Discriminating Rhetorical Analogies in Social Media

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 560–568, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Discriminating Rhetorical Analogies in Social Media Christoph Lofi Christian Nieke Nigel Collier National Inst. of Informatics TU Braunschweig National Inst. of Informatics 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo 101-8430, Japan Mühlenpfordtstr. 23 38106 Braunschweig Germany 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo 101-8430, Japan lofi@nii.ac.jp nieke@ifis.cs.tu-bs.de collier@nii.ac.jp Abstract Analogies are considered to be one of the core concepts of human cognition and communica- tion, and are very efficient at encoding com- plex information in a natural fashion. How- ever, computational approaches towards large- scale analysis of the semantics of analogies are hampered by the lack of suitable corpora with real-life example of analogies. In this paper we therefore propose a workflow for discriminat- ing and extracting natural-language analogy statements from the Web, focusing on analo- gies between locations mined from travel re- ports, blogs, and the Social Web. For realizing this goal, we employ feature-rich supervised learning models which we extensively evalu- ate. We also showcase a crowd-supported workflow for building a suitable Gold dataset used for this purpose. The resulting system is able to successfully learn to identify analogies to a high degree of accuracy (F-Score 0.9) by using a high-dimensional subsequence feature space. 1 Introduction Analogies are one of the core concepts of human cognition (Hofstadter, 2001), and it has been sug- gested that analogical inference is the “thing that makes us smart” (Gentner, 2003). An analogy can be seen as a pattern of speech leading to a cogni- tive process that transfers some high-level mean- ing from one particular subject (often called the analogue or the source) to another subject, usually called the target. When using analogies, one em- phasizes that the “essence” of source and target is similar, i.e. their most discriminating and proto- typical processes and properties are perceived in a similar way. The nature of analogies has been discussed and studied since the ancient Greeks, however compu- tational approaches are still rather limited and in their infancy. One reason for this is that text cor- pora containing analogies are crucial to study the syntactic and semantic patterns of analogies in or- der to make progress on automated understanding techniques. For example, to learn about their dis- tribution and the attribute-value pairs that are compared. However, to the best of our knowledge, no such corpus is freely available. We will there- fore in this paper present a method for creating such a corpus in an efficient fashion, and make our corpus available for further research efforts. As an example, consider this brief statement: “West Shinjuku (a Tokyo district) is like Lower Manhat- tan” It allows readers who know New York, but not Tokyo, to infer some of the more significant properties of the unknown district (e.g., it is an im- portant business district, hosts the headquarters of many companies, features many skyscrapers, etc.). However, automatically understanding anal- ogies is surprisingly hard due to …

Content truncated to 500 words