An Interactive System for Stem-Suffix Discrimination in Italian Words

M. Mennucci, E. MorrealeCOLING 1973

TL;DR

This paper introduces an interactive system designed to aid in the construction of morphological dictionaries for richly inflected languages like Italian. By leveraging a tree-based representation of word forms and an iterative process of subtree extraction and selection, the system automates much of the laborious task of identifying stems and their associated suffixes, significantly streamlining dictionary creation.

What Problem Does This Paper Address?

The automatic processing of natural language, particularly for applications like information retrieval and interactive systems, hinges on understanding the structure and meaning of words. For languages with rich morphology, like Italian, where a single stem can generate numerous word forms through suffixation, building comprehensive morphological dictionaries is a complex and time-consuming endeavor. Traditional approaches often treat words as mere character sequences, failing to capture the underlying stem-suffix relationships. This paper addresses the challenge of efficiently and accurately identifying these relationships to facilitate the creation of morphological dictionaries, which are crucial for deeper linguistic analysis and more sophisticated natural language processing systems.

Methods

The core of the proposed system is an iterative process that operates on a tree-like data structure representing word forms.

  1. Tree Construction: The initial step involves building a tree where each node represents a character, and links represent their sequential occurrence in a word. All forms derived from the same stem naturally form a subtree.
  2. Minimal Subtree Extraction: The system identifies "minimal subtrees," which are branches emerging from a node that don't lead to further branching. These are the initial candidates for flexional structures.
  3. Structure Selection (Filtering): A crucial step where extracted subtrees are filtered based on occurrence counts of structures and their terminations. This process involves operator-controlled thresholds to retain significant structures and discard noise.
  4. Structure Expansion (Widening): This interactive step aims to "widen" the identified flexional structures by matching similar structures and allowing the operator to merge or adjust them. This process helps achieve more stable stem and flexional structure identifications.

The system is designed to be interactive, with human operators guiding steps 3 and 4, ensuring that linguistic nuances and exceptions are handled appropriately.

Datasets

The paper mentions experiments conducted on a small sample of 30 neurological anamneses, comprising approximately 2000 word forms. This dataset was used to test the implemented parts of the system. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Stem-Suffix Relationship is Key: The paper strongly emphasizes that the stem carries the primary semantic weight, while suffixes dictate grammatical and syntactic functions. This insight is fundamental to their approach.
  • Tree Representation for Morphology: Using a tree structure to represent word forms effectively groups related words and facilitates the identification of common prefixes and suffixes.
  • Iterative Refinement: The system's iterative nature, moving from minimal subtrees to wider structures, allows for a gradual and robust discovery of morphological patterns.
  • Human-in-the-Loop Efficiency: By making the system interactive, the authors acknowledge the complexity of natural language and leverage human expertise to validate and refine the automated processes, leading to more accurate results than a fully automatic system might achieve.
  • Italian Morphology Suitability: The system is specifically designed for languages with rich flexional structures like Italian, where this approach offers significant advantages in terms of dictionary compression and analysis.

Key Results

The experiments, though preliminary and not fully automated, showed a "very low fraction of still incorrect groups" of forms that were not derivable from the same stem. This suggests that the process of grouping forms by stem is largely satisfactory. The system's potential lies in its ability to significantly reduce the manual effort required for morphological dictionary construction.

Limitations and Future Work

  • Partial Implementation: At the time of writing, only parts of the system (steps 1 and 2) were fully implemented, with steps 3 and 4 (filtering and expansion) under development.
  • Teletype Interaction: The current interaction method uses a teletype, which is acknowledged as slow. The authors plan to implement an alphanumeric video display with a light pen to speed up interaction.
  • Non-Automatic Steps: Steps 3 and 4, which involve filtering and widening, are not fully automatic and require operator control. While this is a strength for accuracy, it means the system is not entirely hands-off.
  • Future Work: The paper suggests that the completed system will be a valuable tool for building morphological dictionaries for specific applications. Future work will focus on completing the implementation of the interactive steps and further refining the parameters (like thresholds) that control the process.

Summary updated

BibTeX
@inproceedings{mennucci1973interactive,
  author = {M. Mennucci and E. Morreale},
  title = {{An Interactive System for Stem-Suffix Discrimination in Italian Words}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1973},
}
Figure