Segmenting Natural Language by Articulatory Features

David ShillanCOLING 1967

TL;DR

This paper proposes a method for segmenting natural language text into "phrasings" based on articulatory features, specifically stress and tone, rather than traditional sentence or word boundaries. This approach aims to bridge the gap between spoken and written language by leveraging features that are fundamental to speech production and perception, offering a more natural and potentially more useful unit for linguistic analysis and language processing.

What Problem Does This Paper Address?

The paper addresses the limitations of traditional text segmentation units, primarily the sentence and the word. While the sentence is often accepted as a natural unit, it's highly variable in length, not always clearly identifiable in speech, and lacks a consistent definition. Words, on the other hand, are too small to capture meaningful linguistic units. The authors argue that these limitations hinder effective analysis and processing of natural language, especially when bridging the gap between spoken and written forms. They propose that segmenting text based on the rhythmical features of articulated speech, specifically "phrasings" derived from stress and tone patterns, offers a more robust and linguistically grounded alternative.

Methods

The core of the proposed method lies in identifying "phrasings" based on the prosodic features of speech, particularly stress and tone. The authors define a "phrasing" as a unit of speech that is rhythmically structured, with a primary stress-point (nucleus) and a secondary stress-point (head). This unit is seen as being between the word and the sentence in scope.

Key aspects of their approach include:

  • Focus on Articulatory Features: The method is grounded in the idea that language is fundamentally speech, involving skilled bodily movements. Features like stress and tone are considered fundamental to this articulation.
  • "Tone Group" as a Precursor: The concept of "tone groups" or "tone units" from prosodic analysis (e.g., by Crystal & Quirk) serves as a starting point.
  • "Phrasing" Definition: The authors define a "phrasing" as a unit marked by a nucleus (primary stress-point) and a head (secondary stress-point). They also introduce the concept of "silent beats" which may or may not be perceptible pauses but occur in specific semantic locations.
  • Tonetic Marking: A system of marking is proposed, using symbols for nuclear stress (often a high falling tone) and head stress (a stress-mark). Spread stress is indicated by a '+' sign.
  • Algorithmic Production: The paper mentions that this method can be adapted for algorithmic production by computer, allowing for automated phrasing of text. This involves identifying stress-points and unstressed strings (pendants) attached to them.

Datasets

The paper does not explicitly detail specific datasets used for training or evaluation in the traditional machine learning sense. Instead, it refers to:

  • Analysis of Spoken Language: Studies of spoken language, particularly in relation to foreign language teaching, are cited as a basis for understanding tone groups and their boundaries.
  • Analysis of Written Text: The authors have analyzed passages of text in various styles and lengths by hand.
  • Computerized Analysis: A program developed at C.L.R.U. is being applied to a text of 20,000 words from the Canada Year Book of 1962 for algorithmic phrasing.
  • Appendix Data: Appendix IA and IB present histograms of phrasing frequency versus phrasing length in words and syllables, respectively, based on "909 phrasings." The source of these 909 phrasings is not explicitly stated but is implied to be derived from their analytical methods.

Key Insights and Ideas

  • Speech as the Primary Form: The paper strongly advocates for viewing language primarily as speech, with written language being a representation of this underlying articulatory process. This perspective is crucial for understanding the proposed segmentation.
  • "Phrasing" as a Natural Unit: The concept of "phrasing," derived from stress and tone, is presented as a more natural and consistent unit of segmentation than sentences or words, especially for bridging spoken and written language.
  • Bridging Spoken and Written Language: The authors believe that by using articulatory features, they can effectively bridge the gap between how language is spoken and how it is written, aiding in comprehension, particularly for foreign language learners.
  • Semantic Function of Stress: The paper highlights that stress has a semantic function, and intonation can be analyzed as an "audible syntactic bracketing."
  • Universality of Phrasing: The authors suggest that the phrasing hypothesis is unaffected by differences in dialect or accent, implying a degree of universality in this segmentation approach.
  • Potential for Computational Linguistics: The development of algorithmic methods for identifying phrasings opens doors for more sophisticated natural language processing and analysis by computers.

Key Results

The paper focuses more on the conceptual framework and methodology rather than presenting extensive quantitative results. However, some key findings and observations include:

  • Agreement on Terminal Boundaries: Studies of spoken language show high agreement among informants on the terminal boundary of the "tone group" (a precursor to phrasing).
  • Phrasing as a Unit Between Word and Sentence: The authors' "phrasing" unit is found to be a minimal context of not less than one sentence, segmentable into one or more phrasings, thus existing between the word and the sentence.
  • Difficulty for Foreign Learners: Foreign learners with good vocabulary and syntax but lacking speech experience struggle with understanding written English, suggesting the importance of natural speech forms (phrasings) for comprehension.
  • Algorithmic Phrasing: A computer program has been developed that can produce a form of phrasing from text, offering a new and operational segmentation.
  • Histogram Data: Appendix IA and IB show that phrasing lengths in words typically range from 3 to 5 words, and in syllables from 6 to 10 syllables, with a clear peak around these values.

Limitations and Future Work

  • Instrumental Evidence: The paper acknowledges a "lack of satisfactory instrumental evidence on continuous speech" for some aspects of their findings, though research is planned.
  • Definition of "Stress-Point": While defined, the precise differentiation of "stress-point" from syllabic location of stress (which is a complex of amplitude, frequency, and duration) is a nuanced area.
  • "Silent Beat" Perception: The perceptibility of "silent beats" is noted as variable, and their occurrence is described in terms of typical locations rather than strict rules.
  • Future Work: The authors indicate ongoing research at C.L.R.U. in this area, including applying the phrasing program to larger texts and developing a calculus for expressing the patterning of stressed and unstressed words. They also mention planned instrumental documentation of continuous speech.

Summary updated

BibTeX
@inproceedings{shillan1967segmenting,
  author = {David Shillan},
  title = {{Segmenting Natural Language by Articulatory Features}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1967},
}

Loading related content...

Figure