Embedded Sublanguages and Natural Language Processing

Richard KittredgeCOLING 1980

TL;DR

This paper explores the concept of "sublanguages" – specialized subsets of natural language used in specific domains like science or finance. The authors argue that understanding the structure and embedding of these sublanguages is crucial for building more effective natural language processing (NLP) systems, particularly for tasks like information retrieval and translation.

What Problem Does This Paper Address?

Modern NLP systems often struggle with the vast diversity and nuance of human language. Many systems are trained on general language corpora and perform poorly when encountering texts from specialized domains. These specialized texts, or "sublanguages," have their own unique vocabularies, grammatical structures, and ways of expressing concepts. The challenge is that these sublanguages aren't always isolated; they can be embedded within broader texts, creating complex linguistic structures that are difficult for current NLP parsers to handle. For example, a financial report might contain a technical description of trading activity (a narrow sublanguage) embedded within a broader discussion of economic events (a more general language). The paper aims to address how to identify, analyze, and process these embedded sublanguages effectively.

Methods

The paper doesn't detail specific algorithms but focuses on the conceptual framework for analyzing sublanguages. The authors propose that understanding the "discovery of special structural properties, lexical collocations, etc." within functionally and semantically homogeneous corpora is key to defining sublanguages. They highlight the importance of identifying how sublanguages are embedded within each other, drawing parallels to mathematical concepts like "right ideals in a ring" to describe how narrower sublanguage sentences can be adjoined to broader ones. The analysis involves studying textual structure and lexical selection to clarify the defining properties of sublanguages and how they interact.

Datasets

Not discussed in the paper. The authors refer to "a detailed study of several sublanguages of English and French" and mention examples like "technical manuals," "scientific texts," "weather synopses," and "stock market reports," but no specific datasets are named or described.

Key Insights and Ideas

  • Sublanguages are more than just vocabulary: While specialized vocabularies are a hallmark, the real power of sublanguages lies in their unique grammatical structures and lexical co-occurrence restrictions, which reflect the underlying domain's properties and relationships.
  • Embedding is common and complex: Sublanguages rarely exist in isolation. They are often embedded within broader linguistic contexts, creating hierarchical structures that require sophisticated parsing strategies.
  • Domain-specific analysis is crucial for NLP: To build truly intelligent text processors, we need to move beyond general language models and develop methods that can recognize and exploit the specific rules of various sublanguages.
  • Structural properties reveal sublanguage boundaries: Identifying distinct sentence structures, textual linking devices, and the scope of anaphora within different parts of a text can help delineate sublanguage boundaries.

Key Results

The paper presents a conceptual framework and observational findings rather than quantitative results. The key "results" are the insights into the nature of embedded sublanguages and their implications for NLP. The authors demonstrate that:

  • Scientific texts can contain both "science language" (describing objects/relations) and "meta-science language" (describing the observer's relationship to methods/results), which can be embedded within each other.
  • Stock market reports exhibit embedding where descriptions of trading activity are nested within broader discussions of economic and political events.
  • The junctures between embedded sublanguages and their surrounding "matrix" language are critical points for NLP parsers to recognize and exploit.

Limitations and Future Work

The authors acknowledge that the notion of sublanguage is "rather loosely defined" and that there's no rigorous procedure for determining sublanguage membership or defining its limits. They suggest that their proposed paper will present and discuss several of the observed properties, drawing conclusions for the relationship between various sublanguages and language as a whole, and discussing implications for the design of intelligent text-processors. This implies that the work presented is foundational and opens avenues for further research into formalizing sublanguage identification and developing NLP techniques to handle them.

Summary updated

BibTeX
@inproceedings{kittredge1980embedded,
  author = {Richard Kittredge},
  title = {{Embedded Sublanguages and Natural Language Processing}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1980},
}
Figure