論文

In this paper, we propose a non-task-oriented dialogue system controlling the utterance length. The dialogue system can be classified into a task-oriented dialogue system or a non-task-oriented dialogue system. Recently, demand for the non-task-oriented dialogue system is increasing. The utterance length is one of the important information in a dialogue system. In general, our utterance length tends to be long when we are speakers. On the other hand, the length of our utterance tends to be short when we are listeners. In addition, the utterance length differs from person to person, so we change our utterance length for friendly communication. The effect of the utterance length has never considered in dialogue systems using Encoder-Decoder model. Therefore, we propose an utterance length estimator (ULE) and an index of the utterance length. ULE is a neural network which learns the utterance length by training data of dialogue. The index of the utterance length is the parameter considers user's personality and it is calculated during dialogue. Our dialogue system decides the length of system's utterance by ULE and index of the utterance length, and generates output sequences by using a neural encoder-decoder controlling output length. Experimental results show our system can decide the appropriate length of the utterance and makes users more satisfied than the conventional method.

In this paper, we propose an automatic template features extraction method and apply it to utterance generation in a dialogue system. Template-based utterance generation has been widely used in many dialogue systems because of its robustness. Although variety of templates and the appropriate selection are crucial points in the method, they have not been paid attention so far. This paper focuses on the points; first, we propose the new neural network model utilizingLSTM (Long Short-Term Memory) to extract effective and unique features for templates, and then applied it to utterance generation in a dialogue system. To examine the effectiveness of the proposed method, we conduct two kinds of experiments; subjective evaluation and dialogue breakdown detection experiment. In both of the experiments, the proposed method has shown higher accuracy than the conventional methods.

Although word embeddings are powerful, weakness on rare words, unknown words and issues of large vocabulary motivated people to explore alternative representations. While the character embeddings have been successful for alphabetical languages, Japanese is difficult to be processed at the character level as well because of the large vocabulary of kanji, written in the Chinese characters. In order to achieve fewer parameters and better generalization on infrequent words and characters, we proposed a model that encodes Japanese texts from the radical-level representation, inspired by the experimental findings in the field of psycholinguistics. The proposed model is comprised of a convolutional local encoder and a recurrent global encoder. For the convolutional encoder, we propose a novel combination of two kinds of convolutional filters of different strides in one layer to extract information from the different levels. We compare the proposed radical-level model with the state-of-the-art word and character embedding-based models in the sentiment classification task. The proposed model outperformed the state-of-the-art models for the randomly sampled texts and the texts that contain unknown characters, with 91% and 12% fewer parameters than the word embedding-based and character embedding-based models, respectively. Especially for the test sets of unknown characters, the results by the proposed model were 4.01% and 2.38% above the word embedding-based and character embedding-based baselines, respectively. The proposed model is powerful with cheaper computational and storage cost, can be used for devices with limited storage and to process texts of rare characters.

In this paper, we propose an attribute added face image generation system using Deep Convolutional Generative Adversarial Networks(DCGANs). Convolutional Neural Networks(CNNs) can extract important features of an image and attain high precision in image classification tasks. In the proposed system, image features are extracted using CNNs, and attribute features added to image features, and attributes added images are generated by DCGANs. Specifically, we use the attributes of "smile" and "male", and work on a task of generating smile images from non-smile images, and a task of generating male images from female images. Since the training of the proposed system requires image pairs including with and without attributes, we use two extraction methods, 1)Usage of attribute label attached dataset, 2)Usage of cosine similarity. To obtain attribute features, we trained 4-layer CNNs which are the same architecture as Discriminator of GANs, to classify images into two classes, with and without attributes. Here, attribute features are defined as the averaged difference between image features with and without attributes, more specifically, the values in the final convolution layer in the 4-layer CNNs. We performed two kinds of evaluation experiments: the first one is a subjective evaluation experiment on items such as "whether generated images have attributes", the second one is a quantitative evaluation experiment for measuring whether the people shown in the input image and the generated image are the same person. As the results, excellent characteristics were obtained.

In this paper, we propose an improved Hierarchical Temporal Memory (HTM) that can consider long-term dependence. HTM is a temporal sequence prediction model imitating the cerebral cortex structure and learning algorithm. This model is composed of cells of a two-dimensional map representing neurons of the brain, and expresses data by a set of cells in an activated state. Further, the data of the next time is predicted from the set of the cells in the predicted state. HTM learns the time series data by updating synapses connecting each cell according to Hebb's rule and keeps the proper relationship of data. In the conventional model, only the connection with the previous data is learned, but in the proposed model the connection with several former data can be learned. The proposed HTM is modified in terms of structure and learning algorithm. In the structure, we introduced a time axis for the segment which is a collection of synapses. About learning algorithm, the connection with several times ago leads the predicted state. As a result of evaluation experiments, it was confirmed that the proposed model can consider longer-term dependency than the conventional model on temporal sequence prediction.

This paper proposes a quality recovery network (QRNet) that recovers the image quality from distorted images and improves the classification accuracy for image classification using these recovered images as the classifier inputs, which are optimized for image quality loss and classification loss. In certain image classification tasks, classifiers based on deep neural networks achieve higher performance compared to those realized by humans. However, these tasks are based on images that are not distorted. To address distorted images, the classifier is fine-tuned with distorted images for practical applications. However, fine-tuning is insufficient for classifying images that include multiple distortion types with severe distortions and often requires the classifier to be retrained for adapting to distorted images, which is a time-consuming process. Therefore, we propose QRNet that generates recovered images for input to the classifier. To address multiple severe distortions, the proposed network is trained using multiple distortion-type images with our proposed loss, which comprises the image quality and classification losses. Moreover, by training the proposed network with multiple classifiers, the recovered images can be easily classified by a new classifier that is not used for training. The new classifier can classify the recovered images without retraining for adapting to distorted images. We evaluate our proposed network with classifiers on public datasets and demonstrate that it improves the classification accuracy for distorted images. Moreover, the experimental results demonstrate that our proposed network with the new classifier improves the classification accuracy.

The authors propose methods to learn symbolic processing with deep learning and to build question answering systems by means of learned models. Symbolic processing, performed by the Prolog processing systems which execute unification, resolution, and list operations, is learned by a combination of deep learning models, Neural Machine Translation (NMT) and Word2Vec training. To our knowledge, the implementation of a Prolog-like processing system using deep learning is a new experiment that has not been conducted in the past. The results of their experiments revealed that the proposed methods are superior to the conventional methods because symbolic processing (1) has rich representations, (2) can interpret inputs even if they include unknown symbols, and (3) can be learned with a small amount of training data. In particular (2), handling of unknown data, which is a major task in artificial intelligence research, is solved using Word2Vec. Furthermore, question answering systems can be built from knowledge bases written in Prolog with learned symbolic processing, which, with conventional methods, is extremely difficult to accomplish. Their proposed systems can not only answer questions through powerful inferences by utilizing facts that harbor unknown data but also have the potential to build knowledge bases from a large amount of data, including unknown data, on the Web. The proposed systems are a completely new trial, there is no state-of-the-art methods in the sense of 'newest'. Therefore, to evaluate their efficiency, they are compared with the most traditional and robust system i.e., the Prolog system. This is new research that encompasses the subjects of conventional artificial intelligence and neural network, and their systems have higher potential to build applications such as FAQ chatbots, decision support systems and energy-efficient estimation using a large amount of information on the Web. Mining hidden information through these applications will provide great value.

In this paper, we introduce a novel layer designed to be used as the output of pre-trained neural networks in the context of classification. Based on Associative Memories, this layer can help design deep neural networks which support incremental learning and that can be (partially) trained in real time on embedded devices. Experiments on the ImageNet dataset and other different domain specific datasets show that it is possible to design more flexible and faster-to-train Neural Networks at the cost of a slight decrease in accuracy.

Most characters in the Chinese and Japanese languages are ideographic compound characters composed of subcharacter elements arranged in a planar manner. Token-based and image-based subcharacter-level models have been proposed to leverage the subcharacter elements. However, on the one hand, the conventional token-based subcharacter-level models are blind to the planar structural information; on the other hand, the image-based models are weak with respect to ideographic characters that share similar shapes but have different meanings. These characteristics motivate us to explore non-image-based methods to encode the planar structural information of characters. In this paper, we propose and discuss a method to encode the planar structural information by learning embeddings of the categories of the structure types. Our proposed model adds the structure embeddings to the conventional subcharacter embeddings and position embeddings before they are input into the encoder. In this way, the model learns the planar structural and positional information and retains the uniqueness of each character. We evaluated the method in a text classification task. In the experiment, the embeddings were encoded by a CNN encoder, and then the encoded vectors were input into an LSTM classifier to classify product reviews as positive or negative. We compared the proposed model with models that use only the subcharacter embeddings, the structure embedding or the position embeddings as well as with the conventional models in previous works. The results show that adding both structure embeddings and position embeddings leads to more rich and representative features and better fitting on the dataset. The proposed method results in at best 1.8% better recall, 0.63% better F-score, and 0.55% better accuracy on the testing datasets compared to previous methods.

In this paper, we propose an analog value associative memory using Restricted Boltzmann Machine (AVAM). Research on treating knowledge is becoming more and more important such as in natural language processing and computer vision fields. Associative memory plays an important role to store knowledge. First, we obtain distributed representation of words with analog values using word2vec. Then the obtained distributed representation is learned in the proposed AVAM. In the evaluation experiments, we found simple but very important phenomenon in word2vec method: almost all of the values in the generated vectors are small values. By applying traditional normalization method for each word vector, the performance of the proposed AVAM is largely improved. Detailed experimental evaluations are carried out to show superior performance of the proposed AVAM.

The authors propose analogical reasoning systems based on first-order predicate logic using deep learning. The proposed systems consist of a model combining recursive neural networks and Word2Vec. When unknown data is input in this trained model, similar outputs to those of the trained data are obtained. Previous studies on symbolic inference using deep learning have focused on deductive and inductive inferences, and they require rule templates that are manually created and provided to the models in advance. However, it is not enough to infer unknown rules from a large amount of data on the Internet. Analogical reasoning is a mode of inference that attempts to solve a new case based on similarity. Thus, it enables inference of rules that were challenging to comprehend in previous studies. As a result of the experiments, unknown rules can be efficiently inferred from known rules contained in the knowledge bases. Using the proposed systems, the authors can extract unknown rules from five of the eight datasets consisting of three types of knowledge bases described in Prolog. The proposed method is the first case of analogical reasoning based on the first-order predicate logic using deep learning. Furthermore, the models used in the proposed systems have shown high robustness when using Word2Vec. For this reason, it is suggested that the practical applications of the proposed systems enable efficient inferences from a large amount of data that include noisy data on the Internet.

Current image quality assessment (IQA) methods require the original images for evaluation. However, recently, IQA methods that use machine learning have been proposed. These methods learn the relationship between the distorted image and the image quality automatically. In this paper, we propose an IQA method based on deep learning that does not require a reference image. We show that a convolutional neural network with distortion prediction and fixed filters improves the IQA accuracy.