Towards automated axiom generation : a semi-automated approach to generating "knowledge and rule base" corpora from text narratives

Prabhu, Anirudh
Thumbnail Image
Other Contributors
Hendler, James A.
Nirenburg, Sergei
Ma, Xiaogang
Fox, Peter A.
McGuinness, Deborah L.
Issue Date
Multidisciplinary science
Terms of Use
Attribution-NonCommercial-NoDerivs 3.0 United States
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute (RPI), Troy, NY. Copyright of original work retained by author.
Full Citation
With the exponential rise of data in recent years, deep learning has risen to be one of the most prominent forms of artificial intelligence. With many successful applications, deep learning has helped researchers build machines that successfully complete human tasks previously thought to be very difficult. For example, restoring color to black and white photos, image captioning, voice generation, restoring sound in silent videos, lip reading from videos etc. are some very interesting applications being explored and with deep learning. Even with success in a breadth of applications, there are still problems that deep learning has not been able to solve. For example, scalability, understanding context, or examining and understanding the inner workings of deep neural networks themselves remain unsolved problems. The crux of the deep learning approach are “layers” (input, hidden and output). These layers are adjustable to a given corpus and mostly opaque to interpretation or explanation. Current approaches to Artificial Intelligence/Machine Learning rely on an entire corpus, i.e. they use the entire content with noise, bias, etc. These approaches have achieved high success across fields like computer vision, natural language processing, image captioning etc., require a very large amount of training data to accurately understanding the mappings between the input and output embeddings for the deep learning experiment. In this thesis, we ask the question, "What if, intelligent information extraction (both entity and relation) were able to provide a curated corpus for deep network learning?" Curation in this context, addresses eliminating all the "non-essential" parts of text, and simply focusing on the actions, agents and events involved in a text corpus, and the rules that highlight the effects of these actions and change in the narrative. What is needed to achieve this task, is the ability to recognize key entities and map the situational changes occurring in the corpus to specific triggers (such as actions or events), like those seen in axioms in a rule base. Automated axiom creation is a difficult research problem to solve. Most of the work in this area focuses on rules extracting rules from text that explicitly mentions the rule in text. In most old fashioned AI systems, rules are developed with an understanding of the domain and reading between the lines where required to see what action/events could trigger a particular response in the narrative. An automated axiom creation method that completes such a task is still an unexplored and unsolved problem, and the focus of this thesis. The biggest hurdle in exploration of this research problem, is the availability of data (or the lack thereof) where implicit rules are documented for a text narrative. In this thesis, we have developed a novel semi-automated method to generate axioms/rules for a set of text narratives, using crowdsourcing and known natural language processing techniques. We begin with textual narrative such as those in novels, computer manuals, but also in view are scientific works. We then document rules for the given narrative by using Amazon Mechanical Turk, a crowdsourcing platform known to aid in the creation of high-quality datasets. We have found that the usage of a crowdsourcing platform works well for narratives that do not require any expertise, like those seen in novels, and are able to provide textual rules for the narrative which may not be explicitly stated in the text. The next step would be to process these narratives and their rules into knowledge bases and rule bases, where the key concepts and relationships need to be extracted from both the knowledge bases (in the form of triples) and rule base. We have also developed an approach to converting the results of the crowdsourcing experiment (rules in text form), to formal rules in a rule base. These are developed based on known NLP information extraction techniques, like POS tagging, co-reference resolution etc. The overall goal of this thesis is a novel method to extract key information and rules from a narrative , in order to create a set of knowledge bases and rule bases. After examining the results of the crowdsourcing experiment, we found a set of boundaries for the usability of crowdsourcing as tool or means to overcome the automation bottleneck. We also discuss the required distinction of the terms "Humans in the loop" and "Experts in the loop", and provide a platform for fleshing out the framework for experts in the loop for scientific workflows. Finally we also developed a method to evaluate "knowledge base - rule base" corpora for any logical language in any domain. To construct such a "situational narrative", a formalism such as situation calculus stands out as an obvious choice for knowledge representation but is heretofore an unexplored option in explaining “what is going on” in deep learning. At the heart of such a capability may be a learned formalization of the “situation” and perhaps even the identification of changes in state or fluent(s) (situation) over iterations or after learning interactions.
August 2021
School of Science
Multidisciplinary Science Program
Rensselaer Polytechnic Institute, Troy, NY
Rensselaer Theses and Dissertations Online Collection
CC BY-NC-ND. Users may download and share copies with attribution in accordance with a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 license. No commercial use or derivatives are permitted without the explicit approval of the author.