Abstract
In the mid-2000s, the rise of the blogosphere created a need for search engines that could identify human opinions rather than just factual data. This paper introduces a specialized search engine for Japanese blogs that extracts explicit writer opinions using Support Vector Machines (SVM) and a library of nearly 3,000 opinion clue expressions. By incorporating semantic categories and linguistic heuristics tailored to Japanese sentence structure, the system significantly improves opinion extraction performance over traditional keyword-based methods.
Introduction
Early opinion mining primarily focused on sentiment analysis—classifying text as strictly positive or negative. However, human discourse is more nuanced, encompassing advice, neutral judgments, and specific requests. General subjectivity classifiers often suffer from information overload, identifying many subjective sentences that do not represent the core opinion a user seeks. This research addresses the challenge of extracting explicitly stated writer opinions at the sentence level while filtering out noise such as quotes or implicational statements.
System Architecture
The researchers proposed a two-module architecture to balance extraction accuracy with real-time search performance.
1. Opinion Sentence Extraction (Offline)
To enable real-time search, the system pre-processes blog pages using an SVM classifier with three primary feature sets:
- Opinion Clue Expressions: A curated list of 2,936 phrases (e.g., "I think," "terrific," "should go").
- Semantic Categories: 2,715 categories from a Japanese thesaurus to help the system generalize across synonyms.
- Position Heuristics: In Japanese, the predicate usually appears at the end of the sentence. The researchers found that restricting the search for clues to the last ten words of a sentence significantly improved precision.
2. Query-Relevant Sentence Extraction (Online)
Once a user enters a query, the system determines if the indexed opinion sentences are relevant to that query. Strategies tested included checking if the query appears in the sentence itself, in preceding sentences, or within a consecutive "chunk" of opinion sentences.
Experimental Methodology
The system was evaluated using a dataset of Japanese blogs across 10 diverse categories, including Culture, Food, Health, Politics, and Products. A training set of 23,800 sentences and a test set of 5,686 sentences were manually labeled by human evaluators. Relevance was further tested on a set of 9,841 sentences to evaluate the online module's effectiveness.
Key Results and Analysis
The proposed SVM method, particularly when utilizing semantic categories, demonstrated a clear advantage over keyword-based baselines.
As shown in the results, the inclusion of semantic categories boosted precision to 72.5% and accuracy to 79.2%. Furthermore, focusing on the "Quasi-predicate part" (the end of the sentence) was found to be a critical heuristic for Japanese opinion mining. Subjective words in the predicate are more likely to signal an actual opinion rather than a descriptive report (e.g., "This house is beautiful" vs. "We purchased a beautiful house").
Conclusion and Future Work
While the system achieves high precision, recall remains a challenge, as the restrictive nature of the classifier misses some valid opinions. Future work will focus on refining query relevance heuristics, developing user interfaces that automatically categorize opinions into buckets like "Emotion" or "Suggestion," and moving toward automated summarization of the collective blogosphere's viewpoint on specific topics.
Summary updated