This is a TensorFlow implementation accompanying our paper

Lajanugen Logeswaran, Honglak Lee, An efficient framework for learning sentence representations. In ICLR, 2018.

This codebase is based on Chris Shallue's Tensorflow implementation of the SkipThought model. The data preparation, vocabulary expansion and evaluation scripts have been adopted with minor changes. Other code files have been modified and re-structured with changes specific to our model.

• Model configuration files
• Pretrained Models
• Training a Model
• Evaluating a Model

We use json configuration files to describe models. These configuration files provide a concise description of a model. They also make it easy to concatenate representations from different models/types of models at evaluation time.

The description of a sentence encoder has the following format.

{
        "encoder": "gru",                            # Type of encoder
        "encoder_dim": 1200,                         # Dimensionality of encoder
        "bidir": true,                               # Uni/bi directional
        "checkpoint_path": "",                       # Path to checkpoint
        "vocab_configs": [                           # Configuration of vocabulary/word embeddings
        {
                "mode": "trained",                   # Vocabulary mode: fixed/trained/expand
                "name": "word_embedding",
                "dim": 620,                          # Word embedding size
                "size": 50001,                       # Size of vocabulary
                "vocab_file": "BC_dictionary.txt",   # Dictionary file
                "embs_file": ""                      # Provide external embeddings file
        }
        ]
}

Vocabulary mode can be one of fixed, trained or expand. These modes represent the following cases.

• fixed - Use fixed, pre-trained embeddings.
• trained - Train word embeddings from scratch.
• expand - Use an expanded vocabulary. This mode is only used during evaluation on downstream tasks.

checkpoint_path and vocab_file have to be specified only for evaluation.

For concatenating representations from multiple sentence encoders at evaluation time, the json file can be a list of multiple encoder specifications. See model_configs/BC/eval.json for an example.

Models trained on the BookCorpus and UMBC datasets can be downloaded from https://bit.ly/2uttm2j. These models are the multi-channel variations (MC-QT) discussed in the paper. If you are interested in evaluating these models or using them in your tasks, jump to Evaluation on downstream tasks.

The training script requires data to be in (sharded) TFRecord format. scripts/data_prep.sh can be used to generate these files. The script requires a dictionary file and comma-separated paths to files containing tokenized sentences.

• The dictionary file should have a single word in each line. We assume that the first token ("<unk>") represets OOV words.
• The data files are expected to have a tokenized sentence in each line, in the same order as the source document.

The following datasets were used for training out models.

• BookCorpus
• UMBC

The dictionary files we used for training our models are available at https://bit.ly/2G6E14q.

Use the run.sh script to train a model. The following variables have to be specified.

* DATA_DIR      # Path to TFRecord files
* RESULTS_HOME  # Directory to store results
* CFG           # Name of model configuration 
* MDL_CFGS      # Path to model configuration files
* GLOVE_PATH    # Path to GloVe dictionary and embeddings

Example configuration files are provided in the model_configs folder. During training, model files will be stored under a directory named $RESULTS_HOME/$CFG.

The implementation supports using fixed pre-trained GloVe word embeddings. The code expects a numpy array file consisting of the GloVe word embeddings named glove.840B.300d.npy in the $GLOVE_PATH folder.

Once the model is trained, the vocabulary used for training can be optionally expanded to a larger vocabulary using the technique proposed by the SkipThought paper. The voc_exp.sh script can be used to perform expansion. Since Word2Vec embeddings are used for expansion, you will have to download the Word2Vec model. You will also need the gensim library to run the script.

Use the eval.sh script for evaluation. The following variables need to be set.

* SKIPTHOUGHTS  # Path to SkipThoughts implementation
* DATA          # Data directory for downstream tasks
* TASK          # Name of the task
* MDLS_PATH     # Path to model files
* MDL_CFGS      # Path to model configuration files
* CFG           # Name of model configuration 
* GLOVE_PATH    # Path to GloVe dictionary and embeddings

Evaluation scripts for the downstream tasks from the authors of the SkipThought model are used. These scripts train a linear layer on top of the sentence embeddings for each task. You will need to clone or download the skip-thoughts GitHub repository by ryankiros. Set the DATA variable to the directory containing data for the downstream tasks. See the above repository for further details regarding downloading and setting up the data.

To evaluate the pre-trained models, set the directory variables appropriately. Set MDLS_PATH to the directory of downloaded models. Set the configuration variable CFG to one of

• MC-BC (Multi-channel BookCorpus model) or
• MC-UMBC (Multi-channel BookCorpus UMBC model)

Set the TASK variable to the task of interest.

If you found our code useful, please cite us 1.

@inproceedings{
logeswaran2018an,
  title={An efficient framework for learning sentence representations},
  author={Lajanugen Logeswaran and Honglak Lee},
  booktitle={International Conference on Learning Representations},
  year={2018},
  url={https://openreview.net/forum?id=rJvJXZb0W},
}

Contact: [email protected]