In this paper, we investigate the use of discourse-aware rewards with reinforcement learning to guide a model to generate long, coherent text. In particular, we propose to learn neural rewards to model cross-sentence ordering as a means to approximate desired discourse structure. Empirical results demonstrate that a generator trained with the learned reward produces more coherent and less repetitive text than models trained with cross-entropy or with reinforcement learning with commonly used scores as rewards.
Since large knowledge bases are typically incomplete, missing facts need to be inferred from observed facts in a task called knowledge base completion. The most successful approaches to this task have typically explored explicit paths through sequences of triples. These approaches have usually resorted to human-designed sampling procedures, since large knowledge graphs produce prohibitively large numbers of possible paths, most of which are uninformative. As an alternative approach, we propose performing a single, short sequence of interactive lookup operations on an embedded knowledge graph which has been trained through end-to-end backpropagation to be an optimized and compressed version of the initial knowledge base. Our proposed model, called Embedded Knowledge Graph Network (EKGN), achieves new state-of-the-art results on popular knowledge base completion benchmarks.
We develop a technique for transfer learning in machine comprehension (MC) using a novel two-stage synthesis network (SynNet). Given a high-performing MC model in one domain, our technique aims to answer questions about documents in another domain, where we use no labeled data of question-answer pairs. Using the proposed SynNet with a pretrained model from the SQuAD dataset on the challenging NewsQA dataset, we achieve an F1 measure of 44.3% with a single model and 46.6% with an ensemble, approaching performance of in-domain models (F1 measure of 50.0%) and outperforming the out-of-domain baseline of 7.6%, without use of provided annotations.
Teaching a computer to read and answer general questions pertaining to a document is a challenging yet unsolved problem. In this paper, we describe a novel neural network architecture called the Reasoning Network (ReasoNet) for machine comprehension tasks. ReasoNets make use of multiple turns to effectively exploit and then reason over the relation among queries, documents, and answers. Different from previous approaches using a fixed number of turns during inference, ReasoNets introduce a termination state to relax this constraint on the reasoning depth. With the use of reinforcement learning, ReasoNets can dynamically determine whether to continue the comprehension process after digesting intermediate results, or to terminate reading when it concludes that existing information is adequate to produce an answer. ReasoNets have achieved exceptional performance in machine comprehension datasets, including unstructured CNN and Daily Mail datasets, the Stanford SQuAD dataset, and a structured Graph Reachability dataset.
Unsupervised learning is the most challenging problem in machine learning and especially in deep learning. Among many scenarios, we study an unsupervised learning problem of high economic value --- learning to predict without costly pairing of input data and corresponding labels. Part of the difficulty in this problem is a lack of solid evaluation measures. In this paper, we take a practical approach to grounding unsupervised learning by using the same success criterion as for supervised learning in prediction tasks but we do not require the presence of paired input-output training data. In particular, we propose an objective function that aims to make the predicted outputs fit well the structure of the output while preserving the correlation between the input and the predicted output. We experiment with a synthetic structural prediction problem and show that even with simple linear classifiers, the objective function is already highly non-convex. We further demonstrate the nature of this non-convex optimization problem as well as potential solutions. In particular, we show that with regularization via a generative model, learning with the proposed unsupervised objective function converges to an optimal solution.
Monaural source separation is important for many real world applications. It is challenging because, with only a single channel of information available, without any constraints, an infinite number of solutions are possible. In this paper, we explore joint optimization of masking functions and deep recurrent neural networks for monaural source separation tasks, including monaural speech separation, monaural singing voice separation, and speech denoising. The joint optimization of the deep recurrent neural networks with an extra masking layer enforces a reconstruction constraint. Moreover, we explore a discriminative criterion for training neural networks to further enhance the separation performance. We evaluate the proposed system on the TSP, MIR-1K, and TIMIT datasets for speech separation, singing voice separation, and speech denoising tasks, respectively. Our approaches achieve 2.30--4.98 dB SDR gain compared to NMF models in the speech separation task, 2.30--2.48 dB GNSDR gain and 4.32--5.42 dB GSIR gain compared to existing models in the singing voice separation task, and outperform NMF and DNN baselines in the speech denoising task.