Knowledge Representation in Natural Language (KRNL-18)

Natural languages spoken by humans are arguably the most natural medium for humans to express their knowledge about the world. Unfortunately, representing knowledge in the form of natural languages is a challenging task due to the ambiguities that exist in natural languages. Processing knowledge expressed in the form of natural languages, or performing logical inferences directly at the level of human spoken languages is further complicated by the lack of formal structure in human languages.

To avoid these difficulties involved in KR directly at the level of natural languages, the KR community has resort to formal logical representations such as first-order logic or description logics. Once knowledge is extracted from human languages and represented in such abstract formats, a plethora of mathematical tools are at our disposal to perform inferences at scale.

NLP is the branch of AI that considers the problem of extracting and processing knowledge expressed in the form of human languages. Starting from the early work on corpus analysis using word-counting approaches, the NLP community has significantly advanced over the recent years. Accurate semantic representations for elementary lexical units such as words, and compositional approaches that can build semantic representations for larger lexical units such as phrases, sentences or documents have been developed. Moreover, tasks that require some form of logical inference at the level of languages such as recognising textual entailment (RTE), natural language inference in knowledge bases argument mining and semantic parsing have established as central research topics in the NLP community.

KR and NLP communities have so far worked independently and the inter-community communications have been intermittent and sparse. However, given the above-mentioned recent developments, we believe the two communities are at cross-roads, approaching an important junction. The two communities have much to learn from each other and share their experiences in related, yet complementary topics.

To provide one example, the KR community can benefit from the unsupervised knowledge extraction and representation methods developed in the NLP community to overcome the knowledge extraction bottleneck, whereas the NLP community can benefit from the efficient inference algorithms studied over the years in the KR community. We feel the urge to initiate a dialogue between the KR and NLP communities on facilitate this interaction. KRNL-18 is an attempt towards this much needed interaction.