論文

In this paper, we propose a new personalization method for a dialogue system using external knowledge and Graph Convolutional Networks (GCNs). The proposed system is divided into a personalize-part and a response-generation part with Japanese Transformer model. The personalize-part consists of a learned GCN, which can estimate the user’s interests, and a knowledge graph as external knowledge. The personalize-part can add an appropriate template to the response by Transformer according to the dialogue situation and user’s interests. Estimating a topic interesting for each user, the new system can provide an intimate dialogue reflecting individualized interests. We carried out extensive evaluation experiments, and it has shown that the proposed system exceeds the conventional Transformer dialogue model by 35.6% in the variety of information and by 34.0% in the level of desire to use the system.

In this paper, we propose an interactive support system for generating picture books. The proposed system consists of two parts: a sentence generation part and a picture book image generation part. The sentence generation part takes a topic and an opening sentence of a picture book as inputs, and generates sentences of the picture book in an interactive manner. The user can select the most ap-propriate sentence from the generated sentences to create a story for the picture book, and can modify specific parts of the sentences. The image generation part generates a background image of the picture book using the text as an input, and acquires the necessary image objects for the picture book from the web. Using the retrieved images, the user can paste the image objects into the background image to create a picture book image. With these two parts, a picture book can be created without the user having to directly write the text and pictures. Through evaluation experiments, we confirmed that the proposed system can generate a picture book that is comfortable as a story and that can arouse the reader's interest.

In this paper, we propose a CVAM (continuous-valued associative memory for one-to-many associations) with back-propagation learning and analyze the performance in detail. Conventional associative memories often deal with binary patterns, however, most of the data handled today are continuous-valued data. The basic architecture of the proposed CVAM is a three-layer perceptron with multiple sub-layers in the hidden layer. The multiple sub-layers enable one-to-many associations using back-propagation (BP) learning algorithm; each sub-layer memorizes single one-to-one association and the multiple sub-layers enables one-to-many associations. We carried out experiments to analyze the important properties such as memory capacity and noise tolerance performance using continuous-valued data. In addition, we conducted a demonstrative experiment to visually confirm the behavior of the proposed CVAM as an associative memory model using the CIFAR-10 image data set.

Knowledge graphs (KGs) are useful information sources to make machine learning efficient with human knowledge. Since KGs are often incomplete, KG completion has become an important problem to complete missing facts in KGs. Whereas most of the KG completion methods are conducted on a single KG, multiple KGs can be effective to enrich embedding space for KG completion. However, most of the recent studies have concentrated on entity alignment prediction and ignored KG-invariant semantics in multiple KGs that can improve the completion performance. In this paper, we propose a new multiple KG embedding method composed of intra-KG and inter-KG regularization to introduce KG-invariant semantics into KG embedding space using aligned entities between related KGs. The intra-KG regularization adjusts local distance between aligned and not-aligned entities using contrastive loss, while the inter-KG regularization globally correlates aligned entity embeddings between KGs using multi-view loss. Our experimental results demonstrate that our proposed method combining both regularization terms largely outperforms existing baselines in the KG completion task.