Direct Reported Speech in Multilingual Texts: Automatic Annotation and Semantic Categorization

We propose an application for the automatic identification and categorization of quotations. The categorization is based on a semantic map of enunciative modalities. The texts are treated in three languages: Arabic, Korean and French. 1. General presentation and related works Automatic identification of quotations using natural language processing (NLP) is now significantly growing in recent studies (Mourad 2001), (Krestel, Bergler, and Witte 2008), InQuote, (Pouliquen, Steinberger, and Best 2008), (Audebert, Gaubert, and Jaccarini 2009) 3 and (De la Clergerie et al. 2009). We propose in this study an application for the automatic identification and categorization of quotations. This work can be distinguished from the previous ones in many aspects. First of all, our concerns are not to detect the source (holder) of the quotation, neither its anaphoric analysis, but we aim to identify all forms of quotation in texts by taking into consideration of its potential constructions. In addition, by using the theory of enunciation, we aim to automatically categorize the quotations in terms of various semantic criteria (commitment, opinion, judgment...), in a multilingual context (Arabic, French and Korean). Finally, the tool we use for automatic annotation, EXCOM , is a rule-based system that does not deal with any morpho-syntactic analysis or named entities recognition (Alrahabi and Desclés 2009b). EXCOM, implementing the method of Contextual Exploration (Desclés 2006), automatically performs the annotations using the surface forms of certain linguistic markers. In the following sections, we begin by presenting the linguistic analysis of quotations, and then we explain how 1 http://labs.google.com/inquotes/ 2 http://press.jrc.it/NewsExplorer/home/fr/latest.html 3 http://www.ifao.egnet.net/kawakib 4 http://www.excom.fr/ the linguistic markers can be organized in a semantic map. We finish the article by showing the result of the evaluation, and the perspectives. 2. Quotation analysis First, let’s introduce this important distinction between “utterer” (énonciateur) and “speaker” (locuteur). The utterer is the entity that reports the speech, whereas the speaker is the source (holder) of the speech. We consider, on the formal level, that a quotation is any kind of speech delimited by meta-characters (the typographical signs of quotation) and introduced by, at least, one linguistic marker referring to an act of speaking, whether the speaker is explicitly defined or not. We take into consideration any form of direct reported speech, as long as these rules are observed, i.e. the canonical forms and hybrids or mixed forms (such as the direct style introduced by “that”, see (Tuomarla 2000)). In general, we consider that an utterer can report a speaker's discourse in, at least, three ways : • By attributing to a speaker an implicit act of locution (Pour X [As for X] / ربخلا اذه مكيلا [Here is this news...] / [According to X]). This reflects the distance that the utterer takes in relation to the reported content. • By attributing to a speaker a speech as an act of “hearing” (Je me suis laissé entendre [It was intimated to me...] / يلي ام انغلب [This news has reached us] / [heard from X]). This often indicates the spread of information (or rumors). • By attributing to a speaker an explicit act of locution (X a décidé [X decided] / نلاف نلعأ [X declared] / 5 In Korean (Pak et al. 2009), a set of linguistic markers following quotation marks often indicate a real quotation, such as ( / lako, / lakoto / / ko, / koto, / ilako, etc.). 6 Examples in this paper are not identical from one language to another, but they belong to the same semantic categories. 162 Proceedings of the Twenty-Third International Florida Artificial Intelligence Research Society Conference (FLAIRS 2010)