Code and documentation for the SemEval 2018 shared task Capturing Discriminative Attributes:
If you use the code from this repository please cite the following paper:
@inproceedings{taslimipoor2018wolves,
title={Wolves at SemEval-2018 Task 10: Semantic Discrimination based on Knowledge and Association},
author={Taslimipoor, Shiva and Rohanian, Omid and Ha, Le An and Corpas Pastor, Gloria and Mitkov, Ruslan},
booktitle={Proceedings of The 12th International Workshop on Semantic Evaluation},
pages={972--976},
year={2018}
}
For this shared task we developed a classification system to determine whether an attribute word can distinguish one word from another. To model semantic difference, we define a discriminative score as follows:
where w1, w2 and attr are the first, second, and third word respectively. Score is a variable function of relation between two words which we compute from different resources.
We make use of a variety of different association measures derived from huge corpora, and also pre-trained distributional semantic vectors. To augment our method with structured knowledge, we utilise a knowledge-based ontology. We use the feature set in supervised and unsupervised settings.
In summary, we have four sets of different s each used in computing
s.
In order to speed up the evaluation, in this repo we upload all the pre-computed scores in separate folders (ConceptNet, Sketch Engine, Google Ngrams, and Word2vec Similarity). In the script classify.py, we load these scores and feed them as features to the classifier.
The scripts for computing each set of scores will be uploaded in their corresponding directories.