This repository houses the codebase for replicating the experiments detailed in Jorge Martinez-Gil's paper on Context-Aware Semantic Similarity Measurement for Unsupervised Word Sense Disambiguation. Discover more insights and applications through our arXiv preprint and an accessible Medium article.

Word sense disambiguation (WSD) plays a pivotal role in Natural Language Processing (NLP). It involves deciphering the intended meaning of a word in a multi-sense context, which is crucial for enhancing the performance of applications like machine translation and information retrieval.

Our repository offers an innovative unsupervised approach to WSD using context-aware semantic similarity:

1. Preprocessing: Clean and prepare your text data.
2. Context Extraction: Identify the context surrounding the ambiguous word.
3. Semantic Similarity: Utilize pre-trained sentence embeddings and cosine similarity to evaluate semantic parallels.
4. Sense Selection: Choose the sense with the highest similarity score.

Included are the necessary code, pre-trained embeddings, and test data for thorough evaluation.

pip install -r requirements.txt

The CoarseWSD-20 dataset, a well-known resource for coarse-grained WSD, forms the backbone of our experiments. It encompasses 20 commonly ambiguous words.

Follow these steps to apply our method:

1. Clone this repository.
2. Install dependencies (refer to the installation section).
3. Download and position pre-trained word embeddings in the data directory.
4. Execute the script of your choice and observe the results in your console.

Evaluate our approach using the provided test data.

Unsupervised Word Sense Disambiguation (UWSD):

• python uwsd_bert.py - BERT
• python uwsd_elmo.py - ELMo
• python uwsd_use.py - Universal Sentence Encoder (USE)
• python uwsd_wmd.py - Word Mover's Distance (WMD)

Context-Aware Semantic Similarity (CASS):

• python cass-wordnet+bert.py - CASS using WordNet and BERT
• python cass-word2vec+bert.py - CASS using word2vec and BERT
• python cass-webscrapping+bert.py - CASS using webscraping and BERT

Example Scenario UWSD:

• Typed object-oriented programming languages, such as java and c++ , often do not support first-class methods --> options (island, programming language)
  • uwsd_bert: programming language
  • uwsd_elmo: programming language
  • uwsd_use: programming language
  • uwsd_wmd: programming language
  • ChatGPT-4: programming language

Example Scenario CASS:

• Vienna is a nice city situated in the center of the European continent.
  • cass-wordnet+bert: middle
  • cass-word2vec+bert: hub
  • cass-webscrapping+bert: mid
  • ChatGPT-4: middle

The summary of the results in terms of the CoarseWSD-20 dataset disambiguation is:

If you utilize our work, kindly cite us:

@inproceedings{martinez2023b,
  author    = {Jorge Martinez-Gil},
  title     = {Context-Aware Semantic Similarity Measurement for Unsupervised Word Sense Disambiguation},
  journal   = {CoRR},
  volume    = {abs/2305.03520},
  year      = {2023},
  url       = {https://arxiv.org/abs/2305.03520},
  doi       = {https://doi.org/10.48550/arXiv.2305.03520},
  eprinttype = {arXiv},
  eprint    = {2305.03520}
}

Released under the MIT License. View License.