Search for: All records

Different entities with the same name can be difficult to distinguish. Handling confusing entity mentions is a crucial skill for language models (LMs). For example, given the question “Where was Michael Jordan educated?” and a set of documents discussing different people named Michael Jordan, can LMs distinguish entity mentions to generate a cohesive answer to the question? To test this ability, we introduce a new benchmark, AmbigDocs. By leveraging Wikipedia’s disambiguation pages, we identify a set of documents, belonging to different entities who share an ambiguous name. From these documents, we generate questions containing an ambiguous name and their corresponding sets of answers. Our analysis reveals that current state-of-the-art models often yield ambiguous answers or incorrectly merge information belonging to different entities. We establish an ontology categorizing four types of incomplete answers and automatic evaluation metrics to identify such categories. We lay the foundation for future work on reasoning across multiple documents with ambiguous entities. 

Abstract Due to the inaccessibility of Earth's deep interior, geologists have long attempted to estimate the composition of the continental crust from its seismic properties. Despite numerous sources of error including nonuniqueness in the mapping between composition and seismic properties, the corresponding uncertainties have typically been estimated qualitatively at best. We propose a Bayesian approach that uses mineralogical modeling to combine prior knowledge about the composition of the crust with seismic data to give a posterior distribution of the predicted composition at any location, combined with a Monte Carlo simulation to estimate the average composition of the Earth's crust. Our approach yields an estimated composition of 59.5% silica in the upper crust (90% credible interval 58.9 %–60.1%), 57.9% in the middle crust (90% credible interval 57.2%–58.6%), and 53.6% in the lower crust (90% credible interval 53.0%–54.2%). Our estimate exhibits less compositional stratification over depth and a more intermediate composition in the upper and middle crust than previous estimates. Testing our approach on a simulated crust reveals the importance of prior assumptions in estimating the composition of the crust from its seismic properties, and suggests that future work should focus on quantifying those assumptions. 

Geophysical turbulence has a wide range of spatiotemporal scales that requires a multiscale prediction model for efficient and fast simulations. Stochastic parameterization is a class of multiscale methods that approximates the large-scale behaviors of the turbulent system without relying on scale separation. In the stochastic parameterization of unresolved subgrid-scale dynamics, there are several modeling parameters to be determined by tuning or fitting to data. We propose a strategy to estimate the modeling parameters in the stochastic parameterization of geostrophic turbulent systems. The main idea of the proposed approach is to generate data in a spatiotemporally local domain and use physical/statistical information to estimate the modeling parameters. In particular, we focus on the estimation of modeling parameters in the stochastic superparameterization, a variant of the stochastic parameterization framework, for an idealized model of synoptic-scale turbulence in the atmosphere and oceans. The test regimes considered in this study include strong and moderate turbulence with complicated patterns of waves, jets, and vortices.