Joint Reasoning for Multi-Faceted Commonsense Knowledge
This was the subject of my internship at the Max-Planck Insitute for Informatics in Saarbrücken, Germany. I worked under the supervision of Gerhard Weikum and Simon Razniewski. It resulted in a paper published in Automatic Knowledge Base Construction (AKBC) 2020 (see the OpenReview and the PDF). The paper is accompanied by a web demonstration, developped as a Django application.
Commonsense knowledge (CSK) supports a variety of AI applications, from visual understanding to chatbots. Prior works on acquiring CSK, such as ConceptNet, have compiled statements that associate concepts, like everyday objects or activities, with properties that hold for most or some instances of the concept. Each concept is treated in isolation from other concepts, and the only quantitative measure (or ranking) of properties is a confidence score that the statement is valid. This paper aims to overcome these limitations by introducing a multi-faceted model of CSK statements and methods for joint reasoning over sets of inter-related statements. Our model captures four different dimensions of CSK statements: plausibility, typicality, remarkability and salience, with scoring and ranking along each dimension. For example, hyenas drinking water is typical but not salient, whereas hyenas eating carcasses is salient. For reasoning and ranking, we develop a method with soft constraints, to couple the inference over concepts that are related in in a taxonomic hierarchy. The reasoning is cast into an integer linear programming (ILP), and we leverage the theory of reduction costs of a relaxed LP to compute informative rankings. This methodology is applied to several large CSK collections. Our evaluation shows that we can consolidate these inputs into much cleaner and more expressive knowledge.
A former intern of the MPI, Julien Romero, had just published Quasimodo, a new commensense knowledge collection based on smart extraction of query logs. My internship was dedicated to improving this work.
In a first attempt at getting a deep understanding of the subject, I worked a lot on organizing the collection under a taxonomy of concepts, the concepts being the subjects from the facts. It was a good way to explore the data, understanding how balanced it was, which parts were of high quality and which were not. Attempts at cleaning the taxonomy lead to the comparison of facts known about nodes linked together.
The approach was promising. We therefore focused on generalizing and improving the comparisons between facts. Thus, we explored tools for logicial reasoning, before selecting integer linear programming for its bests results. It works by maximizing an objective function, of which we fine-tuned the terms with multi-sourced "evidence" scores, such as word embeddings similarity.
That was the first three months of the internship. Nexts steps consisted in the method generalization, tthe heoretical grounding of our choices, large scale experiments, and finally the paper redaction. The deadline was the submission to the WSDM conference, coinciding with the end of my internship. Unfortunately, this first submission got rejected. In the next months, more fine-tuning was done, leading to the final submission to AKBC.