Towards a Self-aware Intelligent Agent through Automated Output Interpretation and Improvisation

The aim of this thesis is enhance the behaviour of an intelligent agent analysing and suggesting improvements in its self-evaluation of exploration, interpretation and improvisation of general knowledge stored in a knowledge graph. We propose an evaluation of these aspects through metrics that highlight five key components: adherence, cohesion, diversity, serendipity and magnitude. Our work is built on two basic units, namely a triple and a document: in the first step, we extract data from Wikidata, an open structured knowledge graph, pre-processing each page as a document. Each one is composed by statements that we shape in simplified triples that we store and adapt into a graph knowledge base. By drawing strength from both graph theory and modern deep learning technics, utilizing graph and subgraph embeddings, our agent evaluates the input document using adherence, cohesion, diversity, serendipity and magnitude taking into account existing information in the graph. Subsequently, we let the agent applies computational creative combinatory to produce a novel document, keeping on evaluation. Several aspects have to bearing in mind: incompleteness of a Knowledge Graph, context recognition, proper merge between graph and deep learning strategy, time dimension, bias, large amount of data. Multivariate regression and Pareto optimum will be used to analyse correlation between key components of our metrics, in order to establish parameters to drive the agent behaviour. At the same time we will compute the time derivative of this index and find if is perceived as improvisation. Given a score to improvisation is tricky because there is not a stable and shared definition of it and depends on who perceives it. The final work will be evaluated comparing state of the art metrics and user surveys on heterogeneous groups of people. This phase of study is an opportunity to investigate early introspective steps of future conversational agents, which should attempt to become self-awareness, in order to further develop capacity to understand information they store, with the goal to interact with users or other agents in a non-trivial way.