Rachel Rudinger

Script Induction as Language Modeling

The narrative cloze is an evaluation metric commonly used for work on automatic script induction. While prior work in this area has focused on count-based methods from distributional semantics, such as pointwise mutual information, we argue that the narrative cloze can be productively reframed as a language modeling task. By training a discriminative language model for this task, we attain improvements of up to 27 percent over prior methods on standard narrative cloze metrics.

Universal Decompositional Semantics on Universal Dependencies.

We present a framework for augmenting data sets from the Universal Dependencies project with Universal Decompositional Semantics. Where the Universal Dependencies project aims to provide a syntactic annotation standard that can be used consistently across many languages as well as a collection of corpora that use that standard, our extension has similar aims for semantic annotation. We describe results from annotating the English Universal Dependencies treebank, dealing with word senses, semantic roles, and event properties.

Is the Stanford Dependency Representation Semantic

The Stanford Dependencies are a deep syntactic representation that are widely used for semantic tasks, like Recognizing Textual Entailment. But do they capture all of the semantic information a meaning representation ought to convey? This paper explores this question by investigating the feasibility of mapping Stanford dependency parses to Hobbsian Logical Form, a practical, event-theoretic semantic representation, using only a set of deterministic rules. Although we find that such a mapping is possible in a large number of cases, we also find cases for which such a mapping seems to require information beyond what the Stanford Dependencies encode. These cases shed light on the kinds of semantic information that are and are not present in the Stanford Dependencies.

Social Bias in Elicited Natural Language Inferences.

We analyze the Stanford Natural Language Inference (SNLI) corpus in an investigation of bias and stereotyping in NLP data. The SNLI human-elicitation protocol makes it prone to amplifying bias and stereotypical associations, which we demonstrate statistically (using pointwise mutual information) and with qualitative examples.

Learning to predict script events from domain-specific text

The automatic induction of scripts (Schank and Abelson, 1977) has been the focus of many recent works. In this paper, we employ a variety of these methods to learn Schank and Abelson’s canonical restaurant script, using a novel dataset of restaurant narratives we have compiled from a website called “Dinners from Hell.” Our models learn narrative chains, script-like structures that we evaluate with the “narrative cloze” task (Chambers and Jurafsky, 2008).