emnlp2019lsan/lsan

Label-Specific Document Representation for Multi-Label Text Classification

This is the code for our paper "Label-Specific Document Representation for Multi-Label Text Classification"
If you make use of this code or the LSAN algorithm in your work, please cite the following paper:

@inproceedings{xiao2019label,
title={Label-Specific Document Representation for Multi-Label Text Classification},
author={Xiao, Lin and Huang, Xin and Chen, Boli and Jing, Liping},
booktitle={Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)},
pages={466--475},
year={2019}
}

Requirements: Ubuntu 16.0.4;
Python 3.6.7;
Pytorch 1.0.0;
mxnet 1.3.1;

Reproducibility: We provide the processed dataset AAPD, put them in the folder./data/

Train: python classification.py

Processed data download: https://drive.google.com/file/d/1QoqcJkZBHsDporttTxaYWOM_ExSn7-Dz/view

lsan's Issues

eur-lex training parameter

Hi Lin, here are some problems about LSAN code, need your help.

1. the code lack mask mechansim.
2. EUR-lex dataset reproduce bad, is it need more epoch training or extra parameter setting?

what's mean of d_a?

def init(self, batch_size, lstm_hid_dim, d_a, n_classes, label_embed, embeddings):

Averaging label representations

Hi,

Thanks for this work, it's an interesting idea.
I am a bit confused about the final averaging step in your forward method though, this line:

LSAN/attention/model.py

Line 84 in 1d130c5

avg_sentence_embeddings = torch.sum(doc, 1)/self.n_classes

Doesn't averaging all the label representations undo the effect of adaptive fusion? Your code comment says it's also possible to do a linear layer here instead of averaging. Can you please explain more? Just concat'ing the label representations would require the output layer to have shape (batch, num_classes * emb_dim * 2, num_classes), which would never fit on any GPU's memory for any reasonable number of classes. Maybe some max pooling step would help here? Curious to hear your thoughts.
Cheers.

Hope to get full model code

Hi ! I like your work, can you push the full model code? I check the model code and find that it incompletable.

Data

Your method is very inspiring to me.
Can you provide other data sets?
Appreciate it！

How do you get your label embedding?

The label embeddings are not randomly initialized, then how do you get them?

no import data.utils as utils

F1 too low?

I added F1 to the evaluation code and found that the model stopped at max F1 micro of 15%. Can you look at this issue?