SenticNet 7: A Commonsense-based Neurosymbolic AI Framework for Explainable Sentiment Analysis

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3829–3839 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3829 SenticNet 7: A Commonsense-based Neurosymbolic AI Framework for Explainable Sentiment Analysis Erik Cambria,1 Qian Liu,1 Sergio Decherchi,2 Frank Xing,3 Kenneth Kwok4 1 School of Computer Science and Engineering, Nanyang Technological University (NTU), Singapore 2 Fondazione Istituto Italiano di Tecnologia (IIT), Italy 3 School of Computing, National University of Singapore (NUS), Singapore 4 Agency for Science, Technology and Research (A*STAR), Singapore {cambria,liu.qian}@ntu.edu.sg, sergio.decherchi@iit.it, xing@nus.edu.sg, kenkwok@ihpc.a-star.edu.sg Abstract In recent years, AI research has demonstrated enormous potential for the benefit of humanity and society. While often better than its human counterparts in classification and pattern recognition tasks, however, AI still struggles with complex tasks that require commonsense reasoning such as natural language understanding. In this context, the key limitations of current AI models are: dependency, reproducibility, trustworthiness, interpretability, and explainability. In this work, we propose a commonsense-based neurosymbolic framework that aims to overcome these issues in the context of sentiment analysis. In particular, we employ unsupervised and reproducible subsymbolic techniques such as auto-regressive language models and kernel methods to build trustworthy symbolic representations that convert natural language to a sort of protolanguage and, hence, extract polarity from text in a completely interpretable and explainable manner. Keywords: Neurosymbolic AI, sentiment analysis, natural language processing 1. Introduction In 2012, a large-scale, international public effort was launched with the goal of reconstructing the full record of neural activity across complete neural cir- cuits (Alivisatos et al., 2012). Ten years later, we still do not understand much about how the human brain works: we know well its hardware, but we are still pretty clueless about its operating system. Much ear- lier, the field of AI was born as an attempt to emu- late human intelligence into machines. Today, how- ever, most of AI research has regressed to the mim- icking of intelligent behavior. Rather than ‘artificial’, such research should probably go under the name of ‘pareidoliac’ intelligence, as it tends to develop expert systems while claiming that these embed some sort of intelligence. Intelligence, however, is a ‘suitcase’ word that encapsulates several cognitive processes such as learning, attention, memory, decision making, emo- tions, and many more we may not even know about. While recent developments have generated AI models that achieve human-like performance in many classi- fication and pattern recognition tasks, AI still strug- gles with complex tasks that require more than just en- coding joint probabilities or co-occurrence frequencies. Moreover, even the finest AI models are vulnerable to adversarial examples (Goodfellow et al., 2015). The five key problems with AI today are: depen- dency, reproducibility, trustworthiness, interpretabil- ity, and explainability. In this work, we propose a commonsense-based neurosymbolic AI framework that aims to overcome these issues in the context of senti- ment analysis. The framework is neurosymbolic be- cause it leverages both subsymbolic and symbolic AI to perform polarity detection from text. In particular, subsymbolic techniques such as …

Content truncated to 500 words