Analyzing Neural MT Representations

This repository contains the code for our paper on analyzing morphology in neural machine translation models:

"What do Neural Machine Translation Models Learn about Morphology?", Yonatan Belinkov, Nadir Durrani, Fahim Dalvi, Hassan Sajjad, and James Glass, ACL 2017.

Requirements

Torch - http://torch.ch/docs/getting-started.html

Instructions

Run th src/train.lua with the following mandatory arguments:

• model: NMT model trained with seq2seq-attn
• src_dict: source dictionary used for training the NMT model
• targ_dict: target dictionary used for training the NMT model
• cl_train_lbl_file: train file with gold labels
• cl_val_lbl_file: validation file with gold labels
• cl_test_lbl_file: test file with gold labels
• cl_train_source_file: train file with sentences in the source language
• cl_val_source_file: validation file with sentences in the source language
• cl_test_source_file: test file with sentences in the source language
• cl_save: Path to folder where experiment will be saved
• cl_pred_file: Prefix to save prediction files (should be base name, not full path)

Files with labels and sentences (cl_train/val/test_lbl/source_file) should have one sentence per line, with space-delimited words or labels.

For more options, see src/s2sa/beam.lua

Evaluating saved models

Once you've run train.lua and have some models saved, you can run th src/eval.lua with the following arguments to evaluate a model:

• model: NMT model trained with seq2seq-attn to extract embeddings
• src_dict: source dictionary used for training the NMT model
• targ_dict: target dictionary used for training the NMT model
• cl_clf_model: pretrained classifier model to load (trained with train.lua)
• cl_train_lbl_file: train file with gold labels (Used to extract labels and create label-index mapping)
• cl_test_lbl_file: test file with gold labels
• cl_test_source_file: test file with sentences in the source language
• cl_save: Path where the test logs will be saved
• cl_pred_file: Prefix to save prediction files (should be base name, not full path)
• cl_enc_layer: Which layer in the encoder/decoder to use for word representation (0 for word embeddings)

Optional command line arguments

Here are some optional command line arguments that are useful:

• cl_avg_reps: Whether to average the hidden representations states for the sentences. Otherwise we use the right-most hidden state from the forward LSTM encoder and the left most from the backwards LSTM encoder.
• cl_entailment: Whether to train the classifier to classifier sentence pairs for textual entailment
• cl_enc_layer: Which hidden layer to extract the hidden states from. This command line argument needs to be followed by an integer. By defualt, the argument is set to 2. Warning: the integer must be less than or equal to the number of hidden states in the NMT model, otherwise you will run into a error related to indexing out of bounds.
• cl_inferSent_reps: Whether to use the sentence representations as described in Conneau (EMNLP '17)'s InferSet, otherwise the first and last hidden representations will be used and then the sentence reps. will be concatenated.

Citing

If you use this code in your work, please consider citing our paper:

@InProceedings{belinkov:2017:acl,
  author    = {Belinkov, Yonatan  and  Durrani, Nadir and Dalvi, Fahim and Sajjad, Hassan and Glass, James},
  title     = {What do Neural Machine Translation Models Learn about Morphology?},
  booktitle = {Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (ACL)},
  month     = {July},
  year      = {2017},
  address   = {Vancouver},
  publisher = {Association for Computational Linguistics},
}

If you use this code for NLI please cite the following paper as well:

@inproceedings{evaluating-fine-grained-semantic-phenomena-in-neural-machine-translation-encoders-using-entailment,
  author = {Poliak, Adam and Belinkov, Yonatan and Glass, James and {Van Durme}, Benjamin},
  title = {On the Evaluation of Semantic Phenomena in Neural Machine Translation Using Natural Language Inference},
  year = {2018},
  booktitle = {Proceedings of the Annual Meeting of the North American Association of Computational Linguistics (NAACL)}
}

Acknowledgements

This project uses code from seq2seq-attn.