basciple / context_attentive_ir Goto Github PK

PyTorch code for our ICLR 2018 and SIGIR 2019 papers.

• [ICLR 2018] Multi-task Learning for Document Ranking and Query Suggestion
• [SIGIR 2019] Context Attentive Document Ranking and Query Suggestion

The codebase contains source-code of 8 document ranking models, 3 query suggestions models and 3 multi-task context-aware ranking and suggestion models.

• ESM: Embedding Space Model [See the details in our ICLR'18 paper]
• DSSM: Learning Deep Structured Semantic Models for Web Search using Clickthrough Data
• CDSSM: Learning Semantic Representations Using Convolutional Neural Networks for Web Search
• DRMM: A Deep Relevance Matching Model for Ad-hoc Retrieval
• ARCI/ARCII: Convolutional Neural Network Architectures for Matching Natural Language Sentences
• DUET: Learning to Match using Local and Distributed Representations of Text for Web Search
• MATCH TENSOR: Match-Tensor: a Deep Relevance Model for Search

• HREDQS: A Hierarchical Recurrent Encoder-Decoder for Generative Context-Aware Query Suggestion
• Seq2seq with Attention: Effective Approaches to Attention-based Neural Machine Translation
• ACG: Learning to Attend, Copy, and Generate for Session-Based Quer Suggestion

Please note, we have a simplified implementation of ACG.

• MNSRF/M-MATCH-TENSOR: Multi-task Learning for Document Ranking and Query Suggestion
• CARS: Context Attentive Document Ranking and Query Suggestion

• python 3.6
• pytorch >= 0.4 (tested on pytorch 0.4.1)
• spaCy
• tqdm
• prettytable

$ cd  scripts
$ bash SCRIPT_NAME GPU_ID MODEL_NAME

• To train/test document ranking models, use ranker.sh in place of SCRIPT_NAME
• To train/test query suggestion models, use recommender.sh in place of SCRIPT_NAME
• To train/test multitask models, use multitask.sh in place of SCRIPT_NAME

Here is a list of models which you can use in place of MODEL_NAME.

• Document Ranking Models: esm, dssm, cdssm, drmm, arci, arcii, duet, match_tensor
• Query Suggestion Models: seq2seq, hredqs, acg
• Multitask Models: mnsrf, m_match_tensor, cars

For example, if you want to run our CARS model, run the following command.

bash multitask.sh GPU_ID cars

• If GPU_ID is set to -1, CPU will be used.
• If GPU_ID is set to one specific number, only one GPU will be used.
• If GPU_ID is set to multiple numbers (e.g., 0,1,2), then parallel computing will be used.

We are unable to make our experimental dataset publicly available. However, we are sharing scripts to create an artificial dataset from MSMARCO Q&A v2.1 and MSMARCO Conversational Search datasets. Please run the script by going into the /data/msmarco/ directory. Once the data is generated, you should be able to see a table showing the following statistics.

Coming soon!

I borrowed and modified code from DrQA, OpenNMT. I would like to expresse my gratitdue for authors of these repositeries.

If you find the resources in this repo useful, please cite our works.

@inproceedings{Ahmad:2019:CAD:3331184.3331246,
 author = {Ahmad, Wasi Uddin and Chang, Kai-Wei and Wang, Hongning},
 title = {Context Attentive Document Ranking and Query Suggestion},
 booktitle = {Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval},
 year = {2019},
 pages = {385--394}
} 

@inproceedings{uddin2018multitask,
 title={Multi-Task Learning for Document Ranking and Query Suggestion},
 author={Wasi Uddin Ahmad and Kai-Wei Chang and Hongning Wang},
 booktitle={International Conference on Learning Representations},
 year={2018}
}