aakanksha19/odette

Open Domain EvenT Trigger Extractor (ODETTE) using ADA (Adversarial Domain Adaptation)

odette's Issues

What does the arg '10000' stand for?

Hi, would you check Line.275 in da_tester_new.py? I don't understand what the arg '10000' stand for?
model = AdversarialEventExtractor(10000, train_batches[0][0].size()[-1], args.hidden_size, 1, args.adv_size, args.adv_layers, args.num_domains, args.adv_coeff, args.dropout, args.bidir, args.model)

'10000' is clearly the vocab size. But why the vocab size is not len(list(sent_vocab.keys()))?

Thanks!

Question about Transfer from Litbank to Timebank

Hi, can we just use the current code for transfer from litbank to timebank? My concern is the proportion of source/target data in adv_batches. With timebank/litbank as source/target domain, the proportion is 1:1. However, train set of litbank has 5263 sentences, much more than train set of timebank which has 1531 sentences. The following setting doesn't guarantee a 1:1 proportion

adv_sents = unlabeled_sents[:len(labeled_sents)] + labeled_sents
adv_parse = unlabeled_parse[:len(labeled_sents)] + labeled_parse
adv_domains = unlabeled_domains[:len(labeled_sents)] + labeled_domains

Thanks!

Why not use BERT Fine-Tune directly?

Hi, I noticed that you only use BERT's contextual representation as input features for Bi-LSTM. Why didn't you use BERT fine-tune directly? Would you explain the motivation behind your technical design?

Thank you.

Code Bugs

Thanks for your open-sourced project about this work.

However, when running the code, I find some bugs to fix. For example, checkout Line.234 in tester.py, the declaration about model is incorrect and should be modified as follow

model = EventExtractor(len(list(sent_vocab.keys())), args.emb_size, args.hidden_size, 1, args.dropout, args.bidir, args.model, pos_vocab_size=len(list(pos_vocab.keys())))

otherwise the code fails because len(list(pos_vocab.keys()))) is recognized as argument for num_layers and the pos_vocab_size is None.