search

castorini / castor

PyTorch deep learning models for text processing
http://castor.ai/
Apache License 2.0
178 stars 58 forks source link

add DecAtt model #170

Closed Victor0118 closed 5 years ago

Victor0118 commented 5 years ago

Reference: code: https://github.com/lanwuwei/SPM_toolkit/tree/master/DecAtt paper: A Decomposable Attention Model for Natural Language Inference

@likicode @daemon Could you take a look at this PR when you are available?

Victor0118 commented 5 years ago

DecAtt after tuning on SICK dataset

==> decatt_sick_5e-4_0.001_0.5_0.1.log <== INFO - pearson_r spearman_r mse KL-divergence loss INFO - test 0.80094564 0.7184082390455326 0.3711671233177185 0.6171465432980202

daemon commented 5 years ago

Logic seems fine, just nitpicks. Do we have results for *QA now?

Victor0118 commented 5 years ago

I haven't run *QA. Might do that later.

Victor0118 commented 5 years ago

@daemon All of your comments fixed.

With only one trial, the results of DecAtt on

  1. WikiQA is 4 points better than https://arxiv.org/pdf/1806.04330.pdf
  2. TrecQA is 1 point lower than https://arxiv.org/pdf/1806.04330.pdf
daemon commented 5 years ago

Cool. There are still some extraneous comments, but I think it's in good enough shape.