MeSHProbeNet

MeSHProbeNet: a self-attentive probe net for MeSH indexing

Prerequisites

• python==3.6.3
• pytorch==1.2.0
• torchtext==0.2.1
• numpy==1.16.2
• scipy==1.2.1

Input data format

Take ./toy_data/ as an example.

• train.tsv: The training set, where each line is a document. Each document is represented as content word ids separated by spaces + '\t' + journal id + '\t' + MeSH ids separated by spaces
• validation.tsv: The validation set in the same format as train.tsv
• vocab_w.txt: The vocabulary file for context words, where each line is content word id + '\t' + content word
• vocab_j.txt: The vocabulary file for journal names, where each line is journal id + '\t' + journal name
• vocab_m.txt: The vocabulary file for MeSH terms, where each line is MeSH id + '\t' + MeSH term

Validation is optional. Vocabulary id 0 is reserved for the padding token.

Run

Run on the toy data

python main_train.py \
  --do_save \
  --do_eval \
  --train_path ./toy_data/train.tsv \
  --dev_path ./toy_data/validation.tsv \
  --src_vocab_pt ./toy_data/vocab_w.txt \
  --jrnl_vocab_pt ./toy_data/vocab_j.txt \
  --tgt_vocab_pt ./toy_data/vocab_m.txt \
  --expt_path ./toy_data/save \
  --learning_rate 0.0025 \
  --weight_decay 5e-10