This repository contains the code associated to following paper:

• Émile Enguehard, Yoav Goldberg, and Tal Linzen, “Exploring the Syntactic Abilities of RNNs with Multi-task Learning”, in Proceedings of the 21st Conference on Computational Natural Language Learning (ConLL 2017), 2017 (link)

Summary of files

• model.py: models

• utils.py: function to read the data from previous paper

• data.py: other data-processing functions

Dependencies: Keras, pandas, inflect, numpy

Details: models

The file model.py contains the implementation of a number of single or multi-task models.

LanguageModel: Basic language model.

AgreementLM: LM + Agreement.

Supertagger: Tagging (POS, CCG, etc).

AgrSupertagger: Tagging + Agreement.

Agreement: Agreement.

---

Models accept whichever of those parameters are relevant:

• nwords: vocabulary size

• ntags: number of tags

• maxlen: maximum sentence length

• state_size: parameter D from the article

• loss_weights: weights of both losses

• id2word: mapping of integer tokens to words (typically a NP array)

• word2id: mapping of words to integer tokens (typically a dict) Beware: in the embedding layers or the LM output, the 0th dimension is for the null token. To get the embedding or probability of the i-th word you need to look at dimension i+1.

• id2tag: mapping of integer tokens to tags.

• tag2id: mapping of tags to integer tokens.

---

Data is expected to have the following numpy format:

dtype([('word', 'O'), ('pos', 'O'), ('tag', 'O'), ('subj', 'int'), ('verb', 'int')])

The column word contains string tokens, the column pos contains Penn Treebank POS tags in string format, the column tag contains target tags in string format, the columns subj and verb have only zeroes except at the position of the subject and the verb respectively.

Note that models do not actually use most columns in most cases.

---

Models support the following functions, calqued on Keras functions of the same name:

• predict

• fit

• evaluate

They take the data in a list or NP array as first parameter, and a parameter batch_size. There is also a parameter nb_worker which you should set to 1 to avoid weird bugs. fit accepts both training and testing data in that order, and it also takes the arguments nb_epoch (maximum number of training epochs), early_stopping (equivalent to the Keras callback of the same name), verbose (also transferred to Keras).

Models can be saved and loaded back to folders using the save method and the load_base_model function.

Details: data

The function utils.deps_from_tsv is designed to read the following file provided here (which is were utils.py is from as well).

In the file data.py, you'll find among others the following functions:

• tsv_to_numpy(data): transforms the output of deps_to_tsv into our format (as a list of numpy arrays)

• extract_ccg(folder): if folder is the "data" subfolder of the ccg-bank, extract the sentences and the CCG tags. Returns a list of dict's of numpy arrays (sentences are organised in sections and then by ID).

• apply_threshold_pos(data, thres): removes words less frequent than thres (int) and replaces them with their POS tag. Returns a pair of the new data and the new lexicon (as a set).

• apply_threshold_void(data, thres, key='tag'): same as above but replaces infrequent words with '_' and can apply to any key.

• apply_length_threshold(data, thres): remove sentences longer than thres.

• build_dicts(vocab): returns a pair of an int-to-word (np.array) and a word-to-int (dict) token mapping from a lexicon (set or other iterable).