Word Sense Disambiguation Based on Word Similarity Calculation Using Word Vector Representation from a Knowledge-based Graph

Dongsuk O, Sunjae Kwon, Kyungsun Kim, Youngjoong KoCOLING 2018

TL;DR

This paper introduces a novel Word Sense Disambiguation (WSD) method that leverages word similarity calculated from knowledge graph embeddings to select relevant contextual words. By focusing on semantically similar words, the approach significantly improves disambiguation accuracy and outperforms existing unsupervised knowledge-based WSD systems.

What Problem Does This Paper Address?

Natural language is inherently ambiguous, with many words having multiple meanings (senses). Word Sense Disambiguation (WSD) is the crucial task of identifying the correct sense of a word within its given context. While supervised WSD methods require extensive annotated data, which is costly to produce, unsupervised knowledge-based approaches offer a more scalable solution. However, existing knowledge-based methods often struggle with computational complexity and can be overwhelmed by irrelevant contextual information. This paper addresses the challenge of efficiently and accurately disambiguating word senses by proposing a method that intelligently selects context based on semantic similarity, thereby reducing noise and improving performance.

Methods

The proposed WSD system builds upon an iterative subgraph reconstruction approach. The core innovations lie in two key areas:

  • Word Similarity Calculation using Word Vectors from BabelNet: Instead of relying on traditional similarity measures, the authors propose generating word vector representations from BabelNet, a large multilingual lexicalized semantic network. They construct "semantic relational paths" by extending senses of a word by a certain depth and then use Doc2vec to generate word vectors from these paths. Cosine similarity is then used to calculate the similarity between these word vectors. This method aims to capture semantic relationships more effectively than traditional word vectors that might also include syntactic information.
  • Iterative Subgraph Construction with Context Word Selection: The system iteratively disambiguates words. For each ambiguous word, it first identifies contextual words by selecting those whose similarity to the ambiguous word exceeds a predefined threshold. This selective context selection aims to filter out less relevant words, reducing computational complexity and noise. The senses of the ambiguous word and its selected context words are then extracted from BabelNet. A subgraph is constructed by extending these senses and connecting them via semantic relationships. Finally, graph connectivity measures (specifically, Personalized Page Rank) are used to determine the most likely sense for the ambiguous word.

Datasets

The proposed WSD system was evaluated on three publicly available English WSD datasets:

  • SemEval-2007: Used as a development set for parameter tuning. It contains 414 words from BabelNet entries.
  • SemEval-2013: A test set consisting of 13 news articles with 1,931 noun words to be disambiguated.
  • SemEval-2015: A test set comprising four documents with 1,261 content words annotated for WSD and entity linking.

Key Insights and Ideas

  • Semantic Similarity is Key: The paper strongly suggests that semantic similarity between contextual words and the ambiguous word is more critical for WSD than syntactic similarity. By focusing on semantically related words, the system can better capture the true meaning.
  • Knowledge Graph Embeddings for Context Selection: Using word vectors derived from a knowledge graph (BabelNet) provides a richer representation of semantic relationships, leading to more effective selection of contextual words compared to traditional word embeddings.
  • Selective Contextualization Reduces Noise: Filtering contextual words based on a similarity threshold significantly reduces the amount of irrelevant information, leading to improved disambiguation accuracy and potentially faster processing.
  • Iterative Approach with Error Propagation: While the iterative nature of the algorithm allows for efficient subgraph construction, it also introduces a potential limitation where errors in early disambiguation steps can propagate and negatively impact subsequent steps.

Key Results

  • The proposed WSD system, "Wordsim_iter" with the "SRP2vSim" similarity measure, achieved significant improvements over the baseline "Sudoku_iter" system across all tested datasets (p-value < 0.01).
  • The system using the proposed "SRP2vSim" word similarity measure outperformed the system using "W2vSim" (based on traditional Word2vec) on SemEval-2013 and SemEval-2015 datasets.
  • On the SemEval-2013 dataset, the "Wordsim_iterSRP2vSim" system achieved a macro F1 score of 75.0%, outperforming other state-of-the-art unsupervised knowledge-based WSD systems.
  • While performance on SemEval-2015 was comparable to some top systems, it showed higher performance than other unsupervised methods.

Limitations and Future Work

  • Error Propagation: The paper acknowledges that the greedy nature of the iterative subgraph reconstruction approach can lead to mis-disambiguation due to error propagation from previously disambiguated words.
  • Multi-lingual Extension: The authors plan to extend their WSD system to be multi-lingual, researching how to generate multi-lingual word vector representations using BabelNet synset information.
  • Beam Search for Greedy Algorithm: To mitigate the drawbacks of the greedy algorithmic characteristic, they suggest exploring Beam search as a future work direction.

Summary updated

BibTeX
@inproceedings{o2018word,
  author = {Dongsuk O and Sunjae Kwon and Kyungsun Kim and Youngjoong Ko},
  title = {{Word Sense Disambiguation Based on Word Similarity Calculation Using Word Vector Representation from a Knowledge-based Graph}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
}
Figure