This is a pytorch implementation of the ACL 2019 paper "Simple and Effective Text Matching with Richer Alignment Features". The original Tensorflow implementation: https://github.com/alibaba-edu/simple-effective-text-matching.
RE2 is a fast and strong neural architecture for general purpose text matching applications. In a text matching task, a model takes two text sequences as input and predicts their relationship. This method aims to explore what is sufficient for strong performance in these tasks. It simplifies many slow components which are previously considered as core building blocks in text matching, while keeping three key features directly available for inter-sequence alignment: original point-wise features, previous aligned features, and contextual features.
RE2 achieves performance on par with the state of the art on four benchmark datasets: SNLI, SciTail, Quora and WikiQA, across tasks of natural language inference, paraphrase identification and answer selection with no or few task-specific adaptations. It has at least 6 times faster inference speed compared to similarly performed models.
The following table lists major experiment results. The paper reports the average and standard deviation of 10 runs. Inference time (in seconds) is measured by processing a batch of 8 pairs of length 20 on Intel i7 CPUs. The computation time of POS features used by CSRAN and DIIN is not included.
Model | SNLI | SciTail | Quora | WikiQA | Inference Time |
---|---|---|---|---|---|
BiMPM | 86.9 | - | 88.2 | 0.731 | 0.05 |
ESIM | 88.0 | 70.6 | - | - | - |
DIIN | 88.0 | - | 89.1 | - | 1.79 |
CSRAN | 88.7 | 86.7 | 89.2 | - | 0.28 |
RE2 | 88.9±0.1 | 86.0±0.6 | 89.2±0.2 | 0.7618 ±0.0040 | 0.03~0.05 |
Refer to the paper for more details of the components and experiment results.
- install python >= 3.6 and pip
pip install -r requirements.txt
- install PyTorch
- Download GloVe word vectors (glove.840B.300d) to
resources/
Data used in the paper are prepared as follows:
- Download and unzip SNLI
(pre-processed by Tay et al.) to
data/orig
. - Unzip all zip files in the "data/orig/SNLI" folder. (
cd data/orig/SNLI && gunzip *.gz
) cd data && python prepare_snli.py
- Download and unzip SciTail
dataset to
data/orig
. cd data && python prepare_scitail.py
- Download and unzip Quora
dataset (pre-processed by Wang et al.) to
data/orig
. cd data && python prepare_quora.py
- Download and unzip WikiQA
to
data/orig
. cd data && python prepare_wikiqa.py
- Download and unzip evaluation scripts.
Use the
make -B
command to compile the source files inqg-emnlp07-data/eval/trec_eval-8.0
. Move the binary file "trec_eval" toresources/
.
To train a new text matching model, run the following command:
python train.py $config_file.json5
Example configuration files are provided in configs/
:
configs/main.json5
: replicate the main experiment result in the paper.configs/robustness.json5
: robustness checksconfigs/ablation.json5
: ablation study
The instructions to write your own configuration files:
[
{
name: 'exp1', // name of your experiment, can be the same across different data
__parents__: [
'default', // always put the default on top
'data/quora', // data specific configurations in `configs/data`
// 'debug', // use "debug" to quick debug your code
],
__repeat__: 5, // how may repetitions you want
blocks: 3, // other configurations for this experiment
},
// multiple configurations are executed sequentially
{
name: 'exp2', // results under the same name will be overwritten
__parents__: [
'default',
'data/quora',
],
__repeat__: 5,
blocks: 4,
}
]
To check the configurations only, use
python train.py $config_file.json5 --dry
To evaluate an existed model, use python evaluate.py $model_path $data_file
, here's an example:
python evaluate.py models/snli/benchmark/best.pt data/snli/train.txt
python evaluate.py models/snli/benchmark/best.pt data/snli/test.txt
Note that multi-GPU training is not yet supported in the pytorch implementation. A single 16G GPU is sufficient for training when blocks < 5 with hidden size 200 and batch size 512. All the results reported in the paper except the robustness checks can be reproduced with a single 16G GPU.
Please cite the ACL paper if you use RE2 in your work:
@inproceedings{yang2019simple,
title={Simple and Effective Text Matching with Richer Alignment Features},
author={Yang, Runqi and Zhang, Jianhai and Gao, Xing and Ji, Feng and Chen, Haiqing},
booktitle={Association for Computational Linguistics (ACL)},
year={2019}
}
This project is under Apache License 2.0.