Ill-Formed and Non-Standard Language Problems

Stanley C. KwasnyACL 1982 · DOI

Introduction

In the early development of Natural Language Processing (NLP), systems were primarily designed to interpret language that adhered to strict, formal grammatical rules. While these systems performed adequately on "textbook" English, they frequently failed when confronted with the realities of human communication. In his 1981 work, Stan Kwasny argues that for NLP to achieve true utility, it must move beyond rigid syntactic structures and learn to navigate the messy, "ill-formed" nature of real-world dialogue. This paper explores the necessity of graceful degradation and the strategies required to bridge the gap between standard English and non-standard usage.

The Problem of Linguistic Brittleness

By the early 1980s, a significant "brittleness" had been identified in language technology. When users interacted with decision-support systems or early databases, they rarely utilized the perfect syntax expected by the software. Instead, human-computer interaction was—and remains—characterized by fragments, typos, ellipsis (omitted words), and non-standard syntax.

Kwasny identifies that most systems of the era would fail entirely if an input did not perfectly match a predefined grammar. This lack of flexibility created a barrier to adoption, as users were forced to conform to the machine's limitations rather than the machine adapting to the user's intent. The core challenge lies in extracting meaning from technically "ungrammatical" strings while maintaining the integrity of the communication.

Methodological Approaches to Error Handling

To address these failures, Kwasny synthesizes several emerging techniques designed to handle linguistic deviations:

Meta-rules and Grammar Relaxation

One of the most promising approaches involves the use of "meta-rules." Rather than expanding the base grammar to include every possible error, meta-rules describe how to relax or manipulate standard rules when a violation occurs. For instance, if a system encounters a subject-verb disagreement or a missing constituent, a meta-rule can allow the parser to bypass the specific constraint to find the most likely intended structure.

Concept Skimming and Semantic Coherency

Drawing on the work of DeJong, the paper discusses "concept skimming." This method prioritizes the "coherency" of ideas over syntactic perfection. By focusing on identifying key concepts and their relationships within a sentence, a system can often understand the gist of a message even if the formal grammar is broken or incomplete.

Human-in-the-Loop and Speech Processing Strategies

When ambiguity arises from an ill-formed string, some systems are designed to present the user with multiple possible interpretations, allowing the human to clarify the intent. Additionally, Kwasny suggests borrowing non-left-to-right processing strategies from speech recognition. These strategies allow a system to jump over an "errorful" segment of a sentence and use the surrounding context to narrow down legal alternatives for the problematic part.

Key Insights into Human-Computer Dialogue

The Grammaticality Continuum

Kwasny argues that language should not be viewed as a binary of "correct" or "incorrect." Instead, grammaticality exists on a continuum. A robust NLP system must be able to operate across this entire spectrum, recognizing that a sentence can be technically ungrammatical yet perfectly intelligible.

Man-Machine Pidginization

A particularly compelling observation in the research is the phenomenon of "man-machine pidginization." Users naturally adapt to the constraints of the machine. If a user discovers a specific shorthand or simplified syntax that the machine understands, they will consistently use it. This creates a shared, simplified "pidgin" dialect unique to the interaction between that specific human and computer.

The Intuition Trap

The paper cautions linguists against relying solely on their own intuitions regarding what constitutes "standard" language. Empirical studies, such as those by Malhotra (1975) and Thompson (1980), demonstrate that professional linguistic intuition often fails to predict how actual users communicate in practice.

Conclusion and Future Directions

The ultimate goal for NLP is the development of adaptive systems capable of graceful degradation. While the computational costs of considering multiple alternative interpretations were a significant bottleneck in 1981, the theoretical path forward is clear.

Future work in this field must focus on:

  1. Adaptive Learning: Enabling systems to modify their own rules and meta-rules based on historical interaction.
  2. User Modeling: Developing specific profiles for individual users to anticipate their stylistic preferences and common errors.
  3. Architectural Evolution: Utilizing parallel architectures and faster hardware to allow for the creative, non-linear processing required to handle non-standard language in real-time.

True machine intelligence is not found in the rigid enforcement of rules, but in the ability to maintain communication despite the inevitable errors of human expression.

Summary updated

BibTeX
@inproceedings{kwasny1982ill,
  author = {Stanley C. Kwasny},
  title = {{Ill-Formed and Non-Standard Language Problems}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1982},
  doi = {10.3115/981251.981296},
}
Figure