Processing Pragmatics for Computer-Assisted Language Instruction

Keiko HoriguchiAAAI 1994

Abstract

Computer-Assisted Language Instruction (CALI) systems typically excel at syntactic analysis but struggle with the pragmatic dimensions of language, such as politeness and social context. This paper introduces the p-structure (pragmatic structure), a framework designed to model these nuances within the ALICE-chan Japanese tutorial system. By encoding speech acts, social hierarchies, and linguistic "reducing factors" that signal uncertainty or indirectness, the system can provide sophisticated, context-aware feedback. This approach enables the detection of pragmatic errors—such as incorrect social "directionality" in Japanese—that traditional grammar checkers overlook.

Introduction

Traditional Computer-Assisted Language Instruction (CALI) systems are often "syntax-only" machines. While they excel at checking if a subject and verb agree, they frequently fail to account for the immense variability and social requirements of natural language.

In a language like Japanese, the social context (who you are talking to) and the "direction" of an action (who is doing a favor for whom) are as critical as the grammar itself. A sentence might be grammatically perfect but pragmatically "broken"—perhaps it is too blunt, uses the wrong level of politeness, or misrepresents the speaker's intent. Keiko Horiguchi addresses this gap between purely structural processing and the actual social use of language by introducing a pragmatic modeling layer.

The P-Structure Framework

The core innovation of this research is the p-structure. Unlike standard syntactic trees, the p-structure represents more than just the propositional content; it encodes the "how" and "why" of an utterance.

  1. Encoding Pragmatics: The p-structure tracks the speech situation, the speaker’s attitude toward the listener, and "felicity conditions"—the criteria that must be met for a speech act (like a request) to be considered successful.
  2. Reducing Factors: The system treats linguistic markers of uncertainty (e.g., "I wonder if...") as "reducing factors." These factors weaken the force of a speech act, which in many cultures—especially Japanese—is a key component of politeness and indirectness.
  3. Mapping and Matching: A mapping program converts raw syntactic structures into p-structures. An error analysis matcher then compares the student’s p-structure against a "pragmatic analysis template." If the student uses a word that changes the social "direction" of the sentence, the system identifies the specific pragmatic mismatch.
An example of a p-structure representation for a formal request in Japanese.

Key Insights

  • Beyond Binary Correctness: By using p-structures, the system moves away from binary "right or wrong" judgments. It can recognize that multiple different sentence structures might fulfill the same pragmatic goal, giving students more freedom to experiment with expression.
  • Modeling Uncertainty: One of the most significant contributions is the formalization of "uncertainty" as a functional component of a speech act. By identifying words like darou (probably) or ka (question marker) as factors that reduce the "strength" of a request, the system can computationally "understand" politeness.
  • Directionality Matters: In Japanese, "giving" and "receiving" verbs change based on the social relationship between the speaker and the listener. The p-structure specifically tracks the "placement of addressee" (e.g., "higher" status), allowing the system to catch subtle social faux pas that a standard grammar checker would miss.

Results and Feedback

The paper demonstrates the system's ability to provide highly specific, human-like feedback. For example, if a student uses the verb sashiagerarenai (giving to an inferior) instead of itadakenai (receiving from a superior) in a formal request, a standard system might just flag it as a vocabulary error. Horiguchi’s system generates a nuanced explanation:

"You seem to have used the giving verb with the wrong direction. You should have used the verb with inward direction."

This qualitative improvement allows the system to act more like a tutor and less like a spell-checker, reducing the workload for human instructors who otherwise must manually program every possible correct variation of a sentence.

Limitations and Future Work

As an early implementation (1994), the scope of this framework has several limitations:

  • Language Specificity: The current implementation is tailored specifically to Japanese, which has a very rigid and explicit pragmatic system (honorifics, benefactives). Applying this to languages with more implicit pragmatics, like English, would require a different set of features.
  • Scaling: The paper does not discuss how the "pragmatic analysis templates" are created. If they must be hand-authored for every possible lesson, scaling the system to a full curriculum could be labor-intensive.
  • Computational Overhead: The mapping from syntactic structure to p-structure adds a layer of complexity that was an extension of the existing ALICE-chan system, though this is less of a concern for modern hardware.

Summary updated

BibTeX
@inproceedings{horiguchi1994processing,
  author = {Keiko Horiguchi},
  title = {{Processing Pragmatics for Computer-Assisted Language Instruction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}

Loading related content...

Figure