Dialog Intent Structure: A Hierarchical Schema of Linked Dialog Acts

Dialog Intent Structure: A Hierarchical Schema of Linked Dialog Acts Silvia Pareti, Tatiana Lando Google Research Europe {spareti,tlando}@google.com Abstract In this paper, we present a new hierarchical and extensible schema for dialog representation. The schema captures the pragmatic intents of the conversation independently from any semantic representation. This schema was developed to support computational applications, be applicable to different types of dialogs and domains and enable large-scale non-expert annotation. The schema models dialog as a structure of linked units of intent, dialog acts, that are annotated on minimal spans of text, functional segments. Furthermore, we categorise dialog acts based on whether they express a primary or secondary intent and whether the intent is explicit or implicit. We successfully tested the schema on an heterogeneous corpus of human-human dialogs comprising both spoken and chat interactions. Keywords: Dialog Acts, Speech Acts, Intents, Dialog 1. Introduction With the increasing popularity of conversational systems and chatbots, we are faced with the challenge of creating suitable dialog representations and annotated data to sup- port the development of dialog modeling systems. Relying on form-filling strategies, with predefined dialog states se- quences, falls short of providing a flexible and adaptable system that supports natural conversation, where the inter- action is not constrained by a fixed plot with one precise goal. On the other hand, treating utterances or messages in- dependently from each other fails to recognise the intercon- nected dialog structure and phenomena that are crucial to correctly understand and generate dialog interactions. Part of a larger dynamic and interactive language exchange, dia- log messages have both semantic content and communica- tive functions. The latter provide the conversational context needed to correctly interpret the semantic content and de- termine what the current dialog state is. The challenge of representing the communicative intent of dialog segments with dialog acts (DAs) has been addressed multiple times over the past decades. While earlier efforts, e.g. Map Task (Carletta et al., 1997), have focused on repre- senting specific tasks and domains, the attention has shifted to more general purpose schemas such as the Dialogue Act Markup using Several Layers (DAMSL) (Allen, 1997) and the ISO 24617-2(Bunt et al., 2010; Bunt et al., 2012), a stan- dard for dialog act annotation. The latter offers a powerful representation comprising 9 core dimensions and around 60 communicative functions. Communicative functions are linked to functional segments (FSs), defined by (Bunt et al., 2012) as "the unit of dialogue act annotation". These schemas were not specifically developed to represent human-computer conversation. By developing a schema for the specific application, we can ensure that we identify use- ful categories that can help us understand the human dialog input as well as generate a suitable machine reaction while also reducing the tagset complexity. The latter is crucial to conduct large-scale non-expert annotation. In particular, such representation should enable non-constrained interac- tions were the dialog structure is not predefined around a specific scenario with a single task to solve, but transitions between dialog states are learned from data. For …

Content truncated to 500 words