Abstract
Traditional opinion mining often overlooks the temporal context of sentiments, leading to ambiguous signals for predictive tasks. This paper introduces "Time-Sensitive Opinion Mining," a framework that extracts the reference time (RT) of opinions to distinguish between past reports and future predictions. By weighting future-oriented sentiments more heavily, the proposed model significantly enhances the accuracy of stock market predictions using social media data.
Introduction: The Temporal Gap in Opinion Mining
Predicting market trends via social media typically involves extracting sentiment polarity (bullish vs. bearish). However, a single post can contain conflicting temporal references, such as: "The market was bullish today, but tomorrow it may turn bearish." Standard models often fail here, either canceling out the signals or averaging them. For predictive applications, the sentiment regarding the future is inherently more valuable than a summary of the past. The challenge lies in automatically distinguishing these temporal references to prioritize forward-looking information.
Methodology: Extracting Reference Time (RT)
The authors propose a framework to identify the Reference Time (RT) of an opinion, classifying it as "future" or "non-future." This classification relies on four primary feature sets:
- Temporal Expression (TE): Mapping specific dates or time-related words to absolute timestamps.
- Tense Indicator (TI): Utilizing a dictionary of verbs and grammatical markers (e.g., "will" vs. "was") to infer temporal direction.
- Prediction/Summary Indicator (PSI): Identifying keywords with intrinsic temporal meaning, such as "expect" or "forecast."
- Specificity Level (SL): A heuristic based on the observation that reports of the past are often more precise (e.g., specific index values) than abstract future predictions.
Once tagged, future-oriented opinions are assigned a higher weight ($1 + \alpha$) in the prediction model, while past-oriented opinions maintain a baseline weight of $1$.
Experimental Setup
The framework was evaluated using:
- SinaWeibo Data: Over 50,000 microblogs concerning the Shanghai Stock Index (SHI) from early 2013.
- Ground Truth: 1,000 manually labeled opinions to validate RT tagging accuracy.
- Market Benchmarks: Actual Shanghai Market Index movements were used to assess final prediction performance.
Key Results and Discussion
The "Time-Sensitive" approach demonstrated significant improvements in both tagging and prediction tasks.
Tagging Performance
The combined feature model ("ALL") achieved a tagging accuracy of 95% with 98% coverage. As shown in the feature analysis table, while individual indicators like Tense Indicators (TI) had lower coverage (27%), their high precision contributed to a robust ensemble model.
Prediction Accuracy
The time-sensitive model consistently outperformed both the baseline (no sentiment) and traditional sentiment models.
As illustrated in the results, increasing the weight of future opinions ($\alpha$) led to a peak in classification accuracy, confirming that forward-looking statements carry superior predictive power. The "Specificity" heuristic proved particularly insightful, showing that linguistic precision is a strong proxy for temporal reference.
Conclusion and Future Directions
This research highlights that contextual metadata—specifically time—is as critical as sentiment polarity for domain-specific forecasting. Future work may involve refining the binary "future/non-future" classification into more granular segments (e.g., near-term vs. long-term) and developing more sophisticated linguistic models to handle complex structures like sarcasm or conditional statements.
Summary updated