Commonsense Knowledge in Word Associations and ConceptNet

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 481–495 November 10–11, 2021. ©2021 Association for Computational Linguistics 481 Commonsense Knowledge in Word Associations and ConceptNet Chunhua Liu Trevor Cohn Lea Frermann School of Computing and Information Systems The University of Melbourne chunhua@student.unimelb.edu.au {tcohn,lfrermann}@unimelb.edu.au Abstract Humans use countless basic, shared facts about the world to efficiently navigate in their environment. This commonsense knowledge is rarely communicated explicitly, however, un- derstanding how commonsense knowledge is represented in different paradigms is important for both deeper understanding of human cog- nition and for augmenting automatic reason- ing systems. This paper presents an in-depth comparison of two large-scale resources of general knowledge: ConceptNet, an engi- neered relational database, and SWOW a knowl- edge graph derived from crowd-sourced word associations. We examine the structure, over- lap and differences between the two graphs, as well as the extent to which they encode situ- ational commonsense knowledge. We finally show empirically that both resources improve downstream task performance on common- sense reasoning benchmarks over text-only baselines, suggesting that large-scale word association data, which have been obtained for several languages through crowd-sourcing, can be a valuable complement to curated knowledge graphs.1 1 Introduction Humans understand and navigate everyday situa- tions with great efficiency using a wealth of shared, basic facts about their social and physical envi- ronment – a resource often called commonsense knowledge (Liu and Singh, 2004). Advances in ar- tificial intelligence in general, and natural language processing in particular, have led to a surge of in- terest in its nature: what constitutes commonsense knowledge? And despite this knowledge being rarely explicitly stated in text (Gordon and Durme, 2013), how can we equip machines with common- sense to enable more general inference (Davis and 1Code available at https://github.com/ ChunhuaLiu596/CSWordAssociation camping enjoyable bed roll activity glamping encampment fire HasProperty RelatedTo IsA RelatedTo RelatedTo DerivedFrom Synonym nature ground adventure woods trail family marshmallow hike (b) SWOW (a) ConceptNet Figure 1: Sub-graphs centered around ‘camping’ from (a) ConceptNet and (b) SWOW. Nodes in green are common to both KGs. Nodes on the left/blue (right/cyan) are unique to ConceptNet (SWOW). Marcus, 2015)? Recently, large language mod- els (Devlin et al., 2019; Radford et al., 2019) pre- trained on massive text corpora achieved promis- ing results on commonsense reasoning benchmarks with fine-tuning, however, the lack of interpretabil- ity remains a problem. Augmenting them with commonsense knowledge can provide complemen- tary knowledge (Ilievski et al., 2021a; Safavi and Koutra, 2021) and thus make models more robust and explainable (Lin et al., 2019). Attempts to capture general human knowledge include inventories of machine-readable logical rules (Gordon and Hobbs, 2017) and large, cu- rated databases which have been collected with the specific purpose to reflect either domain gen- eral (e.g., ConceptNet; Liu and Singh (2004)) or domain-specific (e.g., ATOMIC; Sap et al. (2019)) commonsense knowledge. Word association norms are a third resource of explicit, basic human knowledge: in a typical study, human participants are presented with a cue word (e.g., ‘camping’) and …

Content truncated to 500 words