Although most web search engines do not publish user statistics, software marketing company HubSpot estimates that industry leader Google now fields 5.6 billion web searches per day. The task of delivering web search results is largely handled by ranking systems based on pretrained language models that learn semantic matching between query and document terms — an approach that does not factor in advanced personalization signals such as user clicks.
In the new paper TPRM: A Topic-based Personalized Ranking Model for Web Search, a research team from the Huawei Technologies Artificial Intelligence Application Research Center widens the scope of ranking systems, proposing a topic-based personalized ranking model (TPRM) that integrates language models’ pretrained contextualized term representations with user profiles constructed by a topic model to tailor a more relevant output ranking list.
The team summarizes their main contributions as:
- Integrate topic model-based user profile with pretrained language model to produce a novel personalized ranking system, outperforming state-of-the-art ad-hoc ranking models and personalized ranking models on a real-world AOL dataset.
- Present the interpretability of user topical profiles by providing a means to visualize users’ preferences in selecting documents under the given query.
- Disclose the effects of user interests and the semantic matching learned from queries and documents, revealing their positive contributions to the performance of TPRM.
The proposed TPRM model architecture comprises four modules: (1) User interest modelling, which uses a topic model based on clicked documents in the search history to model user interest; (2) User-Doc interest matching via a kernel-pooling approach; (3) Query-Doc semantic matching, which leverages the BERT large language model to compute the semantic matching of a query and candidate documents; and (4) Personalized ranking, which uses user-doc and query-doc matching vectors to compute a personalized relevance score.
For their empirical study, the team compared TPRM with the BM25 algorithm and state-of-the-art ad-hoc ranking models such as KNRM, Conv-KNRM, CEDR-KNRM, P-Click, SLTB, etc. Experiments were conducted on the real-world AOL search log and used mean average precision (MAP), mean reciprocal rank (MRR), P@1 (precision in the first position), and A.Clk (average click position) as metrics to evaluate the quality of the generated ranking lists.
The results show that most personalized ranking models will outperform ad-hoc ranking models, indicating the effectiveness of user profiles for improving the performance of ranking systems. In the experiments, TPRM significantly outperformed the TPRM-semantic model, verifying the benefits of user profiles constructed via the proposed topic model approach.
The paper TPRM: A Topic-based Personalized Ranking Model for Web Search is on arXiv.
Author: Hecate He | Editor: Michael Sarazen, Chain Zhang
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.