DIHANA corpus

Abstract

DIHANA is composed of 900 human-computer dialogues in Spanish. The acquisition of the DIHANA corpus was carried out by means of an initial prototype using the Wizard of Oz technique. The operator simulates speech recognition and understanding errors and the answers being synthesized according to a predefined set of templates. This acquisition was only restricted at the semantical level (i.e., the acquired dialogues are related to a specific task domain) and was not restricted at the lexical and syntactical level (spontaneous speech). In the acquisition process, this semantic control was provided by the definition of scenarios that the user must accomplish and by the wizard strategy, which defines the behavior of the acquisition system. The DIHANA task consists of the retrieval of information about Spanish nationwide trains by telephone. Several types of scenarios were defined in order to control the interaction of the user with the system. A scenario is defined by: an objective, the information needed by the user; a situation, the specific circumstances related to the trip request; and the specific requirements of the trip, type of trip, departure city, destination city, and one or more restrictions. The DIHANA corpus contains 5.5 hours of spontaneous speech corresponding to 6278 sentences. In total 225 speakers (153 males and 72 females) recorded 900 dialogues, resulting in 6,278 user turns. Along with the dialogues (speech signals), their full transcript is also provided and a lexicon phonetically containing all the words pronounced in the database. In addition a semantic tagging of the corpus and a labeling of the same corpus in terms of dialog acts is also provided. A more detailed description of DIHANA can be found in the "doc" subfolder and in the following papers: - N. Alcacer, J.M. Benedí, F. Blat, R. Granell, D. Martínez-Hinarejos and F. Torres: "Acquisition and Labelling of a Spontaneous Speech Dialogue Corpus". In proceedings of SPECOM, pages 583-586. Patras (Greece), October 2005. - J.M. Benedí E.Lleida, A. Varona, M.J.Castro, I.Galiano, R.Justo, I. López, and A. Miguel: "Design and acquisition of a telephone spontaneous speech dialogue corpus in spanish: DIHANA". In proceedings of LREC, pages 1636-1639, Genova, Italy, May 2006. Next we describe the contents of each subfolder: README: This file. data: The database: 75 speakers of 3 sites (Basque Country,<br> Aragon and Valencian Country) for 4 scenarios making a<br> total of 225 speakers (153 males and 72 females) with<br> 900 dialogues. For each dialog is provided: - the speech signal of each user turn (*.ul)<br> - the intermediate (*.dis) and final (*.xml) transcriptions<br> - the dialogue Act annotation + Dialogue Act annotation on transcription (*.dia)<br> + Dialogue Act annotation on categorised transcription,<br> without words for each category (*.cad)<br> + Dialogue Act annotation on categorised transcription,<br> with words for each category (*.cwd) semdata: Semantic tagging of the full corpora, in the "data" subfolder,<br> and documentation describing the process of semátinco labeling,<br> in the "doc" folder. doc: Various documents (PDF) related to the design and acquisition<br> processes, the annotation format and event statistics. guides: 5 lists of 45 speakers, which account for 5 leaving-one-out<br> partitions. One of the lists can be alternatively chosen as<br> the test set, the other four joined to form the training set.<br> Under the folder corresponding to each speaker, the speech<br> signals and transcriptions corresponding to four dialogues<br> can be found, so each partition consists of 720 training<br> dialogues an 180 test dialogues. software: Various self-commented programs and utilities.