Lost in Discussion? Tracking Opinion Groups in Complex Political Discussions by the Example of the FOMC Meeting Transcriptions

Proceedings of Recent Advances in Natural Language Processing, pages 747–753, Hissar, Bulgaria, Sep 7–9 2015. Lost in Discussion? – Tracking Opinion Groups in Complex Political Discussions by the Example of the FOMC Meeting Transcriptions C¨acilia Zirn, Robert Meusel and Heiner Stuckenschmidt {caecilia, robert, heiner}@informatik.uni-mannheim.de Data and Web Science Group University of Mannheim Germany Abstract The Federal Open Market Committee (FOMC) is a committee within the central banking system of the US and decides on the target rate. Analyzing the positions of its members is a challenge even for experts with a deep knowledge of the financial do- main. In our work, we aim at automati- cally determining opinion groups in tran- scriptions of the FOMC discussions. We face two main challenges: first, the posi- tions of the members are more complex as in common opinion mining tasks be- cause they have more dimensions than pro or contra. Second, they cannot be learned as there is no labeled data available. We address the challenge using graph cluster- ing methods to group the members, in- cluding the similarity of their speeches as well as agreement and disagreement they show towards each other in discussions. We show that our approach produces sta- ble opinion clusters throughout successive meetings and correlates with positions of speakers on a dove-hawk scale estimated by experts. 1 Introduction In many discussions, participants can easily be divided into two opposing groups, for example people who support democrats versus people who support republicans, or people who are pro or con- tra towards the discussed topic. Having a closer look at the argumentation why people support or defend something however might reveal various stances even within one such group. People might have different opinions and reasons why they support or oppose something. Some discussions have a subject so complex that participants cannot be simply divided in a sup- porting and an opposing group, like the discussion whether abortion should be legal and if yes, up to which status of the pregnancy. In that case the dis- cussants should be grouped by similar positions rather than into pro or contra partitions. In this paper, we are analyzing the discussions of the Federal Open Market Committee (FOMC). The FOMC is a committee within the central banking system of the US and decides on the tar- get rate. The committee meetings are not public, however their transcriptions are released five years later. Understanding the several hundred pages long transcriptions requires a deep knowledge of the financial ecosystem. Apparently even for ex- perts the analysis of those documents is intricate and time-consuming. Our goal is to develop a robust approach to reveal the opinion groups present in discussions where positions are complex to detect and the con- tent is difficult to understand for non-experts. Fur- thermore, we want to assist human readers with a fast automatic system to avoid reading those im- mense amounts of text. There are two major reasons that make it dif- ficult to directly learn a model for the different opinion …

Content truncated to 500 words