Knowledge discovery in task-oriented dialogue

Highlights

• Method for knowledge discovery in task-oriented dialogues.

• Knowledge is represented using folksonomies.

• The knowledge represented by folksonomies can be used to interpret new utterances.

• The folksonomy can be used to discover Topics Addressed by interlocutors.

Abstract

Knowledge discovery is the process of discovering useful knowledge in a broad range of sources, such as relational databases, images, or texts. Dialogues are generated by interaction between people using natural language and can be used as a source of information. Once discovered, knowledge needs to be represented, and there are several approaches to this. In this paper, we propose a method to discover knowledge in task-oriented dialogues by representing these dialogues through folksonomies, using a novel quadripartite model. Folksonomies are knowledge structures composed of users, tags, and resources. Dialogues and folksonomies have a social dimension in common, which renders folksonomies suited to representing knowledge discovered from dialogues. The knowledge represented by folksonomies can be used to interpret new utterances in a dialogue and detect trends, e.g., by discovering Topics Addressed by people at different time intervals, in the dialogues used to learn the folksonomies. The main difference between our approach and past techniques is that we use the characteristics (the content) of each resource in the discovery process. Experiments involving a real-world task-oriented dialogue corpus showed that using our method, learned folksonomies can interpret utterances with an accuracy of 72.32%. Moreover, another experiment showed that it is possible to use our method to determine Topics Addressed by interlocutors in dialogues.

Introduction

The process of extracting useful, implicit, and previously unknown knowledge from large amounts of data is known as Knowledge Discovery in Databases (KDD) (Lara, Lizcano, Martínez, & Pazos, 2014). KDD has been used in a broad range of applications across a variety of domains, such as to improve the analysis of marketing and business databases (Orriols-Puig, Martínez-López, Casillas, & Lee, 2013), extract knowledge from structural medical data (Esfandiary, Babavalian, Moghadam, & Tabar, 2014), and to monitor water quality using hydrological data (Alatrista-Salas et al., 2014). Due to the rapidly growing amounts of digital data, there is a pressing need for theories and tools that can support the extraction of useful information (knowledge) from them. KDD aims to map low-level data into other forms that are more compact, abstract, and useful (Fayyad, Piatetsky-Shapiro, & Smyth, 1996). The knowledge discovered may be used for analysis, trends, classification, group identification, behavior forecasting, etc. Once knowledge has been obtained, it needs to be represented in some form, such as ontologies, frames, or folksonomies.

Various sources may be used for KDD: relational databases, and structured and non-structured texts or images are examples. However, dialogues have not yet been explored as source of knowledge. Dialogues are interactions between speakers and listeners, called interlocutors, and are composed of speech acts (utterances). Of the several types of dialogues, task-oriented dialogue aims to solve a given task in a given domain. Such dialogues generate the solution to a task, requested by someone in order to accomplish something, in a concise sequence. Thus, a remarkable characteristic of task-oriented dialogues is that they involve two kinds of interlocutors: one that asks for help, and another that possesses knowledge of the relevant domain, and assists the former kind in solving the task at hand. Table 1 shows an example of a task-oriented dialogue.

According to Traum and Hinkelman (1992), one of the main characteristics of task-oriented dialogue is the dissemination of knowledge, i.e., the interlocutor with more knowledge transfers it to the one asking for help (Carletta et al., 1997). This kind of dialogue is now common on the Internet. Several companies offer customer support by way of live chat, where an attendant answers questions from customers (Elmorshidy, 2011). In this paper, we refer to as “attendant” the interlocutor who has knowledge of the domain and “user” the interlocutor who is asking for help, as in a customer support center (e.g., help desk system).

Our research in this paper aims to discover knowledge from task-oriented dialogues. Once discovered, the knowledge must be represented. Of the several alternatives available, we choose folksonomies for the following reasons: (i) both dialogue and folksonomies have in common a social dimension; (ii) folksonomies directly reflect the vocabulary of common users, and thus can represent discovered knowledge more faithfully; and (iii) folksonomies are simpler than other knowledge structures, such as ontologies. It is important to mention that since we had some success with folksonomies, we plan in future research to develop a method to learn ontologies from dialogues.

Folksonomies are structures of knowledge representation that emerge from tagging in collaborative tagging systems (Peters, 2009). Tagging is the assignment of tags to resources by users. Thus, folksonomies comprise users, tags, and resources. A resource can be any object that users are interested in tagging, such as photos and videos. In comparison with ontologies, folksonomies are simpler structures to implement and use (Echarte, Astrain, Córdoba, & Villadangos, 2007). According to Hotho, Jäschke, Schmitz, and Stumme (2006a), one of the benefits of tagging is that users do not need experience or a particular skill to participate, i.e., the folksonomies that emerge do not need to be built by knowledge engineers. Moreover, ontologies have a controlled vocabulary derived by consensus, which needs to be attained among the participants of the knowledge-building process (knowledge engineers, domain experts, etc.). By contrast, folksonomies directly reflect the vocabulary of common users, since lay users assign tags to resources (Quintarelli, 2005). As a consequence, folksonomies, unlike ontologies, are untroubled by the large amount of information at hand or the need for consensus. A major characteristic of folksonomies is their social dimension (users), which is also part of dialogues, due to the interaction between users. This characteristic renders folksonomies suitable for representing knowledge discovered from dialogues.

In this paper, we are interested in discovering knowledge in dialogues and introduce a method to learn folksonomies from task-oriented dialogues. The knowledge represented by folksonomies may be used for the interpretation of new dialogue utterances and for trend detection, e.g., discovering Topics Addressed by people at different time intervals in dialogues used to learn the folksonomies. Trending topics are the ones being discussed more than others, and are useful measures of popularity on social networking services, such as Twitter. In order to verify whether the structures created by our method are genuine folksonomies, we performed an experiment to show that they exhibit the small-world phenomenon (Milgram, 1967), which is a characteristic of folksonomies (Cattuto et al., 2007). We also confirmed that our learned folksonomies can interpret dialogue utterances. For this, we performed an experiment to measure the accuracy of the folksonomies in interpreting utterances to determine whether they belong to the domain represented by the folksonomies. Moreover, we conducted an experiment to discover trending topics using the learned folksonomies.

The main contribution of this paper is a method to facilitate the learning of folksonomies from dialogues. To the best of our knowledge, ours is the first published proposal for learning folksonomies from dialogues.

This paper is organized as follows: Section 2 presents the concept of a folksonomy, and Section 3 contains a description of the FolksDialogue method. In Section 4, we present our proposed approach to trend detection in folksonomies. The experiments that we conducted and the results obtained are detailed in Section 5. We survey related works in the area in Section 6, and offer our conclusions as well as directions for future work Section 7.