Noise-Contrastive Estimation for Answer Selection with Convolutional Neural Networks
This repo contains the Torch implementation of noise-contrastive estimation approach for answer selection in question answering with Convolutional Neural Networks, described in the following paper:
- Jinfeng Rao, Hua He, and Jimmy Lin. Noise-Contrastive Estimation for Answer Selection with Deep Neural Networks. Proceedings of the 25th ACM International on Conference on Information and Knowledge Management (CIKM 2016), pages 1913-1916.
Our model was evaluated on two standard QA datasets: TrecQA and WikiQA. On TrecQA, we achieved the best reported results at that time. Another contribution of this paper is to clarify the distinction between the raw and clean versions of the TrecQA test set.
Getting Started
1. Please install the Torch library by following instructions here: https://github.com/torch/distro
2. Checkout our repo:
git clone https://github.com/Jeffyrao/pairwise-neural-network.git
3. Using following script to download and preprocess the Glove word embedding:
$ sh fetch_and_preprocess.sh
Please make sure your python version >= 2.7, otherwise you will encounter an exception when unzip the downloaded embedding file.
4. Currently our tool only supports running on CPUs.
5. Before you run our model, please set the number of threads >= 5 for parallel processing. This is because our model need a large number of computation resource for training.
$ export OMP_NUM_THREADS=5
Running
1. There are several command line paramters to specify for running our model:
-dataset, the dataset you want to evaluate, which can be TrecQA and WikiQA.
-version, the version of TrecQA dataset, which can be raw and clean.
-neg_mode, the negative sampling strategy, 1 is for Random sampling, 2 is for MAX sampling, 3 is for MIX sampling
-num_pairs, the number of negative samples, can be any reasonable value, by default it's set to 8
2. To evaluate on the TrecQA raw dataset, with MAX sampling and number of negative pairs as 8, run:
$ th PairwiseTrainQA.lua -dataset TrecQA -version raw -neg_mode 2 -num_pairs 8
To evaluate on the TrecQA clean dataset, simply change -version to clean. Similarly, if you want to evaluate on the WikiQA dataset, change -dataset to WikiQA (don't need to set the -version). You can also change the -neg_mode and -num_pairs to select different sampling strategies or negative pairs.
3. To run the base convolutional neural network model in [2], please follow the same parameter setting:
$ th trainQA.lua -dataset TrecQA -version raw
Results
You should be able to reproduce some scores close to the numbers in tables below (-num_pairs is set to 8 by default):
1. TrecQA raw
| TrecQA raw | MAP | MRR |
|---|---|---|
| SentLevel [2] | 0.762 | 0.830 |
| Pairwise(Random) | 0.765 | 0.810 |
| Pairwise(MAX) | 0.780 | 0.835 |
| Pairwise(MIX) | 0.763 | 0.813 |
2. TrecQA clean
| TrecQA clean | MAP | MRR |
|---|---|---|
| SentLevel [2] | 0.777 | 0.836 |
| Pairwise(Random) | 0.768 | 0.831 |
| Pairwise(MAX) | 0.801 | 0.877 |
| Pairwise(MIX) | 0.798 | 0.872 |
3. WikiQA
| WikiQA | MAP | MRR |
|---|---|---|
| SentLevel [2] | 0.693 | 0.709 |
| Pairwise(Random) | 0.677 | 0.697 |
| Pairwise(MAX) | 0.681 | 0.705 |
| Pairwise(MIX) | 0.682 | 0.697 |
Though the numbers above don't outperform the base ConvNet model [2] in WikiQA dataset, still they are close. Our best scores in WikiQA dataset are 0.701(MAP), 0.718(MRR), which were obtained in the setting of MAX sampling and num_pairs as 10.
Reference
[1]. Noisy-Contrastive Estimation for Answer Selection with Deep Neural Networks, Jinfeng Rao, Hua He, Jimmy Lin, CIKM 2016
[2]. Multi-Perspective Sentence Similarity Modeling with Convolutional Neural Networks, Hua He, Kevin Gimpel, and Jimmy Lin, EMNLP 2014