Reasoning about actions and change (RAC) is essential to understand and interact with the ever-changing environment. Previous AI research has shown the importance of fundamental and indispensable knowledge of actions, i.e., preconditions and effects. However, traditional methods rely on logical formalization which hinders practical applications. With recent transformer-based language models (LMs), reasoning over text is desirable and seemingly feasible, leading to the question of whether LMs can effectively and efficiently learn to solve RAC problems. We propose four essential RAC tasks as a comprehensive textual benchmark and generate problems in a way that minimizes the influence of other linguistic requirements (e.g., grounding) to focus on RAC. The resulting benchmark, TRAC, encompassing problems of various complexities, facilitates a more granular evaluation of LMs, precisely targeting the structural generalization ability much needed for RAC. Experiments with three high-performing transformers indicate that additional efforts are needed to tackle challenges raised by TRAC.

Table fact verification aims to check the correctness of textual statements based on given semi-structured data. Most existing methods are devoted to better comprehending logical operations and tables, but they hardly study generating latent programs from statements, with which we can not only retrieve evidences efficiently but also explain reasons behind verifications naturally. However, it is challenging to get correct programs with existing weakly supervised semantic parsers due to the huge search space with lots of spurious programs. In this paper, we address the challenge by leveraging both lexical features and structure features for program generation. Through analyzing the connection between the program tree and the dependency tree, we define a unified concept, operation-oriented tree, to mine structure features, and introduce Structure-Aware Semantic Parsing to integrate structure features into program generation. Moreover, we design a refined objective function with lexical features and violation punishments to further avoid spurious programs. Experimental results show that our proposed method generates programs more accurately than existing semantic parsers, and achieves comparable performance to the SOTA on the large-scale benchmark TABFACT.

The logical reasoning capabilities of pre-trained language models have recently received much attention. As one of the vital reasoning paradigms, non-monotonic reasoning refers to the fact that conclusions may be invalidated with new information. Existing work has constructed a non-monotonic inference dataset 𝛿-NLI and explored the performance of language models on it. However, the 𝛿-NLI dataset is entangled with commonsense reasoning. In this paper, we explore the pure non-monotonic reasoning ability of pre-trained language models. We build a non-monotonic reasoning benchmark, named LogicNMR, with explicit default rules and iterative updates. In the experimental part, the performance of popular language models on LogicNMR is explored from the perspectives of accuracy, generalization, proof-based traceability and robustness. The experimental results show that even though the fine-tuned language models achieve an accuracy of more than 94.4% on LogicNMR, they perform unsatisfactorily, with a significant drop, in generalization and proof-based traceability.

Recent methods based on pre-trained language models have shown strong supervised performance on commonsense reasoning. However, they rely on expensive data annotation and time-consuming training. Thus, we focus on unsupervised commonsense reasoning. We show the effectiveness of using a common framework, Natural Language Inference (NLI), to solve diverse commonsense reasoning tasks. By leveraging transfer learning from large NLI datasets, and injecting crucial knowledge from commonsense sources such as ATOMIC 2020 and ConceptNet, our method achieved state-of-the-art unsupervised performance on two commonsense reasoning tasks: WinoWhy and CommonsenseQA. Further analysis demonstrated the benefits of multiple categories of knowledge, but problems about quantities and antonyms are still challenging.

The recent success of neural language models (NLMs) on the Winograd Schema Challenge has called for further investigation of the commonsense reasoning ability of these models. Previous diagnostic datasets rely on crowd-sourcing which fails to provide coherent commonsense crucial for solving WSC problems. To better evaluate NLMs, we propose a logic-based framework that focuses on high-quality commonsense knowledge. Specifically, we identify and collect formal knowledge formulas verified by theorem provers and translate such formulas into natural language sentences. Based on these true knowledge sentences, adversarial false ones are generated. We propose a new dataset named WinoLogic with these sentences. Given a problem in WinoLogic, NLMs need to decide whether the plausible knowledge sentences could correctly solve the corresponding WSC problems in a zero-shot setting. We also ask human annotators to validate WinoLogic to ensure it is human-agreeable. Experiments show that NLMs still struggle to comprehend commonsense knowledge as humans do, indicating that their reasoning ability could have been overestimated.