Despite recent breakthroughs in Machine Learning for Natural Language Processing, the Natural Language Inference (NLI) problems still constitute a challenge. To this purpose, we contribute a new dataset that focuses exclusively on the factivity phenomenon; however, our task remains the same as other NLI tasks, that is prediction of entailment, contradiction, or neutral (ECN). In this paper, we describe the LingFeatured NLI corpus and present the results of analyses designed to characterize the factivity/non-factivity opposition in natural language. The dataset contains entirely natural language utterances in Polish and gathers 2432 verb-complement pairs and 309 unique verbs. The dataset is based on the National Corpus of Polish (NKJP) and is a representative subcorpus in regard to syntactic construction [V][że][cc]. We also present an extended version of the set (3035 sentences) consisting more sentences with internal negations. We prepared deep learning benchmarks for both sets. We found that transformer BERT-based models working on sentences obtained relatively good results ($\approx 89\%$ F1 score on base dataset). Even though better results were achieved using linguistic features ($\approx 91\%$ F1 score on base dataset), this model requires more human labor (humans in the loop) because features were prepared manually by expert linguists. BERT-based models consuming only the input sentences show that they capture most of the complexity of NLI/factivity. Complex cases in the phenomenon—for example, cases with entitlement (E) and non-factive verbs—still remain an open issue for further research.

Taking as a starting point the variation in introspective judgments on embedded gapping in English in the literature, the main goal of this paper is to test the ‘No Embedding Constraint’ experimentally. Building on a first experimental study designed to measure the interaction between that-omission and factivity in English embedded complement clauses, we conducted two experiments testing the role of the complementizer in embedded gapping, paying special attention to the semantic nature of the matrix predicates (non-factives vs semi-factives vs true factives). Our results show, on the one hand, that the ‘No Embedding Constraint’ makes too strong claims that are not backed up by our experimental findings, and, on the other hand, that embedded gapping is affected by both the presence/absence of that and by the semantic class of the matrix predicate in English. In particular, embedded gapping seems to be more acceptable under non-factive verbs, especially in the absence of a complementizer. Both constraints (that-omission and factivity) can be accounted for by a constructionist fragment-based analysis, where the gapped clause is a non-finite phrase that has to address the same Question Under Discussion as its source. This explains, in turn, why embedded gapping under true factive predicates is considered significantly less acceptable. We show that the acceptable cases of embedded gapping involve true syntactic embedding (so, the matrix clause has no parenthetical use). We conclude that English has the same sensitivity to the semantic class of the matrix predicate as other languages, but that the requirements on the presence/absence of that are English specific.

Research on the capacity to understand others' minds has tended to focus on representations of beliefs, which are widely taken to be among the most central and basic theory of mind representations. Representations of knowledge, by contrast, have received comparatively little attention and have often been understood as depending on prior representations of belief. After all, how could one represent someone as knowing something if one does not even represent them as believing it? Drawing on a wide range of methods across cognitive science, we ask whether belief or knowledge is the more basic kind of representation. The evidence indicates that nonhuman primates attribute knowledge but not belief, that knowledge representations arise earlier in human development than belief representations, that the capacity to represent knowledge may remain intact in patient populations even when belief representation is disrupted, that knowledge (but not belief) attributions are likely automatic, and that explicit knowledge attributions are made more quickly than equivalent belief attributions. Critically, the theory of mind representations uncovered by these various methods exhibits a set of signature features clearly indicative of knowledge: they are not modality-specific, they are factive, they are not just true belief, and they allow for representations of egocentric ignorance. We argue that these signature features elucidate the primary function of knowledge representation: facilitating learning from others about the external world. This suggests a new way of understanding theory of mind – one that is focused on understanding others' minds in relation to the actual world, rather than independent from it.