An Enhanced Lesk Word Sense Disambiguation Algorithm through a Distributional Semantic Model

Pierpaolo Basile, Annalina Caputo, Giovanni SemeraroCOLING 2014

TL;DR

This paper introduces an improved Lesk algorithm for Word Sense Disambiguation (WSD) that leverages distributional semantic models to better capture word meaning similarity. By representing word senses and their contexts as vectors in a semantic space, the algorithm achieves state-of-the-art results on English WSD tasks, outperforming both baseline methods and other participants.

What Problem Does This Paper Address?

Word Sense Disambiguation (WSD) is the challenging task of identifying the correct meaning of a word in a given context, especially when a word has multiple senses (e.g., "bank" can refer to a financial institution or a river's edge). Traditional knowledge-based WSD methods, like the Lesk algorithm, rely on matching words in a word's definition (gloss) with words in its context. However, these methods often struggle with the limited expressiveness of glosses and can be computationally complex. While adaptations have improved performance, they still face limitations. This paper aims to overcome these limitations by introducing a more sophisticated similarity measure that goes beyond simple word overlap, using distributional semantics to better understand semantic relationships.

Methods

The core of this work is an enhanced Lesk algorithm that replaces the traditional word overlap with a distributional semantic similarity measure. Here's a breakdown of their approach:

  • Distributional Semantic Model (DSM): They build a semantic space where words are represented as vectors. The proximity of these vectors signifies semantic relatedness. This space is constructed using Latent Semantic Analysis (LSA) on a large corpus, which captures latent semantic structures.
  • Vector Representation: Both the glosses of word senses and the surrounding context are represented as vectors. This is achieved by summing the vectors of the individual words within the gloss or context.
  • Similarity Computation: Instead of counting shared words, they compute the cosine similarity between the gloss vector and the context vector. This allows for a more nuanced measure of semantic relatedness, even if the exact words don't match.
  • Gloss Expansion: To enrich the information available for disambiguation, they expand each sense's gloss by incorporating the glosses of semantically related senses. This is done using the BabelNet API.
  • Sense Distribution: The algorithm also incorporates information about the prior probability of a sense occurring for a given word, derived from sense-tagged corpora (like SemCor). This is combined with the similarity score.
  • Term Scoring: Words within the expanded gloss are weighted. This weighting considers the distance to related senses and an Inverse Gloss Frequency (IGF) score, which prioritizes words that are more discriminative across different senses.

The paper also mentions a simplified version of their approach that reverts to word overlap (with stemming) for comparison.

Datasets

  • Sense Inventory: BabelNet is used as the sense inventory. BabelNet is a large, multilingual semantic network that integrates WordNet and Wikipedia, providing richer glosses and encyclopedic concepts.
  • Corpus for DSM:
    • English: British National Corpus (BNC)
    • Italian: Wikipedia dump
  • Evaluation Dataset: SemEval-2013 Task-12: Multilingual Word Sense Disambiguation. This task involves disambiguating all occurrences of noun phrases in arbitrary texts across different languages.

Key Insights and Ideas

  • Beyond Word Overlap: The most significant insight is that moving beyond simple word overlap to distributional semantic similarity dramatically improves WSD performance. This highlights the power of representing word meanings in a continuous semantic space.
  • BabelNet's Richness: The use of BabelNet, with its integrated Wikipedia content, provides significantly richer glosses compared to traditional resources like WordNet alone, leading to better disambiguation.
  • Gloss Expansion Matters: Expanding glosses with related senses, while carefully managing the contribution of distant senses, helps to capture more comprehensive semantic information.
  • Sense Distribution is Helpful (but not always essential): While incorporating sense distribution generally improves performance, the paper shows that even without it, their DSM-based approach can outperform strong baselines, especially for English.
  • Language Dependence: The authors note that only stemming and the corpus used for building the DSM are language-dependent, suggesting the core algorithm is largely language-agnostic.

Key Results

  • English Performance: The proposed DSM-based algorithm consistently outperforms simplified Lesk approaches. Notably, the run EN.DSM.10 (using sense distribution and the whole text as context) achieved an F-score of 0.715, outperforming all other participants in the SemEval-2013 English task. Even EN.DSM.5 (without sense distribution) surpassed the Most Frequent Sense (MFS) baseline and all task participants.
  • Italian Performance: While not reaching the same peak as English, the Italian results also show the DSM approach outperforming the MFS baseline and some task participants. IT.DSM.10 achieved an F-score of 0.641.
  • Outperforming MFS: A key achievement is that their unsupervised approach, particularly EN.DSM.5, was able to overcome the MFS baseline without using sense frequency information, which is a significant challenge in unsupervised WSD.

Limitations and Future Work

  • Gloss Quality for Italian: The authors attribute the slightly lower performance on Italian to the "poor quality of glosses" in the Italian BabelNet data, particularly from MultiWordNet.
  • Domain Adaptation: The current evaluation is on general domain texts. The authors suggest future work to adapt the approach to specific domains by building domain-specific distributional models and extracting sense frequencies from domain corpora.
  • Language Extension: While the approach is designed to be multilingual, further evaluation on a wider range of languages is planned.
  • Context Size: The optimal context size varied between English and Italian, and further exploration of context representation could be beneficial.

Summary updated

BibTeX
@inproceedings{basile2014enhanced,
  author = {Pierpaolo Basile and Annalina Caputo and Giovanni Semeraro},
  title = {{An Enhanced Lesk Word Sense Disambiguation Algorithm through a Distributional Semantic Model}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2014},
}
Figure