Language and Natural Reasoning
What we do
Knowing what a text means involves drawing inferences based on the information in the text. Our group works on inferential properties of linguistic expressions to enable automated reasoning for NL understanding.
We want to contribute to the theoretical understanding of how language and reasoning interact and to the computational modeling of such interactions. Currently we concentrate on
- the linguistic encoding of temporal and spatial information,
- the linguistic encoding of modality and veridicity,
- local textual inferences,
- natural logic,
- deriving logical forms that allow interaction with structured information and computational reasoners,
- developing data sets for training neural nets to recognize whether a sentence A entails or contradicts sentence B, or whether the two sentences neither entail or contradict each other.
Members
Members in Residence
Johan van Benthem, Cleo Condoravdi, Ignacio Cases, James Collins, Thomas F. Icard III, Lauri Karttunen, Dan Lassiter, Stanley Peters, Annie Zaenen
Off-Campus Members
David Beaver, Graham Katz, Sven Lauer, Shalom Lappin, Larry Moss, Valeria de Paiva, Kyle Richardson, Richard Waldinger
Completed Sponsored Projects
Some of the lexical resources we have been working on for the FAUST project can be browsed by this link. We welcome comments and suggestions. Please let us know if you find the data useful for some project of yours and give us credit if you use the data.
Current Projects
Veridicity
Natural language provides speakers/authors with a variety of ways to signal to hearers/readers what their stance is on the factuality of events or the existence of the entities mentioned in the discourse. Linguists have studied these under the heading of implicatives, factives and epistemic modal expressions. The advent of crowd sourcing techniques allows us to refine and complement these studies. At this point we are looking at the following sub-issues:
Semantics and pragmatics of lucky
The construction be lucky to VP is ambiguous. Some people will be lucky to survive entails its complement clause: some people will survive. But Sam will be lucky to survive can be understood differently: probably Sam will not survive. The idiomatic ‘probably not’ interpretation of lucky is subject to a complex set of conditions. We are running experiments with Amazon's Mechanical Turk to systematically investigate all the factors that are involved.
Factive adjectives
While factive verbs have been studied extensively, there are few studies of factive adjectives. We are currently examining the various constructions in which factive adjectives can be found (e.g. It be ADJ that S, NP be ADJ that S, NP be ADJ to VP, It be ADJ to VP). Corpus and experimental evidence suggests that there is quite a bit of variation in the factivity status of these adjectives.
Natural language inference is core to natural language understanding. Despite recent advances, recognizing the entailment relation between two sentences is still a challenging task for machines to perform. Recently, deep learning approaches have yielded impressive results with models that are trained with simple end-to-end training data.
The goal of this project is to train a neural net to recognize that entailment is a transitive relation and contradiction is symmetric. Whenever A entails B and B entails C, then A entails C as well. This project builds on the successful 2016 project on creating a learning corpus for Implicative Constructions. We used the corpus to train an end-to-end model to classify pairs of sentences with three labels: entails, contradicts, and permits (= neither entails nor contradicts).
Workshop
Workshop on Modality on April, 12 2013
Click here for more information.
Lexical Resources
Here is a link to a collection of lexical resources: lists of adjectives, verbs, and verb-noun collocations with their semantic signatures (factives, counterfactives, and six types of implicatives).
Last modified Thursday, 18-May-2017 16:29:45 PDT
