Jakob Uszkoreit

Lattice-based Minimum Error Rate Training for Statistical Machine Translation

Minimum Error Rate Training (MERT) is an effective means to estimate the feature function weights of a linear model such that an automated evaluation criterion for measuring system performance can directly be optimized in training. To accomplish this, the training procedure determines for each feature function its exact error surface on a given set of candidate translations. The feature function weights are then adjusted by traversing the error surface combined over all sentences and picking those values for which the resulting error count reaches a minimum. Typically, candidates in MERT are represented as N-best lists which contain the N most probable translation hypotheses produced by a decoder. In this paper, we present a novel algorithm that allows for efficiently constructing and representing the exact error surface of all translations that are encoded in a phrase lattice. Compared to N-best MERT, the number of candidate translations thus taken into account increases by several orders of magnitudes. The proposed method is used to train the feature function weights of a phrase-based statistical machine translation system. Experiments conducted on the NIST 2008 translation tasks show significant runtime improvements and moderate BLEU score gains over N-best MERT.

A Decomposable Attention Model for Natural Language Inference

We propose a simple neural architecture for natural language inference. Our approach uses attention to decompose the problem into subproblems that can be solved separately, thus making it trivially parallelizable. On the Stanford Natural Language Inference (SNLI) dataset, we obtain state-of-the-art results with almost an order of magnitude fewer parameters than previous work and without relying on any word-order information. Adding intra-sentence attention that takes a minimum amount of order into account yields further improvements.

Coarse-to-Fine Question Answering for Long Documents

We present a framework for question answering that can efficiently scale to longer documents while maintaining or even improving performance of state-of-the-art models. While most successful approaches for reading comprehension rely on recurrent neural networks (RNNs), running them over long documents is prohibitively slow because it is difficult to parallelize over sequences. Inspired by how people first skim the document, identify relevant parts, and carefully read these parts to produce an answer, we combine a coarse, fast model for selecting relevant sentences and a more expensive RNN for producing the answer from those sentences. We treat sentence selection as a latent variable trained jointly from the answer only using reinforcement learning. Experiments demonstrate state-of-the-art performance on a challenging subset of the WikiReading dataset and on a new dataset, while speeding up the model by 3.5x-6.7x.