A Multilingual Approach to Disambiguate Prepositions and Case Suffixes

Eneko Agirre, Mikel Lersundi, David Martínez

TL;DR

This paper introduces a novel, fully automatic method for disambiguating the meaning of prepositions and case suffixes by leveraging translation equivalences. By finding common interpretations between a target suffix/preposition and its translated equivalent in another language, the system significantly reduces ambiguity, demonstrating promising results for multilingual natural language understanding.

What Problem Does This Paper Address?

Understanding the precise meaning of words, especially those with multiple senses (polysemy), is a fundamental challenge in Natural Language Processing (NLP). This is particularly true for prepositions and case suffixes, which often carry subtle semantic nuances that are crucial for accurate interpretation. For instance, the English preposition "on" can indicate location ("the book is on the table"), time ("on Monday"), or a state ("on fire"). Similarly, case suffixes in languages like Basque, which are agglutinative, can function like prepositions but also mark grammatical roles. Disambiguating these elements is essential for tasks like machine translation, information extraction, and building knowledge bases. This paper focuses on the challenge of disambiguating the instrumental case suffix in Basque by comparing its translations in Spanish and English.

Methods

The core of the proposed method is a fully automatic approach that relies on translation equivalences. The process involves several key steps:

  1. Extraction of Definitions: The system first identifies occurrences of the target case suffix (in this case, the Basque instrumental suffix "-z") within definitions from a Basque monolingual dictionary.
  2. Translation Search: For each identified occurrence, the system searches for its equivalent definitions in Spanish and English monolingual dictionaries. This involves using lemmatizers and bilingual dictionaries to find translations.
  3. Preposition Extraction: From the translated definitions, the system extracts any prepositions that correspond to the original case suffix.
  4. Disambiguation via Intersection: The crucial step is to compare the possible interpretations of the Basque case suffix with the possible interpretations of its translated preposition(s) in Spanish and English. The intersection of these sets of interpretations is then taken as the most likely valid interpretation(s).

The method is designed to be general and can be applied to any preposition or case suffix, provided a table of their possible interpretations is available.

Datasets

The paper utilizes the following resources:

  • Basque Monolingual Dictionary (Euskal Hiztegia, EH): Used as the source for definitions containing the target case suffix.
  • Spanish and English Monolingual Dictionaries: Accessed online (Cambridge, Heritage, Wordsmyth for English; Colmex, Rae, Vox for Spanish) to find translations of the Basque definitions.
  • Basque/English and Basque/Spanish Bilingual Dictionaries: Used to facilitate the translation process.
  • Lemmatizers: For Basque, Spanish, and English.
  • Table of Interpretations: A manually constructed table listing possible interpretations for the Basque instrumental case suffix and its equivalent prepositions in Spanish and English.

No new benchmark datasets were introduced; the focus was on leveraging existing resources.

Key Insights and Ideas

  • Translation as a Disambiguation Tool: The paper demonstrates that translation equivalences can be a powerful, albeit simple, tool for semantic disambiguation. By looking for common ground in how different languages express a concept, we can narrow down its meaning.
  • Intersection of Interpretations: The core idea of taking the intersection of interpretation sets is elegant. It leverages the fact that if a concept is expressed in multiple languages, the shared semantic space between those expressions is likely the intended meaning.
  • Generality of the Approach: The method's independence from specific linguistic features (beyond the need for a table of interpretations) makes it highly adaptable to different languages and grammatical elements.
  • Encouraging Results with Simple Techniques: The authors highlight that even with relatively simple techniques (no deep syntactic parsing or complex semantic features), significant disambiguation can be achieved. This suggests that more sophisticated methods could yield even better results.

Key Results

The algorithm was applied to 4,004 definitions containing the Basque instrumental case suffix. Of these, 125 triples (3.1%) yielded an answer. The hand-tagged "gold standard" was used for evaluation.

  • Accuracy: The system achieved an accuracy of 94.5% on the disambiguated occurrences.
  • Ambiguity Reduction: The initial ambiguity of 7 possible readings for the instrumental case was reduced to an average of 3.1 readings.
  • Coverage: The method's coverage was low at 2.3%, meaning it could not disambiguate most occurrences. This is attributed to variations in dictionary definitions and the simple method for finding translation triples.
  • Comparison to Baseline: The algorithm's accuracy was comparable to a "most frequent baseline" (MF), which is often a strong benchmark in NLP tasks.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • Low Coverage: The primary limitation is the low coverage, stemming from the reliance on finding direct translation triples and variations in dictionary definitions.
  • Future Work: Incorporating surface syntactic analysis to better extract dependency relations and potentially shallow parsing methods are suggested to increase coverage. Exploring more infrequent translations of case suffixes could also help.
  • Errors and Ambiguity: While accuracy is high, the remaining ambiguity (3.1 readings) and sources of error (e.g., incorrect triple pairings, issues with lemmatization) indicate room for improvement.
  • Future Work: Investigating alternative sets of interpretations with coarser or finer grains, and studying the sources of residual ambiguity are planned.
  • Manual Interpretation Table: The need for a manually constructed table of interpretations is a significant upfront cost.
  • Future Work: Applying the method to all case suffixes and prepositions, and potentially using Machine Learning to construct such tables or to directly learn disambiguation rules from automatically disambiguated examples (bootstrapping).
  • Generalization: The current work is limited to definitions and a single case suffix.
  • Future Work: Extending the method to free-running text and exploring its application to other syntactic dependencies beyond prepositions and case suffixes. Coupling the method with semantic type information from the context is also proposed.

Summary updated

BibTeX
@inproceedings{agirre2002multilingual,
  author = {Eneko Agirre and Mikel Lersundi and David Martínez},
  title = {{A Multilingual Approach to Disambiguate Prepositions and Case Suffixes}},
  booktitle = {SENSEVAL},
  year = {2002},
  doi = {10.3115/1118675.1118676},
}

Loading related content...

Figure