Automatic Thesaurus Generation through Multiple Filtering

Kyo Kageura, Keita Tsuji, Akiko N. AizawaCOLING 2000

TL;DR

This paper introduces a novel method for automatically generating bilingual thesauri from parallel or comparable corpora. By combining morphological/lexical processing, translation pair extraction, and graph-theoretic clustering with multiple filtering stages, the approach effectively addresses challenges like ambiguity and syntagmatic mismatches, producing promising results for domain-oriented thesaurus construction.

What Problem Does This Paper Address?

Creating high-quality bilingual thesauri, especially for specialized domains, is a challenging and labor-intensive task. Existing methods often struggle with the inherent ambiguity in language, particularly when dealing with different linguistic structures (like English and Japanese), and the difficulty of accurately identifying and filtering out incorrect translation pairs from large datasets. This paper aims to automate this process, making it more efficient and scalable, particularly for domains where parallel or comparable corpora are abundant.

Methods

The proposed method employs a three-module pipeline:

  1. Morphological and Lexical Processing: This module extracts keyword candidates separately for English and Japanese. It identifies both "minimum units" (basic words) and "maximum units" (complex phrases or multi-word expressions) to capture a richer set of terminology.
  2. Translation Pair Extraction: This stage uses statistical weighting (specifically, log-likelihood ratio) to identify potential translation pairs from the processed keywords. Crucially, it includes a "postprocessing filter" that applies both macro and micro heuristics to restrict these candidates to a more manageable and reliable set, aiming to reduce noise before clustering.
  3. Cleaning and Cluster Generation: The filtered translation pairs are used to construct a bilingual keyword graph. A graph-theoretic method, specifically focusing on minimum edge cuts, is then applied to detect and remove erroneous links. This iterative process aims to decompose the graph into independent clusters, where each cluster represents a coherent set of proper translation pairs.

Datasets

The paper utilizes Japanese and English bilingual parallel corpora. Specifically, it mentions using a corpus of 25,534 title pairs in the field of computer science for experiments. The authors note the abundance of such domain-dependent corpora for Japanese and English.

Key Insights and Ideas

  • Multi-level Unit Extraction: Extracting both single keywords and multi-word expressions (minimum and maximum units) is crucial for capturing comprehensive terminology and handling different linguistic structures.
  • Hierarchical Filtering: The multi-stage filtering process (statistical weighting followed by heuristic-based postprocessing) is key to reducing noise and improving the quality of translation candidates before graph construction.
  • Graph-theoretic Approach for Cleaning: Using minimum edge cuts on a bilingual keyword graph provides a principled way to identify and remove erroneous links, effectively cleaning up the translation pairs and generating coherent clusters.
  • Addressing Syntagmatic Mismatches: The method shows promise in handling issues like syntagmatic unit mismatches, which are common between languages like English and Japanese, by allowing for more flexible patterns in unit extraction.

Key Results

  • Quantitative Evaluation:
    • Precision tokenwise: 84.06%
    • Recall tokenwise: 87.08%
    • Precision typewise: 81.65%
    • After removing 893 pairs, 20,357 bilingual clusters were generated, with the largest cluster containing only 64 pairs (down from 2701).
  • Internal consistency evaluation showed that "more or less 80%" of the clusters were meaningful.
  • Qualitative Findings: The method successfully reduced the number of erroneous links and generated more compact, meaningful clusters. The evaluation highlighted that many errors were due to syntagmatic mismatches, particularly involving borrowed words in Katakana.

Limitations and Future Work

  • Syntagmatic Mismatch: The paper acknowledges that syntagmatic mismatches remain a significant source of errors, especially with Katakana-based borrowed words. Future work includes refining morphological processing and incorporating string proximity information to better determine syntagmatic units.
  • Statistical Alignment Errors: Errors from statistical weighting and filtering could potentially be reduced by applying stemming and orthographic normalization. The authors are also exploring iterative alignment anchoring routines and dynamic parameter adjustments for the cluster generation module.
  • Parameter Sensitivity: The current implementation uses fixed parameter values for all clusters. Future work suggests dynamically adjusting these parameters based on cluster size and pair frequencies to improve performance.
  • Domain Specificity: While the method is applied to computer science titles, future experiments are planned for narrower domains and abstracts to assess its broader applicability.
  • Applications: The authors plan to evaluate the generated thesauri in practical applications like query expansion, user navigation in cross-lingual IR, and online dictionary lookup systems.

Summary updated

BibTeX
@inproceedings{kageura2000automatic,
  author = {Kyo Kageura and Keita Tsuji and Akiko N. Aizawa},
  title = {{Automatic Thesaurus Generation through Multiple Filtering}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2000},
}
Figure