News Citation Recommendation with Implicit and Explicit Semantics

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 388–398, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics News Citation Recommendation with Implicit and Explicit Semantics Hao Peng∗, 1 Jing Liu,2 Chin-Yew Lin2 1School of EECS, Peking University, Beijing, 100871, China 2Microsoft Research, Beijing, 100080, China penghao.pku@gmail.com {liudani, cyl}@microsoft.com Abstract In this work, we focus on the problem of news citation recommendation. The task aims to recommend news citations for both authors and readers to create and search news references. Due to the sparsity issue of news citations and the engineering difficulty in obtaining infor- mation on authors, we focus on content similarity-based methods instead of col- laborative filtering-based approaches. In this paper, we explore word embedding (i.e., implicit semantics) and grounded en- tities (i.e., explicit semantics) to address the variety and ambiguity issues of lan- guage. We formulate the problem as a re- ranking task and integrate different simi- larity measures under the learning to rank framework. We evaluate our approach on a real-world dataset. The experimental re- sults show the efficacy of our method. 1 Introduction When an author writes an online news article, s/he often cites previously published news reports to elaborate a mentioned event or support his/her point of view. For the convenience of the read- ers, the editor usually associates the words with hyperlinks. Through the links the readers can di- rectly access the referenced articles to know more details about the events. If there is no reference for a mentioned event, the readers may search the related news reports for further reading. Hence, it is valuable to have automatic news citation rec- ommendations for authors and readers to create or search news references. In this paper, we focus on the problem of news citation recommendation. As shown in Table 1, ∗ Work done during internship at Microsoft Research. given a snippet of citing context (left), the task aims to retrieve a list of news articles (right) as references. This task differs from traditional rec- ommendation tasks, e.g., citation recommendation for scientific papers, in that: (a) based on the statis- tics from our dataset, the number of references per news article is 4.56 on average, much less than the number of citations per academic paper (typ- ically dozens); (b) the author-topic information is usually unavailable, since it is technically difficult to obtain author information from news articles. These differences make the collaborative filtering- based methods, which have been widely applied to paper citation recommendation, less available in our scenario. Therefore, in this paper we focus on content similarity-based methods to deal with the task of news citation recommendation. Previous studies use string-based overlap (Xu et al., 2014), machine translation measures (Madnani et al., 2012), and dependency syntax (Wan et al., 2006; Wang et al., 2015) to model text similar- ity. More recent work focuses on neural network methods (Yin and Sch¨utze, 2015; He et al., 2015; Hu et al., 2014; dos Santos et al., 2015; Lei et al., 2016). …

Content truncated to 500 words