Analogy Retrieval Through Textual Inference

Analogy-making has been deemed one of the core cognitive mechanisms which play a role in human creative thinking activities such as design and art. Designers can make use of analogies in various stages of design including ideation, planning and evaluation. However, human analogy-making is limited by experience and reliance of human memory on superficial attributes rather than relational or causal structure during analogy retrieval. In this regard, different design-by-analogy tools have been developed to assist designers in analogical reasoning. Analogical reasoning tools can be viewed as either based on hand-coded structured knowledge or natural-language-based design-by-analogy tools. The former are naturally limited in extent and scope to that which was hand coded [1]. Alternatively, natural language analogical reasoning can leverage the abundantly available textual resources. Current text-based analogy research for design have relied on analogies between individual word meanings. This leaves open consideration of the relational structure of the language where the relational similarity of texts can indicate a significant analogy. In this article, we develop four computational models of analogy that capture relational structure of the text. This includes spatial representation of semantics, multi-level deep neural reasoning, graph matching based model and transformation-based model. The models are then combined together into an ensemble model to achieve acceptable level of analogical accuracy for the end-user. The underlying design-related knowledge upon which analogies were drawn includes engineering ontologies, function hierarchy and raw patent texts. Instantiating this analogical reasoning model in design concept analogy retrieval system, we show this approach can help retrieve meaningful analogies from the World Intellectual Property Organization (WIPO) patent repository. We demonstrate this for a particular design problem.