This paper tackles the goal of conclusion-supplement answer generation for non-factoid questions, which is a critical issue in the field of Natural Language Processing (NLP) and Artificial Intelligence (AI), as users often require supplementary information before accepting a conclusion. The current encoder-decoder framework, however, has difficulty generating such answers, since it may become confused when it tries to learn several different long answers to the same non-factoid question. Our solution, called an ensemble network, goes beyond single short sentences and fuses logically connected conclusion statements and supplementary statements. It extracts the context from the conclusion decoder’s output sequence and uses it to create supplementary decoder states on the basis of an attention mechanism. It also assesses the closeness of the question encoder’s output sequence and the separate outputs of the conclusion and supplement decoders as well as their combination. As a result, it generates answers that match the questions and have natural-sounding supplementary sequences in line with the context expressed by the conclusion sequence. Evaluations conducted on datasets including “Love Advice” and “Arts & Humanities” categories indicate that our model outputs much more accurate results than the tested baseline models do.

Generalized Mutual Assignment Problem (GMAP) is a multi-agent based distributed combinatorial optimization where the agents try to obtain the most profitable job assignment. Since it is NP-hard problem, it is challenging to achieve feasible solutions of GMAP. Existing algorithms to solve GMAP are synchronous ones, that is, the performance of the entire system would deteriorate if a certain agent takes a long time to solve her own subproblem. Furthermore, topology of communication networks strictly depends on the structure of a given instance due to the way of decomposing the problem into subproblem. In this paper, we propose a novel distributed asynchronous heuristic algorithm based on the Lagrangian decomposition formulation in order to obtain feasible solutions as good as possible. Our proposed algorithm consists of a couple of parts. One is to check the feasibility of candidate feasible solutions and the other is to solve the Lagrangian dual problem to generate a variety of candidates. Both of them are based on asynchronous gossip algorithms which are sometimes introduced for modeling rumor spreading phenomena or calculating an average value of sensors, where only two agents communicate with each other at one iteration. Our experiments show the effectiveness of the proposed method.

In the teaching and learning process, it is important to understand not only the state of understanding but also the mental state of the learner. However, the mental state of the learner is not always expressed in facial expressions or movements, and it is sometimes difficult to observe or measure from the outside. Therefore, previous studies have tried to estimate the mental state by focusing on biometric information. However, previous studies have ignored the time-series nature of the data, resulting in a model that is difficult to generalize. In addition, the labeling cost of subjects in the previous studies was large, making the model unrealistic from the perspective of widespread use of educational systems. In this study, we experimentally investigated the generalizability of the estimation and the reduction of the labeling cost using deep learning. As a result, we found that the emotion decay model proposed in this study and its response to a small number of samples contributed the most to generalizability. We also confirmed that the combination of these findings with domain adaptation could reduce the labeling cost by up to 80%.

In recent years, the importance of responding to climate change has increased, and various countries have set greenhouse gas emission reduction targets centered on CO₂ (Carbon Dioxide). In this study, we analyzed the relationship between macroeconomic activity and CO₂ emissions amid growing discussions on CO₂ emissions. Recently, while the positive correlation between GDP and CO₂ emissions on a level basis, which had been confirmed in the past, is collapsing mainly in Europe and the United States. However, it this research, we converted the GDP and CO₂ emissions data to a yearly change rate in order to mitigate the effects of structural environmental changes and we confirmed that a positive correlation was maintained between emissions and economic growth rate. Then, in this research, in order to immediately grasp the CO₂ emission status, we used the CO₂ concentration data which was measured using the observation information from artificial satellites (GOSAT, Greenhouse gases Observing SATellite). We proposed an effective method for grasping the macroeconomic situation in real time. In addition, in the proposed method, the estimation accuracy of the economic model was improved by using fine-grained satellite information as a feature quantity.

Visualizing social relationships by a network is useful for understanding the behavior of groups and individuals. The target of this study is a network between employees in the workplace. The construction of this network enables us to understand human relationships and managing a team. To build this network, the questionnaire and E-mail data were conventionally used. However, in this work, we use conversation history data on a chat application(Slack, etc.). We propose a method of quantifying the relationship between employees from conversation data on a chat application and visualizing it as a network between employees. Specifically, we assume that strongly related employees will make remarks at adjacent times on the chat, quantify the relationship by multivariate Hawkes process and build a network. To verify the effectiveness of the proposed model, we used Slack conversation data of a real company and extracted knowledge about team management from the network.

Many learned inference engines have been released as cloud-based AI services. However, learned AI services are black boxes, and it is difficult for users to decide which service to choose. We propose a comparison service to infer the best AI service by learning their different output results as training data. Our model, “AI for selecting the best AI,” involves meta-learning; it learns the output of the cloud-based AI service as metadata. We compared and evaluated the accuracy and cost of two proposed models that recommend the best among several commercial AI services and an ensemble method. The results of our experiments to infer face attributes (i.e., age and gender) on a face image dataset crawled from Wikipedia showed that the accuracy of our system was higher than that of single face classification cloud-based AI service. Notably, results on inferring age and gender, where training data for each service showed a significant difference in the tendency for accuracy, had 6.2% higher accuracy compared to existing cloud-based AIs.

In recent years, there has been a lot of research on building dialogue systems using deep learning, which can generate relatively fluent response sentences to user utterances. Nevertheless, they tend to produce responses that are not diverse and which are less context-dependent. Assuming that the problem is caused by the Softmax Cross- Entropy (SCE) loss, which treats all words equally without considering the imbalance in the training data, a loss function Inverse Token Frequency (ITF) loss, which multiplies the SCE loss by a weight based on the inverse of the token frequency, was proposed and confirmed the improvement of dialogue diversity. However, in the diversity of sentences, it is necessary to consider not only the information of independent tokens, but also the frequency of incorporating a sequence of tokens. Using frequencies that incorporate a sequence of tokens to compute weights that dynamically change depending on the context, we can better represent the diversity we seek. Therefore, we propose a loss function, Inverse N-gram Frequency (INF) loss, which is weighted based on the inverse of the n-gram frequency of the tokens instead of the frequency of the tokens. In order to confirm the effectiveness of the proposed method on INF loss, we conducted metric-based and human evaluations of sentences automatically generated by models trained on the Japanese and English Twitter datasets. In the metric-based evaluation, Perplexity, BLEU, DIST-N, ROUGE, and length were used as evaluation indices. In the human evaluation, we assessed the coherence and diversity of the response sentences. In the metric-based evaluation, the proposed INF model achieved higher scores in Perplexity, DIST-N, and ROUGE than the previous methods. In the human evaluation, the INF model also showed superior values.

Unsupervised image captioning is a task to describe images without the supervision of image–sentence pairs. With the support of pre-trained object detectors, previous work assigned pseudo-captions, i.e., sentences that contain the detected object labels, to a given image. They focused on aligning the pseudo-captions with input images at the sentence level. However, pseudo-captions contain many words that are irrelevant to a given image. To shed light on the problem of partial mismatches between images and pseudo-captions, we focus on removing mismatched words from image–sentence alignment. We propose a simple gating mechanism that is trained to align image features with only the most reliable words in pseudo-captions: the detected object labels. The superior performance of our method empirically demonstrates the importance of removing the partial mismatches. Detailed analysis elucidates that our method successfully improves its performance in predicting the words likely to be mismatched during training. Furthermore, we show that using our method as an initialization method significantly boosts the performance of the previous sentence-level alignment method. These results confirm the importance of careful alignment in word-level details.

Ability to understand surrounding environment based on its components, namely objects, is one of the most important cognitive ability for intelligent agents. Human beings are able to decompose sensory input, i.e. visual stimulation, into some components based on its meaning or relationships between entities, and are able to recognize those components as “object ”. It is often said that this kind of compositional recognition ability is essential for resolving so called Binding Problem, and thus important for many tasks such as planning, decision making and reasoning. Recently, researches about obtaining object level representation in unsupervised manner using deep generative models have been gaining much attention, and they are called ”Scene Interpretation models”. Scene Interpretation models are able to decompose input scenes into symbolic entities such as objects, and represent them in a compositional way. The objective of our research is to point out the weakness of existing scene interpretation methods and propose some methods to improve them. Scene Interpretation models are trained in fully-unsupervised manner in contrast to latest methods in computer vision which are based on massive labeled data. Due to this problem setting, scene interpretation models lack inductive biases to recognize objects. Therefore, the application of these models are restricted to relatively simple toy datasets. It is widely known that introducing inductive biases to machine learning models is sometimes very useful like convolutional neural networks, but how to introduce them via training depends on the models and is not always obvious. In this research, we propose to incorporate self-supervised learning to scene interpretation models for introducing additional inductive bias to the models, and we also propose a model architecture using Transformer which is considered to be suitable for scene interpretation when combined with self-supervised learning. We show proposed methods outperforms previous methods, and is able to adopt to Multi-MNIST dataset which previous methods could not deal with well.

Crowd congestion causes accidents and sometimes leads to hundreds of injuries and deaths. To mitigate crowd congestion, we could design a space where the pedestrians can move smoothly, or intervene and control the crowd movement. Recently, computer simulation is often used to determine optimal spatial design and crowd control. The simulation methods of crowd movement are well organized because it has been an active area of research since the 1990s to understand crowd dynamics. On the other hand, methods and knowledge of spatial design and crowd control are not well organized, because research topics have recently become more active, and the knowledge is distributed among artificial intelligence, information processing, traffic engineering, civil engineering, architecture, physics, and more. The purpose of this paper is to summarize the research trends in spatial design and crowd control from the viewpoint of crowd simulation and optimization, which are often discussed in the field of artificial intelligence. We first introduce the evaluation indices of crowd movement in terms of efficiency and safety. Then, we explain the current findings on spatial design and categorize the methods of crowd control, ordered by the level of pedestrian behavior. In the end, the challenges and approaches are discussed based on the previous research.