Wizard of Search Engine

Access to Information Through Conversations with Search Engines
by Pengjie Ren, Zhongkun Liu, Xiaomeng Song, Hongtao Tian, Zhumin Chen, Zhaochun Ren and Maarten de Rijke

title={Wizard of Search Engine: Access to Information Through Conversations with Search Engines},
author={Ren, Pengjie and Liu, Zhongkun and Song, Xiaomeng and Tian, Hongtao and Chen, Zhumin and Ren, Zhaochun and de Rijke, Maarten},
booktitle={Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval},

Paper summary


Task pipeline for conversational information seeking (CIS)\


Model pipeline for conversational information seeking (CIS)

In this work, we make efforts to facilitate research on conversational information seeking (CIS) from three angles: (1) We formulate a pipeline for CIS with six sub-tasks: intent detection, keyphrase extraction, action prediction, query selection, passage selection, and response generation. (2) We release a benchmark dataset, called wizard of search engine(WISE), which allows for comprehensive and in-depth research on all aspects of CIS. (3) We design a neural architecture capable of training and evaluating both jointly and separately on the six sub-tasks, and devise a pre-train/fine-tune learning scheme, that can reduce the requirements of WISE in scale by making full use of available data.

Running experiments


This code is written in PyTorch. Any version later than 1.6 is expected to work with the provided code. Please refer to the official website for an installation guide.

We recommend to use conda for installing the requirements. If you haven't installed conda yet, you can find instructions here. The steps for installing the requirements are:

  • Create a new environment

    conda create env -n WISE

    In the environment, a python version >3.6 should be used.

  • Activate the environment

    conda activate WISE
  • Install the requirements within the environment via pip:

    pip install -r requirements.txt


We use WebQA, DuReader, KdConv and DuConv datasets for pretraining. You can get them from the provided links and put them in the corresponding folders in ./data/. For example, WebQA datasets should be put in ./data/WebQA, and DuReader datasets in ./data/Dureader and so on. We use the WISE dataset to fine-tune the model, and this dataset is available in ./data/WISE. Details about the WISE dataset can be found here.


  • Run the following scripts to automatically process the pretraining datasets into the required format:
python ./Run.py --mode='data'
  • Run the following scripts sequentially:
python -m torch.distributed.launch --nproc_per_node=4 ./Run.py --mode='pretrain'
python -m torch.distributed.launch --nproc_per_node=4 ./Run.py --mode='finetune'

Note that you should select the appropriate pretrain models from the folder ./output/pretrained, and put them into ./output/pretrained_ready which is newly created by yourself before finetuning. The hyperparameters are set to the default values used in our experiments. To see an overview of all hyperparameters, please refer to ./Run.py.


  • Run the following scripts:
python -m torch.distributed.launch --nproc_per_node=4 ./Run.py --mode='infer-valid'
python -m torch.distributed.launch --nproc_per_node=4 ./Run.py --mode='eval-valid'
python -m torch.distributed.launch --nproc_per_node=4 ./Run.py --mode='infer-test'
python -m torch.distributed.launch --nproc_per_node=4 ./Run.py --mode='eval-test'