Year: 2022
Type: dissertation
Abstract: Human cognition is exciting, it is a mesh up of several neural phenomena which really strive our ability to constantly reason and infer about the involving world. In cognitive computer science, Commonsense Reasoning is the terminology given to our ability to infer uncertain events and reason about Cognitive Knowledge. The introduction of Commonsense to intelligent systems has been for years desired, but the mechanism for this introduction remains a scientific jigsaw. Some, implicitly believe language understanding is enough to achieve some level of Commonsense [90]. In a less common ground, there are others who think enriching language with Knowledge Graphs might be enough for human-like reasoning [63], while there are others who believe human-like reasoning can only be truly captured with symbolic rules and logical deduction powered by Knowledge Bases, such as taxonomies and ontologies [50]. We focus on Commonsense Knowledge integration to Language Models, because we believe that this integration is a step towards a beneficial embedding of Commonsense Reasoning to interactive Intelligent Systems, such as conversational assistants. Conversational assistants, such as Alexa from Amazon, are user driven systems. Thus, giving birth to a more human-like interaction is strongly desired to really capture the user’s attention and empathy. We believe that such humanistic characteristics can be leveraged through the introduction of stronger Commonsense Knowledge and Reasoning to fruitfully engage with users. To this end, we intend to introduce a new family of models, the Relation-Aware BART (RA-BART), leveraging language generation abilities of BART [51] with explicit Commonsense Knowledge extracted from Commonsense Knowledge Graphs to further extend human capabilities on these models. We evaluate our model on three different tasks: Abstractive Question Answering, Text Generation conditioned on certain concepts and aMulti-Choice Question Answering task. We find out that, on generation tasks, RA-BART outperforms non-knowledge enriched models, however, it underperforms on the multi-choice question answering task. Our Project can be consulted in our open source, public GitHub repository (Explicit Commonsense).
Author: Vicente, Frederico Miguel Guerra Paulo Pereira
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7006120048