HEADS: Headline Generation as Sequence Prediction Using an Abstract Feature-Rich Space

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 133–142, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics HEADS: Headline Generation as Sequence Prediction Using an Abstract Feature-Rich Space Carlos A. Colmenares∗ Google Inc. Brandschenkestrasse 110 8002 Zurich, Switzerland crcarlos@google.com Marina Litvak Shamoon College of Engineering Beer Sheva, Israel marinal@sce.ac.il Amin Mantrach Fabrizio Silvestri Yahoo Labs. Avinguda Diagonal 177 08018 Barcelona, Spain {amantrach,silvestr}@yahoo-inc.com Abstract Automatic headline generation is a sub-task of document summarization with many re- ported applications. In this study we present a sequence-prediction technique for learning how editors title their news stories. The intro- duced technique models the problem as a dis- crete optimization task in a feature-rich space. In this space the global optimum can be found in polynomial time by means of dynamic pro- gramming. We train and test our model on an extensive corpus of financial news, and com- pare it against a number of baselines by us- ing standard metrics from the document sum- marization domain, as well as some new ones proposed in this work. We also assess the readability and informativeness of the gener- ated titles through human evaluation. The ob- tained results are very appealing and substan- tiate the soundness of the approach. 1 Introduction Document summarization, also known as text sum- marization, is the process of automatically abridging text documents. Although traditionally the final ob- jective of text summarization is to produce a para- graph or abstract that summarizes a rather large col- lection of texts (Mani and Maybury, 1999; Das and Martins, 2007; Nenkova and McKeown, 2012), the task of producing a very short summary comprised of 10–15 words has also been broadly studied. There have been many reported practical applications for this endeavor, most notably, efficient web browsing ∗Work done during an internship at Yahoo Labs. on hand-held devices (Buyukkokten et al., 2001), generation of TV captions (Linke-Ellis, 1999), dig- itization of newspaper articles that have uninforma- tive headlines (De Kok, 2008), and headline gener- ation in one language based on news stories written in another (Banko et al., 2000; Zajic et al., 2002). In general terms, a headline of a news article can be defined as a short statement that gives a reader a general idea about the main contents of the story it entitles (Borko and Bernier, 1987; Gattani, 2007). The objective of our study is to develop a novel tech- nique for generating informative headlines for news articles, albeit to conduct experiments we focused on finance articles written in English. In this work we make a number of contributions concerning sta- tistical models for headline generation, training of the models, and their evaluation, specifically: • We propose a model that learns how an edi- tor generates headlines for news articles, where a headline is regarded as a compression of its article’s text. Our model significantly dif- fers from others in the way it represents pos- sible headlines in a feature-rich space. The model …

Content truncated to 500 words