Event Prominence Extraction Combining a Knowledge-Based Syntactic Parser and a BERT Classifier for Dutch

Proceedings of Recent Advances in Natural Language Processing, pages 346–357 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_040 346 Event Prominence Extraction Combining a Knowledge-Based Syntactic Parser and a BERT Classifier for Dutch Thierry Desot, Orph´ee De Clercq, and Veronique Hoste LT3, Language and Translation Technology Team Ghent University, Groot-Brittanni¨elaan 45, 9000 Gent, Belgium {thierry.desot,orphee.declercq,veronique.hoste}@ugent.be Abstract A core task in information extraction is event detection that identifies event triggers in sen- tences that are typically classified into event types. In this study an event is considered as the unit to measure diversity and similarity in news articles in the framework of a news rec- ommendation system. Current typology-based event detection approaches fail to handle the variety of events expressed in real-world sit- uations. To overcome this, we aim to per- form event salience classification and explore whether a transformer model is capable of clas- sifying new information into less and more general prominence classes. After compar- ing a Support Vector Machine (SVM) base- line and our transformer-based classifier per- formances on several event span formats, we conceived multi-word event spans as syntac- tic clauses. Those are fed into our promi- nence classifier which is fine-tuned on pre- trained Dutch BERT word embeddings. On top of that we outperform a pipeline of a Conditional Random Field (CRF) approach to event-trigger word detection and the BERT- based classifier. To the best of our knowledge we present the first event extraction approach that combines an expert-based syntactic parser with a transformer-based classifier for Dutch. 1 Introduction Recently, news publishers have shifted from news- papers to digital means which provide news readers easy access to a wide range of news information. However, the challenge is to find the right con- tent that also corresponds to the user’s personal interests. Therefore, many of today’s major media and news websites offer automated news recom- mendation and personalization (Das et al., 2007; Odi´c et al., 2013; Moreira et al., 2019; Feng et al., 2020). News personalization paradigms define good news recommendations in terms of similarity to the user’s previous reading behaviour. Hence, news articles are recommended based on proximity to other articles the user has read (Liu et al., 2010; Adnan et al., 2014). However, this contrasts with the normative concept of journalism that stimulates diversity of topics and events in unfiltered news streams (Pariser, 2011; Joris et al., 2019). In this study we consider the news event as a means to model both diversity and similarity in news articles in the context of a news recommendation system. We present an event extraction approach that will be integrated in a news recommender for Dutch1. As current typology-based event detection fails to handle the variety of events in real-world situations we applied event prominence classification. This allows us to detect unrestricted news events and to overcome the sparsity of a small training data set. Our event extraction approach combines an expert- based syntactic parser with a transformer-based classifier: • Input sentences are first pre-processed using a rule-based syntactic parser in order to …

Content truncated to 500 words