2017 CIKM Learning to Attend, Copy, and Generate for Session-Based Query Suggestion

[ZahraTaherikhonakdar]

Title : Learning to Attend, Copy, and Generate for Session-Based Query Suggestion Venue: Conference on Information and Knowledge Management Year: 2017

Main Problem: This paper used the sequence-to-sequence(seq2seq) models and proposed a new architecture to generate query suggestions. These models consider all queries in each session to suggest queries to a user. The author detects 2 main problems in existing models: 1) directly employing the seq2seq model would lead to losing information at a query level because this model considers a query like a sequence of words. 2) These models are unable to handle out-of-vocabulary words (OOV: terms that are not part of the normal lexicon found in a natural language processing environment). Moreover, these models are less likely to generate very low-frequency words.

Input: • queries in a session Output: • suggested queries

Previous Works and their Gaps:

• Tried to find within-session query pairs co-occurrences and rank suggestions based on the frequency of co-occurrence.
• Tried to first extract candidates and then employed learning-to-rank methods for ranking suggestions considering several features (similar to our BaseRanker baseline)
• Proposed a session-based method based on the Variable Memory Markov model (QVMM) to build a suffix tree to model the user query sequence for query suggestions These methods perform poorly for long-tail queries (long-tail queries: queries consist of three or more words and are rarely searched over a specific period of time ).

Contributions:

• They propose a customized sequence-to-sequence model for session-based query suggestion. The model helps to control the scope of the session from which we infer the suggested next query, which helps not only handle the noisy data but also automatically detect session boundaries.
• They propose evaluation metrics to assess the quality of the generative models for query suggestion.

Proposed Method The suggested architecture model is designed to address the previous seq2seq model that is mentioned above. They employ the seq2seq model with Attention which means this architecture enables the model to attend to the promising scope of the session. In other words, the attention mechanism is used to dynamically change the context of the session to generate the next query. They also incorporate Copy mechanism to copy the terms from the session context to improve the model by keeping terms in a query and handling the OOVs. Finally, the model contains Generater component to generate new queries. Their proposed model is called ACG.

Experiments: They used AOL dataset for training, validation, and testing their model. • Given the anchor query, i.e. the last query in the context, they first select top-20 candidate queries ranked based on the frequency of their co-appearances with the anchor query in the same session in the main training set, as the Most Popular Suggestions (MPS) • In the evaluation metrics, they use ClueWeb09 Category B corpus with over 50 million English documents as the external collection of documents. As the retrieval function, they use JS-divergence retrieval model. They further use the rank-biased overlap (RBO) [47] at level 100 to measure the agreement of two ranked lists.

Code The source code does not exist. The codebase in Context-aware Neural Information Retrieval uses the ACG as a Query Suggestion Model.

Gaps of this work • They just consider the sequence of queries in a session to generate suggested queries. By considering other attributes like time or personalization this model could be improved.

[hosseinfani]

@ZahraTaherikhonakdar Thank you. 1) This summary has a lot of exact text overlap with the paper (copy-paste) which is an instance of plagiarism. This is not acceptable at all. Please revise asap. 2) There are some terms that are used and need more explanations such as "long-tail queries" (what are these), or OOV, etc 3) The venue is not ACM 4) The code is publicly available. Please find it.

The whole idea of the summary is for you to practice on phrasing your own understanding of papers in scientific formal writing. We do not do it just to do a task!

[ZahraTaherikhonakdar]

@ZahraTaherikhonakdar Thank you.

1. This summary has a lot of exact text overlap with the paper (copy-paste) which is an instance of plagiarism. This is not acceptable at all. Please revise asap.

2. There are some terms that are used and need more explanations such as "long-tail queries" (what are these), or OOV,  etc

3. The venue is not ACM

4. The code is publicly available. Please find it.

The whole idea of the summary is for you to practice on phrasing your own understanding of papers in scientific formal writing. We do not do it just to do a task!

I'll revise it and also explain the specific terms