2019 SIGIR Context Attentive Document Ranking and Query Suggestion

[DelaramRajaei]

Main problem
This paper dives into improving query suggestions and document rankings by leveraging users' search context. It recognizes that sometimes queries are unclear, leading to less relevant results. To tackle this, the paper suggests that looking at users' on-task search behavior can reveal the true meaning behind queries. This approach ensures that the solutions provided are custom-made for each user, taking into account their unique information needs and behaviors.

---

Related Works & Their Gaps
They found three ways to understand the search context.:

1- Data-driven solutions:
Methods driven by data that utilize users' past queries or interactions with documents(either click on them or not) to predict how future documents will be ranked. Most of the existing methods handle user on-task behavior in a makeshift way, often relying on manually created rules or extracting keywords and statistical features from past clicks or queries.

2- Model-driven solutions:
Approaches using modeling to create predictive models for guessing how users search. These methods have two main goals: 1. using lots of data and 2. improving the quality of the representation through a regularization effect.

3- Multi-task learning:
Neural network models aim to get search context by representing queries in a distributed manner. These methods merge query classification with document ranking and improve both tasks.

Gap: Many proposed solutions mainly concentrate on the ongoing search task and do not consider the sequential relationships between various retrieval tasks.

This paper focuses on modeling the connection between a user's queries and click sequence during a session by acquiring context-aware representations.

---

Proposed Method
To tackle the issue this paper proposed a context-aware neural retrieval ( Context Attentive document Ranking and query Suggestion- CARS) solution.
Steps:

1. Gather information about the user's previous searches(issued queries and clicks).
2. The user enters a new search query.
3. CARS transforms this information into a search context representation. This involves a two-level hierarchical recurrent neural network:
   • At the lower level, it encodes the queries and documents as sequences of words using bidirectional recurrent neural networks.
   • At the upper level, it captures the overall search task by adding another layer of recurrent states.
   • Attention mechanisms are applied at both levels.
4. The system then predicts the ranking of documents and provides suggestions for the next query.

---

Data Set
They used the AOL data set.

---

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