Multilingual Dynamic Topic Model

Proceedings of Recent Advances in Natural Language Processing, pages 1388–1396, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_159 1388 Multilingual Dynamic Topic Model Elaine Zosa and Mark Granroth-Wilding Department of Computer Science University of Helsinki Helsinki, Finland firstname.lastname@helsinki.fi 1 Abstract Dynamic topic models (DTMs) capture the evo- lution of topics and trends in time series data. Current DTMs are applicable only to monolingual datasets. In this paper we present the multilingual dynamic topic model (ML-DTM), a novel topic model that combines DTM with an existing mul- tilingual topic modeling method to capture cross- lingual topics that evolve across time. We present results of this model on a parallel German-English corpus of news articles and a comparable corpus of Finnish and Swedish news articles. We demon- strate the capability of ML-DTM to track signifi- cant events related to a topic and show that it finds distinct topics and performs as well as existing multilingual topic models in aligning cross-lingual topics. 2 Introduction Dynamic topic models (DTMs, Blei and Lafferty, 2006) capture themes or topics discussed in a set of time-stamped documents and how the words re- lated to these topics change in prominence over time. Other topic models have been proposed that aim to model time series data (Wang and McCal- lum, 2006; Wei et al., 2007; Hall et al., 2008). These models can be used to explore historical document collections to study historical trends, language changes (Frermann and Lapata, 2016) and track the emergence and evolution of certain subjects (Hall et al., 2008; Yang et al., 2011). With the internet becoming more multilingual it is increasingly important to build cross-lingual tools to bridge different linguistic groups online. Fortunately, large multilingual datasets such as Wikipedia, the Europarl parallel corpus (Koehn, 2005) and other datasets assembled from crawl- ing the web (Van Gael and Zhu, 2007) are also becoming widely available to researchers. This has led to the development of several multilin- gual topic models to infer topics from multilin- gual datasets. Examples include the polylingual topic model (PLTM, Mimno et al., 2009), mul- tilingual topic model for unaligned text (MuTo, Boyd-Graber and Blei, 2009), and JointLDA (Ja- garlamudi and Daumé, 2010). What is currently lacking are topic models for multilingual time- stamped data that can model historical and lin- guistic changes in a specific context. Digitaliza- tion efforts in libraries and archives, such as the Europeana collections1, have made available on- line historical document collections from different European countries. Collections such as these are valuable resources for comparing historical trends in different countries. However, scholars and other interested parties may not possess the linguistic skills necessary to explore such data and would benefit from tools to automatically discover con- nections across linguistic boundaries. In this paper, we present the multilingual dy- namic topic model (ML-DTM), a novel topic model that captures dynamic topics from broadly topically aligned multilingual datasets. We extend a DTM inference method by Bhadury et al. (2016) to train this model. In the following sections, we give a broad re- …

Content truncated to 500 words