2021 WWW Towards a Better Understanding of Query Reformulation Behavior in Web Search

**Main Problem:**

This paper analyzes users behavior in web search to answer the key questions:

1: _How do users query reformulating behavior change through the session?_--> They consider the following aspects to answering this question:1. reformulation type, syntactic(e.g. adding/deleting/change/repeat term) and intent-level (new topic, somewhat relevant). 2. reason for reformulation(e.g. satisfaction and dissatisfaction). 3. reformulation entry/from which part of search engine result pages (SERPs), a user chooses to reformulate his query(e.g. query suggestion, input box). 4. what inspired a user to formulate his query (e.g. landing pages or search snippets)

2: _Whether users reformulate their queries differently on different search intents?_--> They consider the following factors: 1. search effort and gain (e.g. dwell time on SERPs, clicks on search results, task satisfaction). 2. user actions (e.g. avg clicks by query, scrolling time)

3- Can we predict the answers to two previous questions?

**Method**
They developed a toolkit for browsers to selected users' search activity for one month (implicit information ). Also, they used questionnaires to collect explicit data from users. 50 users participate in this experiment.

**Input-Output:**

phase1:
Input: Implicit and explicit information related to users web search behavior
Output: Answers of two first questions + predicting model for predicting why and how users reformulate their queries.

**Previous Works and their Gaps:**
previous works proposed query suggestion and auto-completion to predict the user's next query. However, this work couldn't identify factor effect on users' query reformulation. This paper, use both explicit and implicit data of users' web search behavior and collect real data to identify what factors affect users' query reformulation.

**Result:**
They consider three dimensions:
- Intent Trigger--> Interest-driven, Interest and Task-driven, Task-driven (4);
- Intent Specificity: Clear, Board;
- Domain Expertise: Familiar, Unfamiliar.

Question1: In the beginning, users tend to add constraints on the query to limit the search scope. When their search intent is satisfied, they try to start to shift their intent gradually. Another result is, users rely on themselves to reformulate the query on long sessions, the reasons could be: search engines do not guide users properly or existing query suggestions could not satisfy users.

Question2: users' search behavior is affected by users' intent trigger and specificity. Users with interest-driven intents show more satisfaction. On the other hand, in task-driven the search effort is high but the search satisfaction is low. This indicates that query suggestion methods are not helpful in this domain and it is better for QE methods to be personalized to understand the user's search intent better.

question3: they used different predicting models to predict users' search behavior and they conclude that the best model to predict is XGBoost.
**Data Set:**
They used the AOL data set.

Gap of this work:
They could consider search history to understand the user's search behavior better.
This work can be considered as a motivation to personalized QR as the search engines have poor performance in task-driven users' intent.

Code:
http://www.thuir.cn/tiangong-qref/
toolkit: https://github.com/xuanyuan14/Web-Search-Field-Study-Toolkit