ES-Port: a Spontaneous Spoken Human-Human Technical Support Corpus for Dialogue Research in Spanish

ES-Port: a Spontaneous Spoken Human-Human Technical Support Corpus for Dialogue Research in Spanish Laura Garc´ıa-Sardi˜na, Manex Serras, Arantza del Pozo Speech and Natural Language Technologies, Vicomtech Mikeletegi Pasealekua 57, 20009 Donostia-San Sebasti´an {lgarcias, mserras, adelpozo}@vicomtech.org Abstract In this paper the ES-Port corpus is presented. ES-Port is a spontaneous spoken human-human dialogue corpus in Spanish that consists of 1170 dialogues from calls to the technical support department of a telecommunications provider. This paper describes its compilation process, from the transcription of the raw audio to the anonymisation of the sensitive data contained in the transcriptions. Because the anonymisation process was carried out through substitution by entities of the same type, coherence and readability are kept within the anonymised dialogues. In the resulting corpus, the replacements of the anonymised entities are labelled with their corresponding categories. In addition, the corpus is annotated with acoustic-related extralinguistic events such as background noise or laughter and linguistic phenomena such as false starts, use of filler words or code switching. The ES-Port corpus is now publicly available through the META-SHARE repository, with the main objective of promoting further research into more open domain data-driven dialogue systems in Spanish. Keywords: spontaneous dialogue corpus, human-human dialogue, technical support, transcription, anonymisation, named enti- ties 1. Introduction Dialogue systems, often referred to as conversational agents or chatbots, are becoming increasingly popular as they allow users to directly interact with a wide range of in- formation systems in a natural way. Customer support is an application scenario with a strong interest in dialogue sys- tem development, driven by the promise of intelligent digi- tal assistants available 24x7 to resolve customer requests in a fast, cost-effective and consistent manner (Guzm´an and Pathania, 2016). Data-driven approaches to dialogue system development have shown to be more robust than rule-based techniques to variability in user behaviour, the performance of speech and language processing subcomponents and the dynamics of the task domain (Meena, 2015). Despite their promis- ing results in recent years (Young et al., 2013; Wen et al., 2016; Li et al., 2017), most practical dialogue systems are still built by human experts through significant manual en- gineering. In most currently deployed systems, rule-based dialogue managers (DM) are combined with statistical nat- ural language understanding (NLU) models capable of clas- sifying intents and their related entities (Williams et al., 2015). In the customer support domain, this limits their application to frequent use cases in specific areas where solutions are well known, predictable and where scripted answers can be developed (Guzm´an and Pathania, 2016). Lack of annotated corpora is the main problem for the de- velopment of data-driven systems (Serban et al., 2015). To overcome this issue, it is usual to develop rule-based baselines or to employ the Wizard of Oz (WOZ) technique (Benedı et al., 2006; Rieser and Lemon, 2008) in which a human mimics the intended dialogue system, in order to gather interactions with real users. Although this kind of human-machine data is constrained by the employed baseline systems or the scenarios defined in …

Content truncated to 500 words