This paper presents our study of cloze-style reading comprehension by imitating human reading comprehension, which normally involves tactical comparing and reasoning over candidates while choosing the best answer. We propose a multi-choice relational reasoning (McR2) model with an aim to enable relational reasoning on candidates based on fusion representations of document, query and candidates. For the fusion representations, we develop an efficient encoding architecture by integrating the schemes of bidirectional attention flow, self-attention and document-gated query reading. Then, comparing and inferring over candidates are executed by a novel relational reasoning network. We conduct extensive experiments on four datasets derived from two public corpora, Children’s Book Test and Who DiD What, to verify the validity and advantages of our model. The results show that it outperforms all baseline models significantly on the four benchmark datasets. The effectiveness of its key components is also validated by an ablation study.

In this paper, we study the problem of commonsense machine comprehension and propose a new model based on convolutional neural networks and Gated Tanh-ReLU Units. The new model, which serves as an alternative to exiting recurrent models, consists of three layers: input layer, gated convolutional layer, and output layer. The input layer produces representations based on various features, such as part-of-speech and relation embeddings. Gated convolutional layer, the key component of our model, extracts n-gram features at different granularities and models the interactions between different texts (questions, answers, and passages). Bilinear interactions are used as output layer to capture the relations among the final expressive representations and to produce the final answers. We evaluate our model on the SemEval-2018 Machine Comprehension Using Commonsense Knowledge task. Experimental result shows that our model achieves highly competitive results with the state-of-the-art models but is much faster. To our knowledge, this is the first time a non-recurrent approach gains competitive performance with strong recurrent models for commonsense machine comprehension.

Existing sentence alignment methods are founded fundamentally on sentence length and lexical correspondences. Methods based on the former follow in general the length proportionality assumption that the lengths of sentences in one language tend to be proportional to that of their translations, and are known to bear poor adaptivity to new languages and corpora. In this paper, we attempt to interpret this assumption from a new perspective via the notion of collaborative matching, based on the observation that sentences can work collaboratively during alignment rather than separately as in previous studies. Our approach is tended to be independent on any specific language and corpus, so that it can be adaptively applied to a variety of texts without binding to any prior knowledge about the texts. We use one-to-one sentence alignment to illustrate this approach and implement two specific alignment methods, which are evaluated on six bilingual corpora of different languages and domains. Experimental results confirm the effectiveness of this collaborative matching approach.