NEW! Awards have been announed (see below for details).
ClariQ (pronounced as Claire-ee-que) challenge is organized as part of the Conversational AI challenge series (ConvAI3) at Search-oriented Conversational AI (SCAI) EMNLP workshop in 2020. The main aim of the conversational systems is to return an appropriate answer in response to the user requests. However, some user requests might be ambiguous. In IR settings such a situation is handled mainly through the diversification of a search result page. It is however much more challenging in dialogue settings. Hence, we aim to study the following situation for dialogue settings:
The main research questions we aim to answer as part of the challenge are the following:
The detailed description of the challenge can be found in the following document.
question_relevancehave been added. The code is also available on Google Colab.
Google is sponsoring the competition award with the GCP Credits, which we are planning to award as follows:
IMPORTANT! Please note that GCP Credits can only be granted where available and are subject to the Terms and Conditions and product availability. If the competition winner is from a region where the program is not launched, we will, unfortunately, not be able to issue the prize.
The challenge will be run in two stages:
In Stage 1, we provide to the participants the datasets that include:
To answer RQ1: Given a user request, return a score [1 −4] indicating the necessity of asking clarifying questions.
To answer RQ2: Given a user request which needs clarification, returns the most suitable clarifying question. Here participants are able to choose: (1) either select the clarifying question from the provided question bank (all clarifying questions we collected), aiming to maximize the precision, (2) or
choose not to ask any question (by choosing
Q0001 from the question bank.)
The dataset is stored in the following repository https://github.com/aliannejadi/ClariQ, together with evaluation scripts and baseline.
The TOP-5 systems from Stage 1 are exposed to real users. Their responses—answers and clarifying questions—are rated by the users. At that stage, the participating systems are put in front of human users. The systems are rated on their overall performance. At each dialog step, a system should give either a factual answer to the user’s query or ask for clarification. Therefore, the participants would need to:
The participants would need to strike a balance between asking too many questions and providing irrelevant answers.
Note that the setup of this stage is quite different from Stage 1. Participating systems would likely need to operate as a generative model, rather than a retrieval model. One option would be to cast the problem as generative from the beginning and solve the retrieval part of Stage 1, e.g., by ranking the offered candidates by their likelihood.
Alternatively, one may solve Stage 2 by retrieving a list of candidate answers (e.g., by invoking Wikipedia API or the Chat Noir API that we describe above) and ranking them as in Stage 1.
Winners will be announced at SCAI@EMNLP2020 which will take place in November 19-20 (exact details TBD).
Participants’ models will then be compared in two ways after two stages:
The winning will be chosen based on these scores.
There are three types of metrics we will evaluate:
Automated metrics As system automatic evaluation metrics we use MRR, P@[1,3,5,10,20], nDCG@[1,3,5,20]. These metrics are computed as follows: a selected clarifying question, together with its corresponding answer are added to the original request. The updated query is then used to retrieve (or re-rank) documents from the collection. The quality of a question is then evaluated by taking into account how much the question and its answer affect the performance of document retrieval. Models are also evaluated in how well they are able to rank relevant questions higher than other questions in the question bank. For this task, that we call ‘question relevance’, the models are evaluated in terms of Recall@[10,20,30]. Since the precision of models is evaluated in the document relevance task, here we focus only on recall.
Crowd workers given the entrant’s model code, we will run live experiments where Turkers chat to their model given instructions identical to the creation of the original dataset, but with new profiles, and then score its performance. Turkers will score the models between 1-5.
Please send two files per run as to
firstname.lastname@example.org, indicating your team’s name, as well as your run ID.
Each team is allowed to send a maximum of one run per week.
You’ll also need to share your GitHub repository with us. The run files should be formatted as described below.
Each run consists of two separate files:
clarification_needlabel for each topic.
Below we explain how each file should be formatted.
This file is supposed to contain a ranked list of questions per topic. The number of questions per topic could be any number, but we evaluate only the top 30 questions. We follow the traditional TREC run format. Each line of the file should be formatted as follows:
<topic_id> 0 <question_id> <ranking> <relevance_score> <run_id>
Each line represents a relevance prediction.
<relevance_score> is the relevance score that a model predicts for a given
<run_id> is a string indicating the ID of the submitted run.
<ranking> denotes the ranking of the
<topic_id>. Practically, the ranking is computed by sorting the questions for each topic by their relevance scores.
Here are some example lines:
170 0 Q00380 1 6.53252 sample_run 170 0 Q02669 2 6.42323 sample_run 170 0 Q03333 3 6.34980 sample_run 171 0 Q03775 1 4.32344 sample_run 171 0 Q00934 2 3.98838 sample_run 171 0 Q01138 3 2.34534 sample_run
This run file will be used to evaluate both question relevance and document relevance. Sample runs can found in
Also, sample Google Colab Notebookes are available. Please check ClariQ repo for more information.
This file is supposed to contain the predicted
clarification_need labels. Therefore, the file format is simply the
topic_id and the predicted label. Sample lines can be found below:
171 1 170 3 182 4
More information and example run files can be found at https://github.com/aliannejadi/ClariQ.
See https://github.com/aliannejadi/ClariQ for example baseline submissions.
You are free to use any system (e.g. PyTorch, Tensorflow, C++,..) as long as you can wrap your model for the evaluation. The top level README should tell us your team name, model name, and where the eval_ppl.py, eval_hits.py etc. files are so we can run them. Those should give the numbers on the validation set. Please also include those numbers in the README so we can check we get the same. We will then run the automatic evaluations on the hidden test set and update the leaderboard. You can submit a maximum of once per week.
We will use the same submitted code for the top performing models for computing human evaluations when the submission system is locked on September 9, 2020.