An Open-Ended Web Knowledge Retrieval Framework for the Household Domain With Explanation and Learning Through Argumentation

Introduction

The use of Semantic Web Knowledge Bases (KBs) can provide both explainability and scalability to the knowledge representation component of cognitive robotic systems. In some cases, Semantic Web KBs, such as ConceptNet (Liu & Singh, 2004) and WordNet (Fellbaum, 2010), can provide commonsense reasoning skills (see Section 2). Moreover, it is important for a knowledge retrieval framework to be able to argue over and explain the answers it returns.

Ideally, a knowledge retrieval framework should answer as many queries as possible, regardless of the complexity of the involved entities of the query; in practice, high levels of completeness cannot easily be achieved. This is due to the fact that the quality of knowledge stored in any KB that is selected for knowledge retrieval purposes is limited for various reasons, thus, affecting the answers that it can return. A possible solution would be to use external knowledge sources and, specifically, Semantic Web KBs, which can provide commonsense knowledge (Zamazal, 2020).

In addition, it is desirable for the knowledge retrieval framework to justify the answers that it returns in response to a query. One common method to justify an opinion is through argumentation (Vassiliades et al., 2021a). The use of argumentation to support the validity of the answers returned by a knowledge retrieval framework can help the framework provide a human-friendly way of explanation. Moreover, argumentation in many cases can act as a method of learning new knowledge, if one’s opinion is proved wrong through an argumentation dialogue. Thus, argumentation can also assist a knowledge retrieval framework to learn new knowledge that can be used in the future.

Considering the aforementioned observations, the problem that this paper aims to address is, given the contextual information relevant to a household environment, to construct a knowledge retrieval framework for the household domain, enhanced with external knowledge sources, that can argue over the answers that it returns and learn new knowledge, by means of an argumentation process¹. The research questions that emerge are the following: a) “How can a domain-specific knowledge retrieval framework be extended, so as to retrieve knowledge that exists in several knowledge bases?”, b) “How can a knowledge retrieval framework support its answers in a way closer to the human way of reasoning?”, c) “How can a knowledge retrieval framework learn new knowledge with methods closer to human learning?”, and d) “How can a knowledge retrieval framework be evaluated when there is no clear pipeline for the evaluation?”. Question (a) is what this study tries to tackle with the knowledge retrieval component, questions (b), (c) are addressed with the learning-through-argumentation component, and question (d) is addressed in the evaluation section, where flexible methods on how to evaluate a framework with user evaluations are presented.

The framework provides knowledge about sequences of actions on how to perform human tasks in a household environment, answers queries about household objects, and performs semantic matching between entities originating from the web knowledge graph ConceptNet with the ones that exist in the internal knowledge graph, using knowledge from DBpedia (Bizer et al., 2009) and WordNet (see Figure 1). The framework offers a set of predefined SPARQL templates that directly address the ontology of the internal KB, as well as an API for general-purpose querying through SPARQL. The knowledge of the internal KB was extracted from the VirtualHome dataset (Liao et al., 2019; Puig et al., 2018). The framework also features an argumentation component, where the user can argue against the answers of the knowledge retrieval component of the framework under two different scenarios; the missing knowledge scenario, where an entity can be found in the answers of the framework, according to the user, and the wrong knowledge scenario, where an entity does not exist in the answers of the framework. Finally, the framework can learn new knowledge through argumentation, if the argumentation dialogue ends in favor of the human user (see Figure 4).