search

juanmc2005 / SimilarityLearning

Similarity Learning applied to Speaker Verification and Semantic Textual Similarity
12 stars 7 forks source link

Train Contrastive STS model #12

Closed juanmc2005 closed 5 years ago

juanmc2005 commented 5 years ago

Determine a threshold to build positive and negative STS pairs and train a model with contrastive loss. Validate with Spearman score (should discuss which is the best method).

juanmc2005 commented 5 years ago

Training is possible with cosine distance. Euclidean distance is giving NaN weights after a couple of optimizer steps. I'll have to fix that later.

juanmc2005 commented 5 years ago

Reopening as there are other hyperparameters and approaches to evaluate

juanmc2005 commented 5 years ago

After last meeting with @sophierosset , we are considering this experiment finished. A significance test will be done later as suggested by @saharghannay and @sophierosset