Albert Gatt, François Portet. Proceedings of the 13th European Workshop on Natural Language Generation. 2011.
Abstract
This paper introduces a framework for Natural Language Generation (NLG) systems to handle "fuzzy" time data. By combining fuzzy temporal reasoning with cross-linguistic experimental data, the authors demonstrate how to map mathematical uncertainty (Possibility and Necessity) to natural language expressions like "might," "must," or "possibly" in English, French, and Maltese.
Introduction: The Challenge of Temporal Uncertainty
In many real-world applications—such as medical databases or investigative reports—data is rarely precise. We might know a patient was treated "sometime between 2:00 PM and 3:00 PM," but not the exact minute. When an NLG system transforms this raw data into a text summary, it faces a dilemma: if it simply asserts "The patient was treated at 2:30," it is being overly precise and potentially misleading. Conversely, if it ignores the temporal relationship entirely, it loses valuable information.
The challenge is twofold: how to mathematically represent the "subjective" uncertainty a human feels when looking at fuzzy time intervals, and how to choose the right "modal" expression (e.g., may, definitely, perhaps) to communicate that specific level of uncertainty to a reader.
Methodology: Fuzzy Temporal Reasoning
The authors bridge the gap between raw data and text using a two-step approach. First, they employ Fuzzy Temporal Constraint Networks (FTCN). Unlike standard temporal networks that treat events as binary (possible or impossible), FTCNs use Zadeh’s fuzzy set theory to represent constraints as possibility distributions.
This allows the system to calculate two key metrics for any event:
- Possibility ($Π$): The degree to which an event could happen.
- Necessity ($N$): The degree to which it must have happened.
To map these mathematical values to words, the authors used Multinomial Logistic Regression. This treats the choice of a modal expression as a classification problem: given a specific $(Π, N)$ pair, which word is a human most likely to choose?
Experimental Design
The authors conducted a controlled experiment involving 13 distinct scenarios, such as a bank robbery investigation. Each scenario featured variations with no uncertainty, one uncertain event, or two uncertain events.
Thirty-nine native speakers (13 each for English, French, and Maltese) participated. They used a slider (the $Ψ$-scale) to rate their subjective certainty and then selected the best-fitting sentence from a list of modalized options, such as "The robbers must have left..."
Key Results and Cross-Linguistic Insights
The study revealed several critical findings regarding how humans perceive and describe uncertainty:
- Model Validation: The fuzzy reasoning engine’s predictions correlated strongly with human subjective ratings ($r = 0.62$), suggesting that fuzzy logic is a viable proxy for human "gut feelings" about time.
- Universal Intuition, Diverse Expression: Humans across different cultures perceive the degree of uncertainty almost identically. However, the mapping of that feeling to specific words varies by language. For English and French, the logistic regression models showed an excellent fit ($p < .001$).
- The Maltese Necessity Bias: While English and French speakers' choices were predicted by a combination of Possibility and Necessity, Maltese speakers relied almost exclusively on Necessity ($N$). This suggests that different languages have different triggers for their modal verbs.
- Clustering of Modals: Data showed clear clusters: "must" and "definitely" grouped at high Necessity/Possibility, while "may" and "possibly" occupied the lower-certainty space.
Conclusion and Future Work
This research argues that for data-to-text systems, a continuous scale of uncertainty is more computationally tractable and closer to human intuition than binary "possible worlds" theories. While limited by a small sample size and a restricted vocabulary of modal auxiliaries, the study provides a foundation for integrating fuzzy reasoning into full-scale NLG pipelines. Future work aims to expand the vocabulary and explore individual reasoning strategies to produce more nuanced, faithful automated reports.
Summary updated