Probing Multilingual Cognate Prediction Models

Findings of the Association for Computational Linguistics: ACL 2022, pages 343 - 352 May 22-27, 2022 c⃝2022 Association for Computational Linguistics MINER: Multi-Interest Matching Network for News Recommendation Jian Li1, Jieming Zhu1, Qiwei Bi2, Guohao Cai1, Lifeng Shang1, Zhenhua Dong1, Xin Jiang1, Qun Liu1 1Huawei Noah’s Ark Lab 2School of Statistics, Renmin University of China {lijian703, jamie.zhu, caiguohao1, shang.lifeng, dongzhenhua}@huawei.com {jiang.xin, qun.liu}@huawei.com Abstract Personalized news recommendation is an es- sential technique to help users find interested news. Accurately matching user’s interests and candidate news is the key to news recommen- dation. Most existing methods learn a single user embedding from user’s historical behav- iors to represent the reading interest. However, user interest is usually diverse and may not be adequately modeled by a single user embed- ding. In this paper, we propose a poly atten- tion scheme to learn multiple interest vectors for each user, which encodes the different as- pects of user interest. We further propose a dis- agreement regularization to make the learned interests vectors more diverse. Moreover, we design a category-aware attention weighting strategy that incorporates the news category in- formation as explicit interest signals into the at- tention mechanism. Extensive experiments on the MIND news recommendation benchmark demonstrate that our approach significantly out- performs existing state-of-the-art methods. 1 Introduction Online news platforms such as Google News1 and Microsoft News2 have become a prevalent way for users to access news information (Das et al., 2007). The large amount of news generated every day make it hard for users to find their interested news. To alleviate information overload and im- prove reading experience, personalized news rec- ommendation has become an essential part of these platforms (Liu et al., 2010; Phelan et al., 2011). Accurate matching between users’ interests and candidate news is the key to personalized news rec- ommendation. Existing methods usually learn a single user interest vector by aggregating the pre- viously browsed news via sequential or attentive models and then match it with candidate news vec- tors. For example, Okura et al. (2017) employ a 1https://news.google.com/ 2https://microsoftnews.msn.com/ 8.13 Category Title Finance 5 excellent dividend stocks to buy for the holiday season. Sports Should NFL be able to fine players for criticizing officiating? Sports 5 takeaways from the 49ers' dominant win over the Panthers. Movies Francis Ford Coppola says Marvel movies are ‘despicable‘. Sports Magic vs. Cavs Preview: Magic basketball is finally back. Fitness This guy altered his diet and training to drop 65 pounds and pack on muscle. News Click Sequence Figure 1: The news click history of one user, who has various interests including finance, sports and movies. GRU network to model user interest from the se- quence of clicked news with the last hidden state of GRU being the user interest representation. An et al. (2019) also use a GRU network to aggregate clicked news sequence as an interest vector and combine it with the user ID embedding. Wu et al. (2019a) and Wu et al. (2019c) apply attentive pool- ing on the sequence …

Content truncated to 500 words