Introduction
In the field of Natural Language Generation (NLG), early research primarily focused on the distinction between definite and indefinite noun phrases. However, human communication frequently employs subjective, non-singular quantifiers known as "judgment determiners" (JDs), such as many, few, or a lot. Unlike objective cardinals (e.g., "six assignments"), JDs convey an argumentative intent, framing a quantity relative to expectations or goals. For instance, stating a course has "many assignments" suggests a high workload, whereas "six assignments" is a neutral report.
The Empirical Need for Judgment Determiners
The necessity of modeling JDs is underscored by their prevalence in natural dialogue. Analysis of a 40,000-word corpus of academic advising sessions reveals that JDs are used more frequently than exact cardinal numbers when referring to quantities.
As shown in the data, judgment determiners appear significantly more often than cardinals in this domain, justifying the development of computational methods to generate them accurately.
Methodology: Modeling Argumentative Intent
Generating JDs requires moving beyond simple counting. The proposed method, implemented within the ADVISOR II system, utilizes three formal semantic properties:
- Non-Extensionality: The truth value of a JD depends on the speaker's decision to highlight a number as significant within a specific context.
- Monotonicity: Determiners are classified as "increasing" (e.g., many) or "decreasing" (e.g., few). This is represented by an
orientationfeature. - The Intersection Condition: The system distinguishes between the reference set (the baseline) and the intension (the specific subset being evaluated). For example, "Most AI topics are interesting" and "Few of the topics that interest you are in AI" may describe the same set of facts but differ in which set is treated as the reference.
The system uses a functional unification formalism (FUF) to map conceptual representations—enriched with pragmatic annotations—to syntactic features. Key features include orientation (+ or -) and degree (+ or -), which allow the generator to select the appropriate JD based on the desired argumentative strength.
As illustrated above, varying the orientation and degree features allows the system to produce nuanced variations like "a little" versus "almost no," each carrying a different pragmatic weight.
Implementation and Results
The logic was fully integrated into ADVISOR II, a system designed to provide course-scheduling advice. By defining a set of 24 syntactic features, the generator can produce complex determiner sequences such as "all the" or "twice as many." This allows the system to actively support a conversational goal—for example, encouraging a student by framing a set of topics as "most" rather than just a raw count.
Conclusion and Future Directions
While the current model successfully maps argumentative intent to syntactic realization, it relies on a coarse-grained scale for "degree." Future work aims to refine these scales to distinguish between subtler variations like "a lot" and "an incredible number of." Additionally, while the semantic tests are universal, the heuristics determining when a count becomes "many" are domain-specific and require further exploration across different contexts.
Summary updated