kim-sanghoon / finding-implicatures Goto Github PK

The commit d48bf7d (Update query sentence to improve accuracy) surely improves the accuracy of implicature detection from 52% to 62%, when evaluated using VISH-evaluation dataset.

However, the major drawback of the updated query is that it significantly decreases the accuracy of diverse inputs, such as the samples in README file.

Currently, it is hard to decide which versions to use. I reverted the commit and leave it as a limitation of the current NLP model, since the model is not trained on smart-home-related datasets.