On a Certain Distribution of Semantic Units

Wojciech SkalmowskiCOLING 1965

TL;DR

This paper explores the distribution of "semantic units" within Arabic, proposing that their frequency follows a predictable pattern related to their complexity or "semantic value." The authors suggest this regularity might be a general linguistic phenomenon, offering a potential computational approach to understanding and resolving "true" multiple meaning in machine translation.

What Problem Does This Paper Address?

The paper tackles the fundamental challenge of defining and operationalizing "semantic units" in natural language. It highlights the difficulty in creating a robust theory of semantics, leading to much linguistic investigation being "artisanlike" and tentative. A key problem is the phenomenon of "true" multiple meaning, which arises from the inherent "looseness" and non-isomorphic nature of conceptual systems across languages. This ambiguity poses a significant hurdle for machine translation (MT), particularly in handling nuanced or expressive language. The authors aim to investigate whether underlying structural regularities in semantic units can provide a more systematic way to approach this problem.

Methods

The paper proposes a framework for describing semantic concepts and units using generative grammar and semantic theory. A semantic concept is viewed as an ordered set of symbols from grammar and semantics, while a semantic unit is similarly defined but linked to specific word formation rules and grammatical derivations. The core of the empirical investigation involves analyzing the distribution of Arabic verbal roots based on the number of stem-patterns they can form. This number of stem-patterns is hypothesized to be an indicator of the "content" or "semantic value" of the semantic unit. The authors then fit this distribution to a mathematical function and test its validity against empirical data from Arabic dictionaries.

Datasets

The primary data source for the empirical investigation is Arabic verbal roots. The authors utilize:

  • Baranov's Arabic-Russian Dictionary: Used for counting roots and their stem-pattern distributions.
  • Two samples of approximately 700 items each from different Arabic dictionaries (referenced as /7/ and /3/): Used to compare dictionary-specific distributions with the overall distribution and assess potential differences arising from compilation materials.
  • Chambers' 20th Century Dictionary: Used in an illustrative example to estimate the "value" of English verbs based on the number of their sub-entries.

No new datasets or benchmarks are introduced; the focus is on analyzing existing linguistic data through a new lens.

Key Insights and Ideas

  • "Semantic Value" as a Distributional Property: The paper posits that the frequency of semantic units is not random but follows a predictable distribution, potentially linked to their inherent complexity or "semantic value."
  • Arabic as a "Laboratory Example": The clear grammatical structure of Arabic makes it an ideal language for testing hypotheses about semantic unit distribution.
  • Stem-Patterns as a Proxy for Semantic Value: The number of stem-patterns a verbal root can form in Arabic is used as a quantifiable measure of its "semantic value."
  • Mathematical Regularity in Language: The observed distribution of Arabic roots suggests that linguistic structures, even at the semantic level, might adhere to mathematical principles.
  • Potential for MT: If such regularities are generalizable, they could offer a computational basis for resolving semantic ambiguities in machine translation by matching "semantic values" across languages.

Key Results

The investigation into Arabic verbal roots revealed a striking regularity in their distribution based on the number of stem-patterns. The relationship between the number of stem-patterns ($s$) and the number of roots belonging to that class ($r$) was found to be approximately described by the function $r/s/ = N/As² +Bs +C/$, where $N$ is the total number of roots and $A, B, C$ are constants. The goodness of fit was tested using the chi-square distribution and found to be statistically significant, with differences between empirical data and the theoretical distribution not exceeding 0.3 significance level, except for one value. The distributions from different dictionaries showed remarkable similarity, supporting the general validity of this pattern in Arabic and suggesting the existence of "natural hierarchies" of semantic units.

Limitations and Future Work

The authors acknowledge that the applied "method" is unsystematic and ad hoc, limiting the possibility of broad generalizations from the illustrative example of English verbs. They suggest that future work should focus on:

  • Discovering similar regular distributions in other languages: This would require wider "trial and error" investigations but could be mechanized.
  • Developing a more comprehensive set of "semantic parameters": The current approach is a simplification.
  • Establishing objective, numeric measures of semantic content: This would involve ordering subsets of synonyms within concepts based on the intrinsic value of semantic units.
  • Investigating interconnections between syntactic and semantic structures: This is a recognized area in structural linguistics and computational linguistics.

The paper implies that a deeper understanding of these semantic regularities could lead to more robust solutions for machine translation, particularly in handling the complexities of "expressive language" and "true" multiple meaning.

Summary updated

BibTeX
@inproceedings{skalmowski1965certain,
  author = {Wojciech Skalmowski},
  title = {{On a Certain Distribution of Semantic Units}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1965},
}
Figure