Skip to main content

Assessing Hypotheses in Multi-Agent Systems for Natural Language Processing

Abstract

In multi-agent systems (MAS) different agents work on a common problem. Such systems are also used in natural language processing (NLP). Agents of an MAS for natural language can generate results with confidence, so called hypotheses. These hypotheses reflect the ambiguity of natural language. If agents are dependent on each other, a wrong hypothesis can quickly lead to error propagation into the hypotheses of the dependent agents. The exploration of hypotheses offers the chance to improve the results of agents. This thesis improves the results of agents of a MAS for NLP by a controlled exploration of the hypothesis search space. Therefore, a framework for the exploration and evaluation of hypotheses is developed. In an evaluation with three agents promising results regarding the improvement could be achieved. For example, Top-X Exploration achieved an average improvement of the F1 score of the Topic Detection agent from originally 40% to now 49% and of the Ontology Selection agent from originally 74% to 79%.

Download PDF

BibTeX

@masterthesis{fuchss_assessing_2020,
author = {Fuch{\ss}, Dominik},
school = {Karlsruher Institut für Technologie (KIT)},
title = {{Assessing Hypotheses in Multi-Agent Systems for Natural Language Processing}},
type = {Master's Thesis},
doi = {10.5445/IR/1000126806},
keywords = {Natural Language Processing, Hypotheses, Multi-Agent Systems},
year = {2020}
}