Skip to content
This repository has been archived by the owner on Apr 14, 2021. It is now read-only.

Latest commit

 

History

History
155 lines (141 loc) · 6.62 KB

README.md

File metadata and controls

155 lines (141 loc) · 6.62 KB

Sogou Machine Reading Comprehension Toolkit

Introduction

The Sogou Machine Reading Comprehension (SMRC) toolkit was designed for the fast and efficient development of modern machine comprehension models, including both published models and original prototypes.

Toolkit Architecture

avatar

Installation

$ git clone https://github.com/sogou/SMRCToolkit.git
$ cd SMRCToolkit
$ pip install [-e] .

Option -e makes your installation editable, i.e., it links it to your source directory

This repo was tested on Python 3 and Tensorflow 1.12

Quick Start

To train a Machine Reading Comprehension model, please follow the steps below.

For SQuAD1.0, you can download a dataset with the following commands.

$ wget https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json
$ wget https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json
$ wget https://nlp.stanford.edu/data/glove.840B.300d.zip #used in DrQA
$ unzip glove.840B.300d.zip

Prepare the dataset reader and evaluator.

train_file = data_folder + "train-v1.1.json"
dev_file = data_folder + "dev-v1.1.json"
reader = SquadReader()
train_data = reader.read(train_file)
eval_data = reader.read(dev_file)
evaluator = SquadEvaluator(dev_file)

Build a vocabulary and load the pretrained embedding.

vocab = Vocabulary(do_lowercase=False)
vocab.build_vocab(train_data + eval_data, min_word_count=3, min_char_count=10)
word_embedding = vocab.make_word_embedding(embedding_folder+"glove.840B.300d.txt")

Use the feature extractor,which is only necessary when using linguistic features.

feature_transformer = FeatureExtractor(features=['match_lemma','match_lower','pos','ner','context_tf'],
build_vocab_feature_names=set(['pos','ner']),word_counter=vocab.get_word_counter())
train_data = feature_transformer.fit_transform(dataset=train_data)
eval_data = feature_transformer.transform(dataset=eval_data)

Build a batch generator for training and evaluation,where additional features and a feature vocabulary are necessary when a linguistic feature is used.

train_batch_generator = BatchGenerator(vocab,train_data, training=True, batch_size=32, \
    additional_fields = feature_transformer.features,feature_vocab=feature_transformer.vocab)
eval_batch_generator = BatchGenerator(vocab,eval_data, batch_size=32, \
    additional_fields = feature_transformer.features, feature_vocab=feature_transformer.vocab)

Import the built-in model and compile the training operation, call functions such as train_and_evaluate for training and evaluation.

model = DrQA(vocab, word_embedding, features=feature_transformer.features,\
 feature_vocab=feature_transformer.vocab)
model.compile()
model.train_and_evaluate(train_batch_generator, eval_batch_generator, evaluator, epochs=40, eposides=2)

All of the codes are provided using built-in models running on different datasets in the examples. You can check these for details. Example of model saving and loading.

Modules

  1. data
    • vocabulary.py: Vocabulary building and word/char index mapping
    • batch_generator.py: Mapping words and tags to indices, padding length-variable features, transforming all of the features into tensors, and then batching them
  2. dataset_reader
    • squad.py: Dataset reader and evaluator (from official code) for SQuAD 1.0
    • squadv2.py : Dataset reader and evaluator (from official code) for SQuAD 2.0
    • coqa.py : Dataset reader and evaluator (from official code) for CoQA
    • cmrc.py :Dataset reader and evaluator (from official code) for CMRC
  3. examples
    • Examples for running different models, where the specified data path should provided to run the examples
  4. model
    • Base class and subclasses of models, where any model should inherit the base class
    • Built-in models such as BiDAF, DrQA, and FusionNet
  5. nn
    • similarity_function.py: Similarity functions for attention, e.g., dot_product, trilinear, and symmetric_nolinear
    • attention.py: Attention functions such as BiAttention, Trilinear and Uni-attention
    • ops: Common ops
    • recurrent: Wrappers for LSTM and GRU
    • layers: Layer base class and commonly used layers
  6. utils
    • tokenizer.py: Tokenizers that can be used for both English and Chinese
    • feature_extractor: Extracting linguistic features used in some papers, e.g., POS, NER, and Lemma
  7. libraries

Custom Model and Dataset

  • Custom models can easily be added with the description in the tutorial.
  • A new dataset can easily be supported by implementing a Custom Dataset Reader and Evaluator.

Performance

F1/EM score on SQuAD 1.0 dev set

Model toolkit implementation original paper
BiDAF 77.3/67.7 77.3/67.7
BiDAF+ELMo 81.0/72.1 -
IARNN-Word 73.9/65.2 -
IARNN-hidden 72.2/64.3 -
DrQA 78.9/69.4 78.8/69.5
DrQA+ELMO 83.1/74.4 -
R-Net 79.3/70.8 79.5/71.1
BiDAF++ 78.6/69.2 -/-
FusionNet 81.0/72.0 82.5/74.1
QANet 80.8/71.8 82.7/73.6
BERT-Base 88.3/80.6 88.5/80.8

F1/EM score on SQuAD 2.0 dev set

Model toolkit implementation original paper
BiDAF 62.7/59.7 62.6/59.8
BiDAF++ 64.3/61.8 64.8/61.9
BiDAF++ + ELMo 67.6/64.8 67.6/65.1
BERT-Base 75.9/73.0 75.1/72.0

F1 score on CoQA dev set

Model toolkit implementation original paper
BiDAF++ 71.7 69.2
BiDAF++ + ELMo 74.5 69.2
BERT-Base 78.6 -
BERT-Base+Answer Verification 79.5 -

Contact information

For help or issues using this toolkit, please submit a GitHub issue.

Citation

If you use this toolkit in your research, please use the following BibTex Entry

@ARTICLE{2019arXiv190311848W,
       author = {{Wu}, Jindou and {Yang}, Yunlun and {Deng}, Chao and {Tang}, Hongyi and
         {Wang}, Bingning and {Sun}, Haoze and {Yao}, Ting and {Zhang}, Qi},
        title = "{Sogou Machine Reading Comprehension Toolkit}",
      journal = {arXiv e-prints},
     keywords = {Computer Science - Computation and Language},
         year = "2019",
        month = "Mar",
          eid = {arXiv:1903.11848},
        pages = {arXiv:1903.11848},
archivePrefix = {arXiv},
       eprint = {1903.11848},
 primaryClass = {cs.CL},
       adsurl = {https://ui.adsabs.harvard.edu/\#abs/2019arXiv190311848W},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}

License

Apache-2.0