Detection of Opinions and Facts. A Cognitive Approach

International Conference RANLP 2009 - Borovets, Bulgaria, pages 150–154 Detection of opinions and facts. A cognitive approach Yann Vigile Hoareau† Human and Artificial Cognition Lab. University of Paris 8 – 2 rue de la Liberté. 93526 St Denis Cedex 02 -FRANCE hoareau@lutin-userlab.fr Adil El-Ghali Human and Artificial Cognition Lab. University of Paris 8 – 2 rue de la Liberté. 93526 St Denis Cedex 02 -FRANCE elghali@lutin-userlab.fr Charles Tijus Human and Artificial Cognition Lab. University of Paris 8 – 2 rue de la Liberté. 93526 St Denis Cedex 02 -FRANCE tijus@lutin-userlab.fr Abstract A model of episodic memory is derived to propose algorithms of text categorization with semantic space models. Performances of two algorithms named Target vector and Sub-target vector are contrasted using textual material of the text-mining context ‘DEFT09’. The experience reported here have been realized on the english corpus which is composed of articles of the economic newspaper “The Financial Times”. The aim of the task was to categorize texts in function of the factuality or subjectivity they expressed. Results confirm (i) that the episodic memory metaphor provides a convenient framework to propose efficient algorithm for text categorization, and (ii) that Sub-target vector algorithm outperforms the Target vector algorithm. Keywords Random Indexing, episodic memory, text-mining, categorization. 1. Introduction Since its early introduction, the model that is now named Latent Semantic Analysis [14] has been proposed as a method of matrix reduction and vectorial representation of information for indexing textual documents. The model was known as Latent Semantic Indexing [3] at that time. Originally only concerned by indexing tasks, LSA has been extended to the area of human memory simulation. Researchers in cognitive psychology got interested in it and then proposed it as a plausible model of human behavior in different tasks such as synonymy test [14] and problem solving [17]. The most famous application in cognitive psychology is the coupled CI-LSA model of text comprehension [12], which combines the previous “Construction-Integration” model of reading [11] with LSA as model of semantic memory. Whereas research involving LSA has been split in two main fields with the text-mining on the one hand and cognitive psychology on the other hand, our paper deals with both of those fields. Discussions of MINERVA 2 model of human episodic memory [6][7] allow proposing an operative algorithm for texts categorization. LSA has been known to perform in synonymy test and other equivalent thematic classification tasks [14]. The model has been recently successfully applied on opinion judgment task [1]. There are very important differences between thematic classification, and opinion judgment classification. Firstly, thematic classification is directly connected to the distributional hypothesis, which states that “words that appear in similar contexts have similar meanings”. Here is the reason why LSA is able to find words that share the same thematic, ie “appear in equivalent contexts”. Secondly, in opinion judgment classification, different thematics could possibly belong to the same category of opinion. For example, I have a good opinion of different movies, which do not deal with the same topic. …

Content truncated to 500 words