Automatically Merging Lexicons that have Incompatible Part-of-Speech Categories

Daniel Ka-Leung Chan, Dekai WuEMNLP 1999

Abstract

Part-of-speech (POS) lexicons are critical resources in Natural Language Processing, yet they often employ disparate and incompatible tagsets. Merging these resources manually is labor-intensive and error-prone. This paper presents an automated method for merging lexicons with different POS categories by leveraging internal lexical information and a novel "anti-lexicon" concept. The approach significantly improves merging precision from 15.9% to nearly 90%, albeit at the cost of recall.

Introduction

In the field of Natural Language Processing (NLP), lexicons serve as foundational data structures providing grammatical information about words. However, a persistent challenge arises from the lack of standardization in part-of-speech (POS) tagging. Different projects and dictionaries often utilize distinct tagsets; for instance, one might use a granular set like the Penn TreeBank (e.g., NN for singular nouns), while another uses a coarse-grained set (e.g., N for all nouns).

Example of incompatible POS tags for the word "boy" in Brill' and Moby' lexicons.

When researchers attempt to combine these resources to broaden lexical coverage, they encounter the "incompatible categories" problem. Manual mapping is unscalable, and previous automated methods typically required expensive hand-annotated corpora. This research explores whether these lexicons can be merged using only the information contained within the lexicons themselves.

Methodology

The proposed algorithm follows a multi-step process to align disparate tagsets:

1. Feature Vector Representation

Each POS tag is represented as a vector. For a given word-tag pair, the corresponding dimension in the vector is marked based on its presence in the lexicon: 1 for present, 2 for explicitly absent (never occurs), and 0 for unknown.

2. Lexicon Intersection and Anchor Words

To establish a common ground, the algorithm identifies "anchor words"—lemmas that appear in both dictionaries. By focusing on this intersection, the system reduces noise from words unique to a single source.

3. Similarity Scoring (SimScore)

The algorithm calculates a similarity score between tags from different lexicons. If tag T1 in Lexicon A and tag T2 in Lexicon B consistently appear with the same set of anchor words, they are considered likely candidates for mapping.

4. The Anti-Lexicon

The core innovation of this work is the anti-lexicon. While standard lexicons define what a word is, the anti-lexicon defines what a word cannot be. The authors calculate "cohesion," a metric determining the likelihood of a word having a specific tag based on its other known tags. Word-tag pairs with extremely low cohesion are added to the anti-lexicon. This negative information serves as a filter, pruning incorrect mapping rules that might otherwise pass a simple similarity check.

A sample lexicon structure showing lemmas and their associated POS tags.

Experimental Setup and Datasets

The method was evaluated using four major English monolingual lexicons:

  • Brill’s Lexicon: 43 tags (Penn TreeBank set).
  • Collins Lexicon: 32 tags.
  • Moby Lexicon: 15 tags (coarse-grained).
  • Oxford Machine-Readable Dictionary: 20 tags.

The researchers conducted 12 pairwise merging tasks (e.g., Brill-to-Moby) to evaluate the algorithm's robustness across varying levels of tagset granularity.

Results and Discussion

The introduction of the anti-lexicon yielded a dramatic improvement in performance. Without it, the baseline precision for mapping POS rules was a mere 15.9%. With the anti-lexicon, precision rose to 66.1%. In the final lexicon merging task, the system achieved an average precision of 89.6%, with some specific tasks reaching 100% precision.

However, this high precision comes with a trade-off in recall, which averaged 21.7%. The authors argue that in the context of lexicon construction, precision is paramount. It is significantly more efficient for a human editor to add missing entries (addressing low recall) than to manually verify and delete thousands of incorrect entries (addressing low precision).

Conclusion

This research demonstrates that "negative knowledge"—knowing what a word cannot be—is a powerful tool for aligning incompatible grammatical categories. By formalizing this intuition through the anti-lexicon, the authors provide a viable path for automated lexicon merging without the need for external annotated corpora. Future work will focus on improving recall and handling cases where one-to-one tag mappings do not exist.

Summary updated

BibTeX
@inproceedings{chan1999automatically,
  author = {Daniel Ka-Leung Chan and Dekai Wu},
  title = {{Automatically Merging Lexicons that have Incompatible Part-of-Speech Categories}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {1999},
}
Figure