2019 SIGIR Context Attentive Document Ranking and Query Suggestion

[ZahraTaherikhonakdar]

Main Problem:
This paper identified users' search context to improve query suggestion and document ranking. To identify the search context, they consider users' previous queries and document clicks (users feedback) in the search log. To do this, they consider previous queries related to a search task in the search log and consider it as a search session. For example, looking for a job in computer science is a search task, this paper considers all user's queries in the search log related to a computer science job. Previous works that consider search log in QE usually define search session as a timestamp that user is looking for information.
They proposed a model named CARS- Context Attentive document Ranking and query Suggestion.

Input-Output:

phase1:
Input: a user's query + the user's past search behavior like previous queries and clicks
Out-put: the CARS encode the inputs to represent them as a search context.

Phase2:
Input: search context representation
Output: documents ranking+query suggestion

Previous Works and their Gaps:
They identified three solutions in understanding search context:
1- Data-driven solutions which use users' past queries or clicked documents to predict future documents ranking.
2- Model-driven solutions which develop predictive models to predict user's search behavior.
3- Neural network models to understand search context by distributed representation of queries.

In none of the above solutions, researchers considered a user's query and sequence of clicks in the ongoing search session to understand the search context and improve both documents ranking and query suggestion.

Result:
They used evaluation methods like MAP, MRR, NDCG to compare their model with other document ranking models. As the below figure shows the CARS model improved the ranking task based on the evaluation methods.

They also compared the model in terms of query suggestion and the performance enhance significantly

Data Set:
They used the AOL data set.

Gap of this work:
To identified users' search context this research considered users' previous clicks and queries in the search log instead of considering queries and clicks in one session in a specific timestamp which is a good point. To improve the query suggestion and document ranking performance we could use users' social information in addition to users' previous clicks and queries in the search log. Also considering query type in QE would be a good point.

Code:
https://github.com/wasiahmad/context_attentive_ir

[hosseinfani]

@ZahraTaherikhonakdar
Same comment. The summary is vague. It shows that you have no clear understanding of the work.
Also, the code is available in the first author's github

[ZahraTaherikhonakdar]

@ZahraTaherikhonakdar Same comment. The summary is vague. It shows that you have no clear understanding of the work. Also, the code is available in the first author's github

Done