yangzw/seqmatchseq

SeqMatchSeq

Implementations of three models described in the three papers related to sequence matching:

• Learning Natural Language Inference with Lstm by Shuohang Wang, Jing Jiang

• Machine Comprehension Using Match-LSTM and Answer Pointer by Shuohang Wang, Jing Jiang

• A Compare-Aggregate Model for Matching Text Sequences by Shuohang Wang, Jing Jiang

Learning Natural Language Inference with Lstm

Requirements

• Torch7
• nn
• nngraph
• optim
• Python 2.7

Datasets

• The Stanford Natural Language Inference (SNLI) Corpus
• GloVe: Global Vectors for Word Representation

Usage

sh preprocess.sh snli
cd main
th main.lua -task snli -model mLSTM -dropoutP 0.3 -num_classes 3

sh preprocess.sh snli will download the datasets and preprocess the SNLI corpus into the files (train.txt dev.txt test.txt) under the path "data/snli/sequence" with the format:

sequence1(premise) \t sequence2(hypothesis) \t label(from 1 to num_classes) \n

main.lua will first initialize the preprossed data and word embeddings into a Torch format and then run the alogrithm. "dropoutP" is the main prarameter we tuned.

Docker

You may try to use Docker for running the code.

• Docker Install
• Image: docker pull shuohang/seqmatchseq:1.0

After installation, just run the following codes (/PATH/SeqMatchSeq need to change):

docker run -it -v /PATH/SeqMatchSeq:/opt --rm -w /opt      shuohang/seqmatchseq:1.0 /bin/bash -c "sh preprocess.sh snli"
docker run -it -v /PATH/SeqMatchSeq:/opt --rm -w /opt/main shuohang/seqmatchseq:1.0 /bin/bash -c "th main.lua"

Machine Comprehension Using Match-LSTM and Answer Pointer

Requirements

• Torch7
• nn
• nngraph
• optim
• parallel
• Python 2.7
• Python Packages: NLTK, collections, json, argparse
• NLTK Data: punkt
• Multiple-cores CPU

Datasets

• Stanford Question Answering Dataset (SQuAD)
• GloVe: Global Vectors for Word Representation

Usage

sh preprocess.sh squad
cd main
th mainDt.lua 

sh preprocess.sh squad will download the datasets and preprocess the SQuAD corpus into the files (train.txt dev.txt) under the path "data/squad/sequence" with the format:

sequence1(Doument) \t sequence2(Question) \t sequence of the positions where the answer appear in Document (e.g. 3 4 5 6) \n

mainDt.lua will first initialize the preprossed data and word embeddings into a Torch format and then run the alogrithm. As this code is run through multiple CPU cores, the initial parameters are written in the file "main/init.lua".

• opt.num_processes: 5. The number of threads used.
• opt.batch_size : 6. Batch size for each thread. (Then the mini_batch would be 5*6 .)
• opt.model : boundaryMPtr / sequenceMPtr

Docker

You may try to use Docker for running the code.

• Docker Install
• Image: docker pull shuohang/seqmatchseq:1.0

After installation, just run the following codes (/PATH/SeqMatchSeq need to change):

docker run -it -v /PATH/SeqMatchSeq:/opt --rm -w /opt      shuohang/seqmatchseq:1.0 /bin/bash -c "sh preprocess.sh squad"
docker run -it -v /PATH/SeqMatchSeq:/opt --rm -w /opt/main shuohang/seqmatchseq:1.0 /bin/bash -c "th mainDt.lua"

A Compare-Aggregate Model for Matching Text Sequences

Requirements

• Torch7
• nn
• nngraph
• optim
• Python 2.7

Datasets

• The Stanford Natural Language Inference (SNLI) Corpus
• MovieQA: Story Understanding Benchmark
• InsuranceQA Corpus V1: Answer Selection Task
• WikiQA: A Challenge Dataset for Open-Domain Question Answering
• GloVe: Global Vectors for Word Representation

For now, this code only support SNLI and WikiQA data sets.

Usage

SNLI task (The preprocessed format follows the previous description):

sh preprocess.sh snli
cd main
th main.lua -task snli -model compAggSNLI -comp_type submul -learning_rate 0.002 -mem_dim 150 -dropoutP 0.3 

WikiQA task:

sh preprocess.sh wikiqa (Please first dowload the file "WikiQACorpus.zip" to the path SeqMatchSeq/data/wikiqa/ through address: https://www.microsoft.com/en-us/download/details.aspx?id=52419)
cd main
th main.lua -task wikiqa -model compAggWikiqa -comp_type mul -learning_rate 0.004 -dropoutP 0.04 -batch_size 10 -mem_dim 150 

• model (model name) : compAggSNLI / compAggWikiqa
• comp_type (8 different types of word comparison): submul / sub / mul / weightsub / weightmul / bilinear / concate / cos

Docker

You may try to use Docker for running the code.

• Docker Install
• Image: docker pull shuohang/seqmatchseq:1.0

After installation, just run the following codes (/PATH/SeqMatchSeq need to change):

For SNLI:

docker run -it -v /PATH/SeqMatchSeq:/opt --rm -w /opt      shuohang/transition:1.1 /bin/bash -c "sh preprocess.sh snli"
docker run -it -v /PATH/SeqMatchSeq:/opt --rm -w /opt/main shuohang/transition:1.1 /bin/bash -c "th main.lua -task snli -model compAggSNLI -comp_type submul -learning_rate 0.002 -mem_dim 150 -dropoutP 0.3"

For WikiQA

docker run -it -v /PATH/SeqMatchSeq:/opt --rm -w /opt      shuohang/seqmatchseq:1.0 /bin/bash -c "sh preprocess.sh wikiqa"
docker run -it -v /PATH/SeqMatchSeq:/opt --rm -w /opt/main shuohang/seqmatchseq:1.0 /bin/bash -c "th main.lua -task wikiqa -model compAggWikiqa -comp_type mul -learning_rate 0.004 -dropoutP 0.04 -batch_size 10 -mem_dim 150"

Copyright

Copyright 2015 Singapore Management University (SMU). All Rights Reserved.