering,
  author = {Lu Chen and Justin Martineau and Doreen Cheng and Amit Sheth},
  title = {{Clustering for Simultaneous Extraction of Aspects and Features from Reviews}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2016},
}
Figure