An Empirically Based Approach Towards a System Of Semantic Features

Cornelia Zelinsky-WibbeltCOLING 1986

TL;DR

This paper introduces a novel, empirically grounded system for generating semantic features, crucial for machine translation. It proposes a formal rule-based system to represent complex relationships between semantic features, aiming for a coherent and operational description of lexical units.

What Problem Does This Paper Address?

A fundamental challenge in machine translation (MT) is the accurate and consistent semantic description of words (lexical units). Existing MT systems often struggle with semantic representation, lacking a coherent and operational system that can guarantee consistency across different lexical items. This paper addresses the need for a robust semantic feature system that is both theoretically sound and practically applicable within the demanding context of MT. The authors highlight the shortcomings of existing feature inventories in various MT systems, such as insufficient definitions, vague feature descriptions, and a lack of clarity on how features relate to each other.

Methods

The authors propose a generative system for semantic features based on a context-free rule system. This system formalizes three key relationships between semantic features:

  • Hierarchy: Defines the derivation of features, where a dominating feature refines into subordinate ones.
  • Alternation: Allows only one feature from a set to apply (represented by '/').
  • Disjunction: Indicates features that obligatorily occur together (represented by '*').

The formalism uses a rule format like X = (A/B)*(C/D), where X is the dominating feature, and A, B, C, D are refinements. Terminal features are represented by rules like X = 0.

To handle co-occurrence restrictions, the system introduces attributes. Derived attributes (on the left-hand side of a rule) are "inherited" by subordinate features (on the right-hand side), influencing rule application in disjunct branches. This attribute mechanism is illustrated with examples like Y = A[X], where X is an inherited attribute. The authors emphasize that each semantic feature is defined only once, with "or" relations combined by brackets when prior to "and" relations.

Datasets

Not discussed in the paper. The authors mention using the "ESPRIT-corpus" for empirical work and consulting semantic feature inventories from other MT systems (METAL, JAPAN, SYSTRAN, SUSY), but these are not presented as datasets for training or evaluation in the context of this paper.

Key Insights and Ideas

  • Empirical grounding is key: The proposed system is not purely theoretical but is built upon empirical work with a corpus and consultation of existing linguistic and computational semantic resources.
  • Coherent and operational system: The primary goal is to create a semantic feature system that is both logically consistent and directly usable by MT systems.
  • Formalizing feature relationships: The paper introduces a clear, rule-based formalism to explicitly define hierarchical, alternating, and disjunctive relationships between semantic features.
  • Attributes for context-specific restrictions: The introduction of derived and inherited attributes provides a mechanism to manage complex co-occurrence restrictions, enhancing the expressiveness of the system.
  • Linguistic motivation: The organization of the semantic features is explained through cognitive principles like prototypicality and gestalt psychology, linking the formal system to human conceptualization.
  • Tree-like representation: The rule system naturally generates a tree structure, which the authors found to be a transparent way to illustrate the complex relationships within the semantic feature set.

Key Results

The paper presents a formal framework and a methodology for constructing a semantic feature system. While specific quantitative results are not detailed, the authors demonstrate the system's capability to:

  • Represent complex semantic relationships.
  • Handle co-occurrence restrictions through attributes.
  • Generate a structured semantic description for lexical units.
  • The authors state they have defined 87 features for "ENTITIES" and 87 for "SITUATIONS" so far.

The output of this generating system is intended as input for dictionaries and for disambiguation processes in MT.

Limitations and Future Work

  • Incomplete feature definitions: The authors acknowledge that the semantic features are not always sufficiently defined, especially in the EUROTRA proposal, making their conceptual structure and usage unclear.
  • Vague feature descriptions: Even with examples, understanding the precise meaning and application of some features remains challenging.
  • Underexplored dependencies: The paper notes that dependencies between features are not always fully explained, particularly in existing systems.
  • Future work on synthesis: The authors suggest that the rule system could be used in synthesis to identify dependencies between semantic features, which would be necessary for semantic generalizations in the target language. This area, however, has not yet been worked out.
  • Efficiency of translation: The possibility of incorporating redundancies by including other dominating features in the list is mentioned as a potential avenue for a more efficient translation process.

Summary updated

BibTeX
@inproceedings{zelinskywibbelt1986empirically,
  author = {Cornelia Zelinsky-Wibbelt},
  title = {{An Empirically Based Approach Towards a System Of Semantic Features}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1986},
}

Loading related content...

Figure