For many complex tasks, humans are far more intelligent than the best AI systems. While the large language models (LLMs), notably ChatGPT, have demonstrated outstanding performance in performing many NLP tasks, their ability of performing logical reasoning and explanation of their behavior is quite limited due to the limitation of their underlying neural network architecture. This limitation significantly limits their utility since both unreliable inferences (halluciation) and lack of explanation (provenance) would undermine the trustworthiness of any such intelligent system. The more critical an application is, the higher level of trustworthiness is needed. To break this limitation, we need to study how to build human-like intelligent systems. Human brains are known to have two somewhat separate systems, i.e., System 1 and System 2. System 1 is a fast intuitive but unreliable system, quite similar to the current neural networks. System 2 is a slow symbolic system that can perform logic reasoning and planning. The current LLMs and intelligent systems in general appear to be able to successfully simulate human System 1, but how to extend it to further simulate System 2 remains a difficult challenge, which requires revision to or extension of the current transformer-based deep neural network architecture, and many neuro-symbolic models/architectures have already been develoed precisely for this purpose, in the broad context of neuro-symbolic AI. However, we are far from having a neuro-symbolic system that can simulate both System 1 and System 2 of human brains in an integrated manner. We are interested in tackling this problem from the following perspectives:

• Competence-based Analysis and Evaluation of Language Models: LLMs are currently evaluated primarily based on their performances on many specific tasks. Such a pure performance-based evaluation method generally does not tell us the generalization capacity of an intelligent system; indeed, we can never be sure whether a model would perform similarly well or make any prediction about their (future) behavior when the model is to be used on a future application scenario, especially if the scenario has not been "seen" by the mode in the training data. To be able to predict the behaviors of an LLM, we must "open up" an LLM, examine its internal representation, and perform analysis of the competence of an intelligent system. To this end, we have proposed a general framework for analyzing and evaluating competence of language models, called CALM. It enables quantitative evaluation of competence of LLMs using task-based counterfactural probing.
• Learning Interpretable Knowledge Representation with Neural Networks: The current LLMs encode knowledge in a non-interpretable way, which hinders their ability to explain their predictions. Humans are able to learn interpretable knowledge from the observed data. We are interested in studying how to use neural networks to learn more interpretable representations (e.g., a knowledge graph). Such an interpretable knowledge reresentation can also be expected to be more generalizable than the current implicit non-interpretable knowledge representation, leading to more predictable behaviors of the model. Moreover, it would also enable humans to directly evaluate the learned representation and improve it directly. A small step toward this direction that we have made is the development of an extension of BERT to incorporate Sparse Latent Type (SLT) regularization, published as an EMNLP 22 paper.
• Computational models of human cognition: Ultimately, in order to build a human-like intelligent system, we must have a good understanding of how human brains work. We are thus also interested in how to use a computational model to accurately simulate human cognition and problem solving since such a computational model will be able to provide insights about how to design a more human-like neuro-symbolic neural network architecture so as to add reasoning ability to the current LLMs. Here's a very old work(a failed submission to Cognitive Science 1998) on a novel computational model that can simultaneously simulate language acquisition and concept attainment using discrimination and generalization on a semantic memory structure for representing language and concepts.