The Design And Implementation Of Search Recommendation System Based On Machine Translation Model

A Search engine is an important means for users to find useful information.When users use search engines,they often fail to accurately construct queries to express their real intentions,which results in search engines failing to retrieve accurate results.Relevant search recommendation refers to recommend relevant queries to users based on user's query.Relevant search recommendation on the mobile side is an important way to understand the behavior of mobile phone users,help users to build queries and improve their efficiency in using mobile phones for information retrieval.Xiaomi mobile phone's current relevant search recommendation system indexes the search logs,and obtains relevant queries by searching the index.Such practices have poor timeliness,low coverage,and unintended generalization effects.The historical query has certain relevance to the corresponding click document name,we can use neural network machine translation model extracts the correlation law between the historical query and the click data.Therefore,this thesis proposes a neural network machine translation model to design and implement a relevant search recommendation system,using the model to translate the new queries into relevant query,and recommend them to the user through the data index and retrieval system.We train the neural network machine translation model and applies the model to the relevant search recommendation system.The system includes an offline data processing module and an online search terminal module.The offline data processing module includes using Spark Streaming to extract daily top query,model translate top query to relevant query,and build relevant query index based on Lucene.The online retrieval module retrieves relevant query from the index,establishes a Response through Backend,and then sorts and tunes the Response through the Rank and Tunner sub-modules,and finally presents the results to the user.Such a design and implementation can balance the conflict between the cost of log data mining and the high update frequency of index,improve the correlation between recommended query and original query,thereby improving coverage;recommend query based on top query,improving the timeliness;using the intent generalization ability of the model,can generate certain intent generalization effects for new query.