Extracting Resource Terms for Sentiment Analysis

Lei Zhang, Bing LiuIJCNLP 2011

Introduction

Traditional sentiment analysis systems primarily focus on explicit opinion words such as "good," "bad," or "expensive." However, a significant portion of consumer sentiment is expressed implicitly through factual descriptions of resource consumption. For example, the statement "This car uses very little gas" conveys a positive sentiment, whereas "This washer uses a lot of water" implies a negative one. These "resource terms"—such as gas, water, ink, or memory—are often domain-specific and do not appear in standard sentiment lexicons.

The underlying logic of resource consumption: low consumption implies positive sentiment, while high consumption implies negative sentiment.

Methodology

The authors propose a novel approach to automatically extract these terms from unstructured text. The process involves identifying (verb, quantifier, noun_term) triples and ranking them using a graph-based approach.

The MRE Algorithm

The core innovation is the Mutual Reinforcement based on Expected values (MRE) algorithm. Unlike the standard HITS algorithm, which treats all links equally, MRE utilizes expected values to calculate the probability that a verb is associated with a resource based on co-occurrence frequency. To ensure robustness against noise and rare terms, the authors incorporate Lidstone smoothing. This allows the system to distinguish between general nouns and genuine resource terms that are frequently modified by resource-usage verbs.

Evaluation and Results

The MRE algorithm was tested across seven diverse product domains, including automotive, appliances, and electronics. The evaluation focused on precision at various ranks (P@5, P@10).

Precision@5 comparison across seven datasets, showing MRE's superior performance over baseline methods like TF and HITS.

As illustrated in the performance table, MRE significantly outperforms baseline methods such as Triple Frequency (TF) and the original HITS algorithm. MRE achieved an average Precision@5 of 0.77 across all domains. Notably, in the "Car" domain, the algorithm achieved a perfect precision of 1.00, successfully identifying key terms like gas, fuel, and oil. Additionally, the iterative algorithm demonstrates high efficiency, typically converging within 6 to 9 iterations.

Conclusion

This research demonstrates that resource consumption is a powerful, universal indicator of sentiment. By leveraging the MRE algorithm, sentiment analysis systems can move beyond simple adjective-based detection to capture the nuanced, fact-based opinions that characterize real-world consumer reviews. Future research will focus on integrating these extracted terms into end-to-end sentiment classification pipelines.

Summary updated

BibTeX
@inproceedings{zhang2011extracting,
  author = {Lei Zhang and Bing Liu},
  title = {{Extracting Resource Terms for Sentiment Analysis}},
  booktitle = {IJCNLP},
  year = {2011},
}
Figure