Generating Adjectives to Express the Speaker's Argumentative Intent

Michael ElhadadAAAI 1991

Michael Elhadad’s 1991 paper from Columbia University is a foundational contribution to Natural Language Generation (NLG). It challenges the traditional view of adjectives as simple property labels, proposing instead that they are strategic tools used to shape a listener's perspective.

Beyond Simple Property Mapping

In early NLG systems, adjective selection was treated as a direct mapping from a database to a word. If a database entry for an object included the attribute "red," the system would generate the word "red." While this approach suffices for referential tasks (identifying an object) or purely attributive tasks (stating a fact), it fails to capture the pragmatic complexity of human communication.

In real-world scenarios, such as academic advising, adjectives are frequently argumentative. When an advisor describes a course as "hard," they are not merely stating an objective property; they are often attempting to discourage a student from enrolling. Traditional systems struggled to distinguish between "hard" (which carries argumentative weight) and "difficult" (which can be neutral). Furthermore, they could not account for relativity—the fact that a course "hard" for a freshman might be "easy" for a senior.

A Multi-Layered Model for Selection

Elhadad proposes a model that integrates pragmatic goals with linguistic constraints through several key components:

  1. Argumentative Orientation (AO): This formalizes the speaker's intent. The input to the generator includes a "scale" (e.g., difficulty), an "orientation" (positive or negative), and a "focus."
  2. Scalar and Relative Representation: Adjectives are treated as points on a scale. The system incorporates a hearer model to handle relativity, defining terms like "advanced" based on the specific background of the listener.
  3. Functional Unification Grammars (FUF): The system uses FUF to manage "non-local" decisions. Adjective selection is not an isolated event; it interacts with verb selection and syntax. For instance, if a system chooses the verb "to struggle," the argumentative intent may already be satisfied, rendering an additional adjective redundant.
  4. Lexical Encoding: The lexicon distinguishes between "marked" adjectives (carrying argumentative weight) and "neutral" ones.

Empirical Analysis and Taxonomy

The research is grounded in an empirical study of a 40,000-word corpus of transcripts from human academic advising sessions. The researchers identified approximately 700 occurrences of 150 distinct adjectives, focusing specifically on those modifying "courses."

Table 1: Adjective occurrences in Importance and Evaluative classes.
Table 2: Adjective occurrences in Difficulty and Domain classes.

As shown in the tables above, the paper categorized these adjectives into four primary semantic classes: Difficulty, Domain, Importance, and Evaluative. A key finding was that 58 out of 69 adjectives modifying courses expressed properties that were not present in any standard database (such as "good" or "interesting"). These are evaluations rather than objective facts.

Key Insights: The "Hard" vs. "Difficult" Distinction

One of the paper's most compelling examples is the distinction between "hard" and "difficult." While often considered synonyms, the corpus revealed they serve different pragmatic functions. "Hard" was consistently used to discourage, whereas "difficult" appeared in neutral contexts where the advisor avoided a strong evaluation. This highlights the necessity of an "evaluation norm"—adjectives require a context relative to a specific set of expectations or student experiences.

Implementation and Legacy

The model was implemented within ADVISOR, an expert system designed to assist university students with course planning. By using Functional Unification Grammars, Elhadad demonstrated that systems could successfully navigate the complex interplay between high-level communicative goals and low-level grammatical rules. While the paper acknowledged limitations—such as the need for larger corpora—it remains a landmark study in moving NLG toward more human-like, persuasive communication.

Summary updated

BibTeX
@inproceedings{elhadad1991generating,
  author = {Michael Elhadad},
  title = {{Generating Adjectives to Express the Speaker's Argumentative Intent}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure