neural-sentence-matching

This repo contains Theano implementations of our original models and the models described in the following papers:

Semi-supervised Question Retrieval with Gated Convolutions. NAACL 2016

These codes are based on the following codes:

Recurrent & convolutional neural network modules

Similar Question Retrieval

These neural methods are used to calculate text similarity, for applications such as similar question retrieval in community-based QA forums.

Data

The data can be downloaded at this repo.

Dependencies

To run the code, you need the following extra packages installed:

• PrettyTable
• Numpy and Theano

Usage

1. Clone this repo
2. Run python -m sim_q_ranking.main.main --help to see all running options

Example Comand

• Basic Model: python -m sim_q_ranking.main.main --corpus path/to/data/text_tokenized.txt.gz --embeddings path/to/data/vector/vectors_pruned.200.txt.gz --train path/to/data/train_random.txt --dev path/to/data/dev.txt --test path/to/data/test.txt --layer rcnn
• Attention Model: python -m sim_q_ranking.main.main --corpus path/to/data/text_tokenized.txt.gz --embeddings path/to/data/vector/vectors_pruned.200.txt.gz --train path/to/data/train_random.txt --dev path/to/data/dev.txt --test path/to/data/test.txt --layer rcnn --model attention
• Decomposition Attention Model: python -m sim_q_ranking.main.main --corpus path/to/data/text_tokenized.txt.gz --embeddings path/to/data/vector/vectors_pruned.200.txt.gz --train path/to/data/train_random.txt --dev path/to/data/dev.txt --test path/to/data/test.txt --layer rcnn --model attention --model_type 1
• Alignment Model: python -m sim_q_ranking.main.main --corpus path/to/data/text_tokenized.txt.gz --embeddings path/to/data/vector/vectors_pruned.200.txt.gz --train path/to/data/train_random.txt --dev path/to/data/dev.txt --test path/to/data/test.txt --layer rcnn --model alignment