Chunk Different Kind of Spoken Discourse: Challenges for Machine Learning

5164 Chunk Different Kind of Spoken Discourse: Challenges for Machine Learning Iris Eshkol-Taravella, Mariame Maarouf, Flora Badin, Marie Skrovec, Isabelle Tellier MoDyCo UMR7114, LLL UMR7270, Lattice UMR8094 200 avenue de la République, 92001 Nanterre, France, 10 Rue de Tours, 45065 Orleans, France 1 rue Maurice Arnoux, 92120 Montrouge, France, ieshkolt@parisnanterre.fr, maarouf.mariame@gmail.com, {marie.skrovec, flora.badin @univ-orleans.fr} Abstract This paper describes the development of a chunker for spoken data by supervised machine learning using the CRFs, based on a small reference corpus composed of two kinds of discourse: prepared monologue vs. spontaneous talk in interaction. The methodology considers the specific character of the spoken data. The machine learning uses the results of several available taggers, without correcting the results manually. Experiments show that the discourse type (monologue vs. free talk), the speech nature (spontaneous vs. prepared) and the corpus size can influence the results of the machine learning process and must be considered while interpreting the results. Keywords: chunking, machine learning, CRF, automatic segmentation of spoken data, oral corpus, kind of spoken discourse 1 Introduction The notion of sentence is generally considered as irrelevant to the analysis and treatment of spoken data (Blanche- Benveniste et al, 1990; Fribourg Group, 2012). Researchers suggested different segmentation units for spoken data as part of projects such as Rhapsody or Orfeo. The project SegCor1 aims to segment the data of talk-in-interaction at different levels. Its first level segmentation concerns minimal syntactic units, which are called chunks. A "chunk" is a non-recursive constituent of linguistic units (Abney 1991). Chunking (or shallow parsing) identifies the surface syntactic structure of a sentence and it can be done automatically. The purpose of a chunking is to identify constituents of the sentence without specifying their internal structure and their syntactic function, which is based on previous morphosyntactic labeling. Chunking is a good way to perform syntactic analysis automatically on spoken data, for which it is not always feasible to provide a full syntactic parsing. Chunkers are well adapted for transcribed oral data in which “sentences” are not always syntactically fulfilled. The chunk is supposed to be a relevant unit for spontaneous speech. Blanche-Benveniste (1997) has shown that it is in the chunks where the reparation markers often occur in spoken data. Some software tools provide this type of analysis but their performance is usually low on oral data. There are several strategies to develop a chunker. There have been attempts to build a chunker that is particularly adapted to French data by using symbolic methods (Blanc et al, 2008, 2010, Antoine et al, 2008). The method consists of iteratively applying finite-state transducers, together with lexical and syntactic resources. The supervised machine learning seems to be particularly effective in this task as shown by (Sha and Pereira, 2003 Tellier et al, 2012, 2014, 1A Franco-German project, funded by the National Research Agency (ANR-15-FRAL-0004) Tsuruoka et al, 2009). The present research continues the work of (Tellier et al, 2014) and uses the method of supervised learning. Oral data are characterized by discursive variety: situational …

Content truncated to 500 words