Research On Key Techniques Of Entity Recommendation In Web Search Engines

Search engine is an important tool for people to find information on Internet.Over the past few years,to better meet users' information needs,in addition to the returned search results in response to a query,major commercial Web search engines have enriched and improved user experience by proactively providing related information suggestions to the query.Entity recommendation,providing suggestions at entity level,is capable of providing suggestions with the richest information at the finest level of granularity amongst all the suggestions.Entity recommendation aims to provide a list of related entities that have direct or indirect connections with a given query,which can help users to broaden their knowledge and has been more and more valued by users.Therefore,entity recommendation has not only become an indispensable feature of today's Web search engine,but has also attracted increasing attention from industry and researchers.The building of an entity recommendation system in Web search engines typically involves:(1)providing entity recommendations related to a user's query;and(2)providing plausible explanations for recommendations to make them easily understood by the user.To address these two sub-tasks,the following two directions should be studied:(1)entity recommendation,which aims to find a set of related entities to a query,and then rank these entities;and(2)recommendation captioning,which aims at generating explanations for recommendations,so as to enhance the understandability of the recommendation results.Our research covers both directions.First,we study the problems of building an entity recommendation framework and context-aware entity recommendation models.Second,we study the problems of extracting explanations for entity pairs and generating entity highlights for individual entities.To be more specific,the main contributions of this research can be summarized as follows:1.A learning to recommend framework for entity recommendation with serendipity.There are four challenges in the task of building a large-scale entity recommendation system in Web search engines:(1)there are large amounts of queries and entities;(2)the queries are completely domain-free;(3)there exists data sparsity problem in entity click logs;and(4)it is difficult to recommend entities with serendipity.To address these challenges,we propose a learning to recommend framework for entity recommendation.Features and learning objectives that correlate with the important aspects of serendipity are employed in the proposed framework.This framework gives us flexibility to optimize the performance of candidate finding and entity ranking in a two-phase way.It also employs domain-independent features at different levels that are capable of boosting serendipity performance.The proposed method is domain-agnostic and is capable of recommending both personalized and serendipitous entities for any queries.The experiments show that our method can significantly improve the recommendation performance as well as the user engagement upon several strong baseline methods.2.A deep multi-task learning framework for context-aware entity recommendation.Existing studies on entity recommendation typically only consider the current query while ignoring the in-session preceding queries.In addition,there exists data sparsity problem for context-aware recommendations in entity click logs.To address these problems,we propose a multi-task learning framework with deep neural networks to improve entity recommendation by leveraging the abundant search logs of context-aware document ranking task.The proposed method can alleviate the data sparsity problem by using multi-task learning for knowledge transfer.The experiments show that incorporating context information significantly improves entity recommendation,and learning the model in a multi-task learning setting can bring further improvements.3.Learning to explain entity relationships with convolutional neural networks.Providing a plausible explanation for the relationship between two related entities is an important way to enhance the understandability of entity recommendation results.Existing studies highly rely on manually labeled training data and costly handcrafted features,which makes the quality of the produced sentences likely to be lowered.To address these problems,we propose a pairwise ranking model with convolutional neural networks(CNNs)to produce entity relationship explanations.We first construct largescale training data by leveraging the clickthrough data of a Web search engine.Then,we employ a CNN to automatically learn relevant features.The experiments show that our method significantly outperforms several strong baseline methods in terms of sentence quality.4.Generating entity highlights with machine translation based models.If there are no defined relationships between two entities,providing entity highlights for individual entities may increase the understandability of entity recommendations.Entity highlight refers to a short,concise,and characteristic description to an entity,which can help users figure out the connections between a query and the entity.However,this task has not been well studied.To address this task,we propose several machine translation based models to generate entity highlights.Specifically,we propose an entity enhanced sequence-tosequence model to generate entity-specific highlights.The proposed model can capture and retain the salient information w.r.t.a given entity from the source text,and generate entity-specific highlights.The experiments show that it can generate entity highlights of highest-quality.Our research has been successfully applied to Baidu search engine(a widely used commercial Web search engine),which has shown promising improvements in terms of both recommendation effectiveness and user engagement.Our research is a part of the project “A Large-Scale Recommendation Engine for Web Search”,which won the Baidu Highest Prize in 2014.In addition,it is also a part of the project “Knowledge Graph and its Applications”,which won the First Prize for Scientific and Technological Progress by the Chinese Institute of Electronics in 2017.