TL;DR
This paper introduces a novel "lemma-based" approach to Word Sense Disambiguation (WSD) for Dutch, significantly improving accuracy and reducing system complexity compared to traditional wordform-based methods. By treating all inflected forms of a word as a single unit (its lemma), the system leverages more training data per classifier, leading to more robust and accurate disambiguation.
What Problem Does This Paper Address?
Word Sense Disambiguation (WSD) is a fundamental challenge in Natural Language Processing (NLP) that aims to identify the correct meaning of a word in a given context. This is crucial for applications like machine translation, information retrieval, and document extraction. Traditional WSD systems often build individual classifiers for each ambiguous wordform. However, in morphologically rich languages like Dutch, this leads to a proliferation of classifiers, each with limited training data, resulting in lower accuracy and larger system footprints. The paper addresses this by proposing a more efficient and effective approach using lemmas.
Methods
The core of the paper's contribution lies in its lemma-based approach to WSD. Instead of creating separate classifiers for each inflected form of an ambiguous word (e.g., "loop," "loopt," "liep" for the verb "to walk"), the system groups all these forms under their common lemma (e.g., "lopen"). This allows for a single classifier to be trained on the aggregated data from all inflected forms.
The system employs a Maximum Entropy (MaxEnt) classification algorithm, which is well-suited for combining various linguistic features (lemmas, part-of-speech tags, context words) without making independence assumptions. To handle data sparseness, especially with limited training data for Dutch, the paper utilizes Gaussian priors for smoothing the MaxEnt model.
A crucial preprocessing step is lemmatization, performed using a dictionary-based lemmatizer for Dutch that leverages the CELEX lexical database. This lemmatizer is efficient, compact, and accurate, using finite state automata (FSA) for fast lookups. Part-of-speech (PoS) tagging is also performed, as Dutch WSD data is ambiguous in both senses and PoS.
Datasets
The experiments were conducted on the Dutch SENSEVAL-2 test data. This dataset is described as being ambiguous in both word senses and part-of-speech, making it a challenging benchmark for WSD systems. The training section of this dataset contains approximately 120,000 tokens and 9,300 sentences, while the test section has about 40,000 tokens and 3,000 sentences.
Key Insights and Ideas
- Lemmatization as a powerful generalization tool: Grouping inflected forms under their lemma dramatically increases the training data available for each classifier, leading to better generalization and accuracy.
- Reduced classifier complexity: The lemma-based approach significantly reduces the number of classifiers needed, making the WSD system smaller and more efficient, which is beneficial for real-world applications.
- Robustness through aggregation: By pooling data from various inflected forms, the lemma-based approach makes the system more robust, capable of classifying more ambiguous wordforms compared to the wordform-based model.
- MaxEnt's flexibility: The choice of Maximum Entropy is justified by its ability to integrate diverse linguistic features seamlessly, overcoming the limitations of models that assume feature independence.
Key Results
The lemma-based WSD system achieved a significant improvement in accuracy over the wordform-based model. Specifically:
- The lemma-based classifiers achieved an accuracy of 84.15% on ambiguous words and 92.45% overall, compared to 83.66% and 92.37% respectively for wordform classifiers.
- The number of classifiers built was drastically reduced: 669 for the lemma-based approach versus 953 for the wordform-based approach.
- The lemma-based approach was able to classify 438 unique ambiguous wordforms in the test data, compared to 410 for the wordform model.
- The paper also notes that the lemma-based system achieves comparable accuracy to a previously published Memory-Based WSD system for Dutch, but with significantly less effort (no extensive per-classifier parameter optimization).
Limitations and Future Work
The authors acknowledge that the degree of generalization through lemmatization depends on the training data. If a training corpus contains only a limited variety of inflected forms for a particular lemma, the compression rate might be lower.
Future work suggested includes:
- Exploring more specialized feature selection techniques.
- Optimizing settings for each ambiguous wordform individually, rather than using a uniform strategy.
- Testing the lemma-based approach within a multi-classifier voting scheme.
Summary updated