Tracking Information Flow in Financial Text

Tracking Information Flow in Financial Text Will Radford †‡ Ben Hachey ‡⋄ James R. Curran †‡ Maria Milosavljevic ‡ School of Information Technologies † Capital Markets CRC ‡ Centre for Language Technology ⋄ University of Sydney 55 Harrington Street Macquarie University NSW 2006, Australia NSW 2000, Australia NSW 2109, Australia {wradford,james}@it.usyd.edu.au {bhachey,maria}@cmcrc.com Abstract Information is fundamental to Finance, and un- derstanding how it flows from official sources to news agencies is a central problem. Readers need to digest information rapidly from high vol- ume news feeds, which often contain duplicate and irrelevant stories, to gain a competitive ad- vantage. We propose a text categorisation task over pairs of official announcements and news stories to identify whether the story repeats an- nouncement information and/or adds value. Us- ing features based on the intersection of the texts and relative timing, our system identifies infor- mation flow at 89.5% F-score and three types of journalistic contribution at 73.4% to 85.7% F- score. Evaluation against majority annotator de- cision performs 13% better than a bag-of-words baseline. 1 Introduction Financial news is an important resource for capital market participants and plays a central role in how they interact with the market. Companies must contin- uously disclose any information “a reasonable person would expect to have a material effect on the price or value of the entity’s securities” (ASX, 2008). News agencies publish Finance stories that report on a broad range of events. Some stories report facts directly from announcements and may add value by present- ing background knowledge, expert analysis or edito- rial commentary. The financial environment rewards participants that are alert and responsive to incoming information (Zaheer and Zaheer, 1997) and automated analysis of information flow is highly advantageous. We define information flow between a pair of docu- ments as when one document repeats information from the other. Textual similarity is central to this and has been addressed in a variety of research areas. Plagia- rism detection concentrates on verbatim duplication of sections of text (Brin et al., 1995; Wise, 1996), while Information Retrieval techniques assess similarity at the broader topic level (Manning et al., 2008). Text Reuse examines a finer-grained notion of similarity (Metzler et al., 2005) between the verbatim copying and topic similarity. Topic detection and tracking (Al- lan et al., 1998a) focuses on tracking emerging events at a topic level over a news feed. We examine two sources: the Australian Securi- ties Exchange (ASX)1 official announcements and the Reuters NewsScope Archive (RNA)2, both of which release time-stamped documents tagged with one or more company stock ticker codes. In this research, ASX-RNA document pairs that share the same ticker and were published within a time window are ex- tracted. The ASX-RNA pairs are annotated to indicate information flow (LINK) and, if so, whether the story is the first to report an announcement (FIRST), back- ground (BACK) or analysis (ANLY) content. We for- mulate four tasks classifying whether each label ap- plies to ASX-RNA pairs. We design textual and temporal features to model …

Content truncated to 500 words