Closed ZahraTaherikhonakdar closed 2 years ago
@ZahraTaherikhonakdar thank you for the summary. About the gaps of this work, they consider the time, right? so they can capture the constant change to the user's interest. Don't you think so?
@ZahraTaherikhonakdar You answered my comment in our meeting but you were supposed to discuss it here too. Any update?
@ZahraTaherikhonakdar You answered my comment in our meeting but you were supposed to discuss it here too. Any update?
Done @hosseinfani
@ZahraTaherikhonakdar Please explain how time info has been used in this paper.
@ZahraTaherikhonakdar Please explain how time info has been used in this paper.
Done @hosseinfani
Main Problem: This paper proposed new approaches in neural Query Auto Completion (QAC) to address the limitations of previous methods. This paper integrated personalization and time-sensitive information in neural QAC to address the inability of the previous methods to predict unseen queries.
Input: • concatenation of these to vectors: a vector of a user's previous query search (personalization) and a vector of the time of the query Output: • suggested queries
Previous Works and their Gaps: Previous neural auto-completion models: semantic models use Convolutional Neural Nets (CNN) and language models using Recurrent Neural Nets (RNN), are disabled to handle unseen queries.
Proposed Method They suggested queries by using a concatenation of a vector of a user's previous query search (personalization) and a vector of the time of the query as an input for the neural language model. They used AOL data set and biochemical dataset( to test the model on a specific area as unseen queries)
Results: A user's previous queries as user information made improvement for seen queries (+ 23%). Time sensitivity helps unseen queries (+21%) however hurting the quality for seen queries (-47%).
Code https://github.com/ncbi-nlp/NQAC
Gaps of this work This paper used a user's search history for personalization which is not a reliable source for personalization because a user's interests could change constantly. The paper considered time to obtain the quantity of query occurrence. Based on this research, queries about whether or sports have been searched by a user at 10:00 pm (at night time). In order to effectively use time in personalized query suggestions, we should consider the semantic change of words over time together with a user's profile. To be more specific, assume a user whose job or interest is diet or agriculture based on his/her profile. The user search for "Apple". The retrieved documents would be about "Apple as a fruit" (assume that Apple Company didn't come to existence yet). After some years, the user major in Computer Science or become interested in technology. Considering the user's profile and the Apple company establishment, retrieved documents would be about Apple company products. In this perspective, we consider the semantic of the words over time.