Efficient Neural Query Auto Completion

Abstract

Query Auto Completion (QAC), as the starting point of information retrieval tasks, is critical to user experience. Generally it has two steps: generating completed query candidates according to query prefixes, and ranking them based on extracted features. Three major challenges are observed for a query auto completion system: (1) QAC has a strict online latency requirement. For each keystroke, results must be returned within tens of milliseconds, which poses a significant challenge in designing sophisticated language models for it. (2) For unseen queries, generated candidates are of poor quality as contextual information is not fully utilized. (3) Traditional QAC systems heavily rely on handcrafted features such as the query candidate frequency in search logs, lacking sufficient semantic understanding of the candidate.In this paper, we propose an efficient neural QAC system with effective context modeling to overcome these challenges. On the candidate generation side, this system uses as much information as possible in unseen prefixes to generate relevant candidates, increasing the recall by a large margin. On the candidate ranking side, an unnormalized language model is proposed, which effectively captures deep semantics of queries. This approach presents better ranking performance over state-of-the-art neural ranking methods and reduces ∼95% latency compared to neural language modeling methods. The empirical results on public datasets show that our model achieves a good balance between accuracy and efficiency. This system is served in LinkedIn job search with significant product impact observed.

Session-Aware Query Auto-completion using Extreme Multi-Label Ranking

“…Of the models that have latencies suitable for realtime systems with a latency budget of the order of few milliseconds, PrefXMRtree outperforms other baselines such as MFQ, and MFQ + seq2seq-GRU. Consistent with previous work [39], MFQ outperforms MFQ + seq2seq-GRU which re-ranks wrt likelihood conditioned on previous query using a seq2seq-GRU model (except for shorter prefixes as in Fig. 3).…”

“…we lowercase all queries, replace periods with spaces, and remove non-alphanumeric characters. Further, we split a sequence of queries into sessions using a 30-min idle time between consecutive queries to define the end of a session [36,39]. Given a session consisting of a sequence of queries < q 1 , .…”

“…Traditional query auto-completion systems operate in the retrieveand-rank framework where a fixed set of queries are retrieved given the input prefix followed by ranking based on various sources of information. The simplest yet effective approach is to suggest the top-k most frequent queries matching the input prefix [39]. Other approaches further improve over this by ranking suggestions based on previous queries, user profile, time-sensitivity, or coherence with the typed prefix [5,34,36,39].…”

“…The simplest yet effective approach is to suggest the top-k most frequent queries matching the input prefix [39]. Other approaches further improve over this by ranking suggestions based on previous queries, user profile, time-sensitivity, or coherence with the typed prefix [5,34,36,39]. Recent work also leverages deep learning models to generate additional features such as query likelihood using language models [30], personalized language models [12,14], or previous queries [36] to rank queries using a learning to rank framework [41].…”

“…These retrieved queries are then ranked based on features like frequency, similarity to the previous query, user profile, etc. [5,34,36,39]. Finally, the top-k ranked queries are shown as suggestions to the user.…”