No abstract

We introduce jiant, an open source toolkit for conducting multitask and transfer learning experiments on English NLU tasks. jiant enables modular and configuration-driven experimentation with state-of-the-art models and implements a broad set of tasks for probing, transfer learning, and multitask training experiments. jiant implements over 50 NLU tasks, including all GLUE and SuperGLUE benchmark tasks. We demonstrate that jiant reproduces published performance on a variety of tasks and models, including BERT and RoBERTa. jiant is available at https:// jiant.info. * Equal contribution. 1 The name jiant stands for "jiant is an NLP toolkit".

In this paper we propose a neural network model with a novel Sequential Attention layer that extends soft attention by assigning weights to words in an input sequence in a way that takes into account not just how well that word matches a query, but how well surrounding words match. We evaluate this approach on the task of reading comprehension (on the Who did What and CNN datasets) and show that it dramatically improves a strong baseline-the Stanford Reader-and is competitive with the state of the art.

The international investment working paper seriesincluding policies and trends and the broader implications of multinational enterpriseis designed to make available to a wide readership selected studies by the OECD Investment Committee, OECD Investment Division staff, or by outside consultants working on OECD Investment Committee projects. The papers are generally available only in their original language English or French with a summary in the other if available. The opinions expressed in these papers are the sole responsibility of the author(s) and do not necessarily reflect those of the OECD or the governments of its member countries.

Using data from 1946-2014, we show that audio features of lawyers' introductory statements and lawyers' facial attributes improve the performance of the best prediction models of Supreme Court outcomes. We infer face attributes using the MIT-CBCL human-labeled face database and infer voice attributes using a 15-year sample of human-labeled Supreme Court advocate voices. We find that image features improved prediction of case outcomes from 63.8% to 65.6%, audio features improved prediction of case outcomes from 66.8% to 68.8%, image and audio features together improved prediction of case outcomes from 66.9% to 67.7%, and the weights on lawyer traits are approximately half the weight of the most important feature from the models without image or audio features. Predictions of Justice votes with image and/or audio features however remained more similar relative to their baselines. We interpret this difference to suggest that human biases are more relevant in close cases.