Skip to content
ACL 2019: Incorporating Syntactic and Semantic Information in Word Embeddings using Graph Convolutional Networks
Python C++ Makefile
Branch: master
Clone or download
Latest commit 0e1ad47 Aug 10, 2019

Readme.md

Incorporating Syntactic and Semantic Information in Word Embeddings using Graph Convolutional Networks

Source code for ACL 2019 paper: Incorporating Syntactic and Semantic Information in Word Embeddings using Graph Convolutional Networks.

...

Overview of SynGCN: SynGCN employs Graph Convolution Network for utilizing dependency context for learning word embeddings. For each word in vocabulary, the model learns its representation by aiming to predict each word based on its dependency context encoded using GCNs. Please refer Section 5 of the paper for more details.

Dependencies

  • Compatible with TensorFlow 1.x and Python 3.x.
  • Dependencies can be installed using requirements.txt.
  • Install word-embedding-benchmarks used for evaluating learned embeddings.
    • The test and valid dataset splits used in the paper can be downloaded from this link. Replace the original ~/web_data folder with the provided one.
    • For switching between valid and test split execute python switch_evaluation_data.py -split <valid/valid>

Dataset:

  • We used Wikipedia corpus. The processed version can be downloaded from here.

  • The processed dataset includes:

    • voc2id.txt mapping of words to to their unique identifiers.

    • word2freq.txt contains frequency of words in the corpus.

    • de2id.txt mapping of dependency relations to their unique identifiers.

    • data.txt contains the entire Wikipedia corpus with each sentence of corpus stored in the following format:

      <num_words> <num_dep_rels> tok1 tok2 tok3 ... tokn dep_e1 dep_e2 .... dep_em
      • Here, num_words is the number of words and num_dep_rels denotes the number of dependency relations in the sentence.
      • tok_1, tok_2 ... is the list of tokens in the sentence and dep_e1, dep_e2 ...is the list of dependency relations where each is of form source_token|destination_token|dep_rel_label.

Training SynGCN embeddings:

  • Download the processed Wikipedia corpus (link) and extract it in ./data directory.
  • Execute make to compile the C++ code for creating batches.
  • To start training run:
    python syngcn.py -name test_embeddings -gpu 0
  • The trained embeddings will be stored in ./embeddings directory with the provided name test_embeddings .
  • Note: As reported in TensorFlow issue #13048. The current SynGCN's TF-based implementation is slow compared to Mikolov's word2vec implementation. For training SynGCN on a very large corpus might require multi-GPU or C++ based implementation.

Fine-tuning embedding using SemGCN:

...

  • Pre-trained 300-dimensional SynGCN embeddings can be downloaded from here.
  • For incorporating semantic information in given embeddings run:
    python semgcn.py -embed ./embeddings/pretrained_embed.txt 
                     -semantic synonyms -embed_dim 300 
                     -name fine_tuned_embeddings -gpu 0
  • The fine-tuned embeddings will be saved in ./embeddings directory with name fine_tuned_embeddings.

Extrinsic Evaluation:

For extrinsic evaluation of embeddings the models from the following papers were used:

Citation:

Please cite the following paper if you use this code in your work.

@inproceedings{vashishth-etal-2019-incorporating,
    title = "Incorporating Syntactic and Semantic Information in Word Embeddings using Graph Convolutional Networks",
    author = "Vashishth, Shikhar  and
      Bhandari, Manik  and
      Yadav, Prateek  and
      Rai, Piyush  and
      Bhattacharyya, Chiranjib  and
      Talukdar, Partha",
    booktitle = "Proceedings of the 57th Conference of the Association for Computational Linguistics",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-1320",
    pages = "3308--3318"
}

For any clarification, comments, or suggestions please create an issue or contact shikhar@iisc.ac.in.

You can’t perform that action at this time.