Abstract
Traditional text mining for relationship discovery often relies on simple word co-occurrence (collocation). This paper proposes a more sophisticated approach by analyzing the evolution of public opinion over time. By mining "opinion tracking" curves from over a million web documents and integrating them with traditional metrics, the researchers demonstrate a highly precise method for identifying latent correlations between entities.
Introduction
In text mining, relationship discovery typically assumes that if two entities (e.g., "Apple" and "Foxconn") appear frequently in the same context, they are related. However, this "collocation" approach is shallow; it often misses deeper, event-driven connections where two entities are affected by the same market forces but are not explicitly mentioned together. The authors hypothesize that correlated companies will exhibit similar "sentiment trajectories" in public opinion over time, providing a proxy for shared underlying events.
Methodology
The researchers developed a pipeline consisting of three primary components:
- Opinion Tracking & Burst Detection: The system calculates daily opinion scores for target companies from web documents. To filter out noise and identify significant events, they utilize Kleinberg’s burst detection algorithm to create a smoothed sentiment curve.
- Relationship Models: Three categories of models were evaluated:
- Collocation-based: Standard statistical measures like Mutual Information (MI) and t-tests.
- Opinion-based: Novel metrics including Curve Overlap (CO), which measures trajectory similarity, and a Chi-square ($\chi^2$) approach based on daily sentiment tendencies.
- Integration Models: Hybrid approaches (CO+MI and CO $\cap$ MI) combining sentiment similarity with co-occurrence data.
- Filtering: A "general name" filter was implemented to remove companies with names that are also common adjectives (e.g., "Happy"), which frequently cause false positives in sentiment mining.
Datasets and Evaluation
The study utilized a massive web corpus of over 1.2 million economics-related documents. To establish a "Gold Standard," the authors used real-world financial data from 1,078 Taiwanese companies. Company pairs showing statistically significant correlations in stock price movements (relative to the market index) were used as the ground truth for related entities.
Key Results
The integration of sentiment trajectories with collocation data proved superior to either method in isolation. The CO+MI model emerged as the most effective approach.
As shown in the precision curves, the CO+MI model achieved perfect precision (1.00) for the top 25 discovered relationships and maintained significantly higher precision than traditional MI as the number of proposed pairs increased. Specifically, it achieved 0.92 precision at the top 50 and 0.79 at the top 100.
Discussion and Insights
The core finding is that public opinion serves as a powerful reflection of underlying events. If two companies react similarly to the same sequence of events, their opinion curves will overlap, revealing a relationship even if they are never mentioned in the same article. While collocation models are effective for explicit partnerships (like supplier-buyer), opinion tracking excels at identifying companies in the same sector that react similarly to macro-economic shifts.
However, the system remains sensitive to document relevance. Sentiment models can be easily skewed by irrelevant mentions, making high-quality Information Retrieval and name filtering essential for accuracy. Future work may involve applying these models to non-corporate targets, such as politicians, or using opinion curves as leading indicators for short-term stock market movements.
Summary updated