Search for: All records

Abstract We present an event structure classification empirically derived from inferential properties annotated on sentence- and document-level Universal Decompositional Semantics (UDS) graphs. We induce this classification jointly with semantic role, entity, and event-event relation classifications using a document-level generative model structured by these graphs. To support this induction, we augment existing annotations found in the UDS1.0 dataset, which covers the entirety of the English Web Treebank, with an array of inferential properties capturing fine-grained aspects of the temporal and aspectual structure of events. The resulting dataset (available at decomp.io) is the largest annotation of event structure and (partial) event coreference to date. 

We investigate which patterns of lexically triggered doxastic, bouletic, neg(ation)-raising, and veridicality inferences are (un)attested across clause-embedding verbs in English. To carry out this investigation, we use a multiview mixed effects mixture model to discover the inference patterns captured in three lexicon-scale inference judgment datasets: two existing datasets, MegaVeridicality and MegaNegRaising, which capture veridicality and neg-raising inferences across a wide swath of the English clause-embedding lexicon, and a new dataset, MegaIntensionality, which similarly captures doxastic and bouletic inferences. We focus in particular on inference patterns that are correlated with morphosyntactic distribution, as determined by how well those patterns predict the acceptability judgments in the MegaAcceptability dataset. We find that there are 15 such patterns attested. Similarities among these patterns suggest the possibility of underlying lexical semantic components that give rise to them. We use principal component analysis to discover these components and suggest generalizations that can be derived from them. 

There is growing evidence that the prevalence of disagreement in the raw annotations used to construct natural language inference datasets makes the common practice of aggregating those annotations to a single label problematic. We propose a generic method that allows one to skip the aggregation step and train on the raw annotations directly without subjecting the model to unwanted noise that can arise from annotator response biases. We demonstrate that this method, which generalizes the notion of a mixed effects model by incorporating annotator random effects into any existing neural model, improves performance over models that do not incorporate such effects. 

There is growing evidence that the prevalence of disagreement in the raw annotations used to construct natural language inference datasets makes the common practice of aggregating those annotations to a single label problematic. We propose a generic method that allows one to skip the aggregation step and train on the raw annotations directly without subjecting the model to unwanted noise that can arise from annotator response biases. We demonstrate that this method, which generalizes the notion of a mixed effects model by incorporating annotator random effects into any existing neural model, improves performance over models that do not incorporate such effects.