Looking for Trouble

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 185–192 Manchester, August 2008 Looking for Trouble Stijn De Saeger Kentaro Torisawa Language Infrastructure Group National Institute of Information and Communications Technology {stijn,torisawa}@nict.go.jp Jun’ichi Kazama School of Information Science Japan Advanced Institute of Science and Technology kazama@jaist.ac.jp Abstract This paper presents a method for mining potential troubles or obstacles related to the use of a given object. Some exam- ple instances of this relation are ⟨medicine, side effect⟩ and ⟨amusement park, height restriction⟩. Our acquisition method con- sists of three steps. First, we use an un- supervised method to collect training sam- ples from Web documents. Second, a set of expressions generally referring to trou- bles is acquired by a supervised learning method. Finally, the acquired troubles are associated with objects so that each of the resulting pairs consists of an ob- ject and a trouble or obstacle in using that object. To show the effectiveness of our method we conducted experiments using a large collection of Japanese Web doc- uments for acquisition. Experimental re- sults show an 85.5% precision for the top 10,000 acquired troubles, and a 74% pre- cision for the top 10% of over 60,000 ac- quired object-trouble pairs. 1 Introduction The Stanford Encyclopedia of Philosophy defines an artifact as “. . . an object that has been inten- tionally made or produced for a certain purpose”. Because of this purpose-orientedness, most human actions relating to an object or artifact fall into two broad categories — actions relating to its in- tended use (e.g. reading a book), and the prepa- rations necessary therefore (like buying the book). Information concerning potential obstacles, harm- ful effects or troubles that interfere with this in- tended use is therefore highly relevant to the user. c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. While some such troubles are self-evident, others represent a genuine obstacle whose existence was thusfar unknown to the user. For example, in early 2008 a food poisoning case caused a big media stir in Japan when dozens of people fell ill after eating Chinese-imported frozen food products containing residual traces of toxic pesticides. While suppos- edly the presence of toxic chemicals in imported frozen foods had already been established on sev- eral occasions before, until the recent incidents public awareness of these facts remained low. In retrospect, a publicly available system suggesting “residual agrichemicals” as a potential danger with the consumption of “frozen foods” based on in- formation mined from a large collection of Web documents might have led to earlier detection of this crisis. From the viewpoint of manufacturers as well, regularly monitoring the Internet for prod- uct names and associated troubles may allow them to find out about perceived flaws in their products sooner and avoid large scale recalls and damage to their brand. For a less dramatic example, searching for “Tokyo Disneyland” on the Internet typically yields many commercial sites offering travel deals, but little …

Content truncated to 500 words