Automatic Dialogue Flow Extraction and Guidance

EasyChair Preprint 8132, version 1

Abstract

Communication between clients and call-centers has undergone numerous changes over the years, in response to different challenges, such as long response times and difficulties in assuring solutions for the problems posed by the clients. Nowadays, human chat service agents are frequently replaced by conversation software agents, designed to communicate with humans by means of natural language, often based on Artificial Intelligence, namely Natural Language Processing (NLP) and Machine Learning (ML). This work will begin by identifying, collecting and annotating dialogue sets, written in Portuguese, to be used in the experimentation and made available to the community. It consists of researching and implementing a solution with the objective of aiding communication between participants, suggesting appropriate responses, thus anticipating their interventions. This guidance can be supported by the history of interactions, where information is extracted and frequent dialog flows are discovered, allowing an automatic extraction and representation of them to guide humans. The approach will be divided into three components: Extraction to process dialogues and use the information to describe/classify interactions; Representation for the discovery of the most frequent dialogue flows, represented by graphs of interaction classes; Guidance to guide the agent during a new dialogue. All of the previous are often tackled in the scope of Dialog Modelling. Research in the area of dialogs is currently booming, with interest in chatbots. However, most systems are English-focused, so the focus of this work has innovative potential in the area and context.

Keyphrases: Assisted guidance in dialog, Dialog Information Extraction dialog, Natural Language Processing, Representation of dialog flows, dialog analysis

@booklet{EasyChair:8132,
  author    = {Patrícia Ferreira},
  title     = {Automatic Dialogue Flow Extraction and Guidance},
  howpublished = {EasyChair Preprint 8132},
  year      = {EasyChair, 2022}}