A Multilayered Approach to the Handling of Word Formation

Wolfgang HoeppnerCOLING 1982

TL;DR

This paper introduces a novel, multilayered approach to handling word formation (derivatives and compounds) within a natural language dialogue system. Instead of treating word formation as a separate module, it's integrated across multiple processing levels, from lexical analysis to semantic interpretation and dialogue context, aiming to improve linguistic coverage and reduce vocabulary errors.

What Problem Does This Paper Address?

Natural language processing (NLP) systems often struggle with word formation, which is a fundamental aspect of language for expressing complex relationships and ideas. While linguistic theories acknowledge word formation, AI research has historically neglected it due to its perceived complexity. This paper argues that word formations, like derivatives and compounds, are crucial for increasing a system's vocabulary and reducing errors. The challenge lies in the fact that the relationships between word components are often implicit, leading to ambiguities on various linguistic and cognitive levels. Existing approaches often treat word formation in isolation, failing to capture its dynamic nature within a dialogue.

Methods

The paper proposes a multilayered approach integrated into the HAM-ANS dialogue system. This approach processes word formations across several levels:

  • Lexical Analysis: The system first attempts to reduce inflected word forms and identify potential derivatives or compounds if a word isn't found in the lexicon. This involves using a table of derivative suffixes, graphemic rules, and lexicon entries.
  • Segmentation and Identification: For unknown words, the system identifies derivative suffixes and then attempts to split words into two components, both of which must be transformable into canonical forms. This process is recursive, allowing for multiple compounding and derivation, though the current implementation is limited to two-part compounds and singular derivation.
  • Semantic Interpretation: Identified word formations are mapped onto a logic-oriented representation language called SURF. For derivatives, this involves using SURF expressions associated with suffixes. For compounds, inference strategies are employed to discover semantic relations between constituents, drawing on a conceptual semantic network. These relations can include physical object properties, material, location, and combinations of objects.
  • Dialogue Context Evaluation: The system evaluates interpretations of word formations within the ongoing dialogue. This allows for reinterpreting or rejecting initial assumptions based on the conversational context, similar to how elliptical utterances are handled.
  • Generation: The system can also generate word formations. One approach involves creating semantic interpretations of compounds and derivatives and storing them in lexical memory for later use in answer generation. Another approach uses patterns to bind relation identifiers and canonical word forms, allowing the system to decide whether to use a word formation, derive, or compound.

Datasets

Not explicitly discussed in the paper. The examples provided are drawn from dialogues within the hotel-reservation and motion analysis domains of the HAM-ANS system.

Key Insights and Ideas

  • Integrated Approach is Key: Word formation shouldn't be a separate module but rather woven into the fabric of NLP processing, from initial lexical analysis to final dialogue evaluation.
  • Multilayered Processing: Analyzing word formations requires considering multiple levels: the meaning of constituents, their interpretation in context, and their situational context.
  • Implicit Relations: The core challenge is inferring the implicit semantic relationships within word formations, which requires leveraging knowledge sources and inferential capabilities.
  • Dialogue Context is Crucial: The ultimate evaluation and refinement of word formation interpretations happen within the dynamic context of a dialogue.
  • Generation as a Mirror of Analysis: Generating word formations can be viewed as a process analogous to generating elliptical utterances, leveraging stored semantic interpretations.

Key Results

The paper focuses on the methodology and approach rather than presenting quantitative results. The key "results" are the successful integration of a multilayered word formation handling mechanism into the HAM-ANS system, enabling it to:

  • Identify and segment derivatives and compounds.
  • Generate semantic interpretations for these word formations.
  • Utilize this knowledge within a dialogue context for both understanding and generation.

The paper demonstrates the approach with examples like "STRASSENFEGER" (street cleaner) and "STUHLBEIN" (chair leg), illustrating the lexical representation and semantic interpretation processes.

Limitations and Future Work

  • Limited Scope of Current Implementation: The current HAM-ANS system is limited to handling compounds with two parts and singular derivation.
  • Degrees of Lexicalization: The sharp distinction between lexicalized and analyzable words is acknowledged as not fully capturing the nuances of lexicalization and is slated for improvement.
  • Ambiguity Handling: While the system aims to handle ambiguities, the paper notes that the parser more frequently encounters ambiguity with compounds and derivatives, requiring more sophisticated mechanisms for selection.
  • Future Work: The authors suggest that the analytical capacity can be extended to handle more complex word formations. The integration of semantic interpretations into knowledge sources for subsequent processing and generation is a key area of ongoing development. The paper also hints at the need for more robust methods for analogy-driven resolution of word formations.

Summary updated

BibTeX
@inproceedings{hoeppner1982multilayered,
  author = {Wolfgang Hoeppner},
  title = {{A Multilayered Approach to the Handling of Word Formation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1982},
}
Figure