Post-Specialisation

Post-Specialisation: Retrofitting Vectors of Words Unseen in Lexical Resources (Vulić et al., NAACL-HLT 2018)

This repository contains the code and data for the post-specialisation method in the NAACL-HLT 2018 paper. The method is implemented in Keras (Python 2.7)

Contact: Ivan Vulić (iv250@cam.ac.uk)

Configuring the Tool

The post-specialisation tool reads all the experiment config parameters from the experiment_parameters.cfg file in the root directory. An alternative config file can be provided as the first (and only) argument to code/post-specialisation.py.

Note that the tool assumes that input distributional vectors have already been initially specialised with one of the standard specialisation models such as retrofitting (Faruqui et al., NAACL-HLT 2015), counter-fitting (Mrkšić et al., NAACL-HLT 2016), or Attract-Repel (Mrkšić et al., TACL 2017). In our experiments, we use the most recent and state-of-the-art post-processor Attract-Repel, but the post-specialisation model is equally applicable to any other post-processor. If you use any of these tools, please cite the corresponding paper(s).

The config file specifies the following:

1. The location of the initial word vectors (distributional_vectors)

• In the default setup, we use SGNS vectors with bag-of-words contexts trained on Wikipedia, available here

• the location of the training data; training data contains word vectors (x_i, x_o) changed by the initial specialisation (i.e., seen words x)

2. We have to specify the location of the distributional vectors (x_i) in distributional_training_data as well as the location of the specialised vectors (x_o) in specialised_training_data

• The two training data files follow the standard format for word vectors (word dim_1 dim_2 ... dim_N), but note that they have to contain representations of exactly the same words and have exactly the same number of items.
• We have provided toy sample training data files containing 5000 training pairs to illustrate the data format.

3. The config file also specifies the hyperparameters of the post-specialisation procedure (set to their default values in config/experiment_parameters.cfg).

• Right now, we provide support in the config file only for the best-performing max-margin loss function, but it should be fairly easy to customise the objective function using some of the Keras pre-built losses.

Running Experiments

python code/post-specialisation.py config/experiment_parameters.cfg

Running the experiment loads the word vectors specified in the config file and learns the mapping/regression function using a deep feed-forward network as specified in the config file. The procedure prints the updated word vectors to the results directory as results/final_vectors.txt (standard format: one word vector per line)

References

The paper which introduces the Post-Specialisation procedure and the problem of specialising unseen words:

 @inproceedings{Vulic:2018,
  author    = {Ivan Vuli\'{c} and Goran Glava\v{s} and Nikola Mrk\v{s}i\'c} and Anna Korhonen,
  title     = {Post-Specialisation: Retrofitting Vectors of Words Unseen in Lexical Resources},
  booktitle   = {Proceedings of NAACL-HLT},
  year      = 2018,
 }