Proceedings of the Annual Conference of JSAI
Current issue
Showing 1-50 articles out of 892 articles from the selected issue
  • Koji MORIKAWA, Akisato KIMURA, Daisuke KATAGAMI
    Pages 1-6
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS
    Preface and committee members of the 34th annual conference
    Download PDF (234K)
  • Ryoma KUBO, Takuya FUJITA, Takehito UTSURO, Akio KOBAYASHI, Hiromitsu ...
    Session ID: 1B5-GS-6-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    This paper proposes how to mine concerns and reviews that are relevant to TV dramas from the results of collecting tweets on those TV dramas. The task studied in this paper is to identify tweets which include reviews and impressions of actors and characters of those TV dramas. In order to formalize this task, we apply the machine comprehension framework of BERT, where each tweet is considered as the context of machine comprehension and the name of an actor or a character is given to this tweet as a question and an adjective representing its reviews or impressions is returned as the answer to the question. Experimental evaluation results show that the proposed approach achieved over 70% recall and precision.

    Download PDF (593K)
  • SHIAU CHOUJEN, Ochi MASANAO, Nagahama KEN, Sakaki TAKESHI, Mori JUNICH ...
    Session ID: 1B5-GS-6-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In recent years, "viral spreading social issues" which are social issues that emerge in public consciousness through the fast spread of information online, have been rapidly increasing. This type issues sometimes yield unexpected side effects, such as product boycott. Therefore, it is important to recognize them as early as possible. Existing researches on social issue extraction have mainly focused on news and newspapers. However, the epicenter of viral spreading social issues is not these media but online public opinion. In this study, we propose a constructive approach-based method, which is inspired by the social issues research approach, called "Constructivism", for early extraction of viral spreading social issues. Our experiments revealed that the proposed method successfully extracted six out of ten viral spreading social issues earlier than the first TV news coverage and the average lead time was 20.5 days to the first national TV news coverage.

    Download PDF (486K)
  • Tetsuro KAMURA, Hideaki TAKEDA
    Session ID: 1B5-GS-6-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this study, we extracted 262 artists who are considered well known based on the data for 4,730 Japanese oil painters collected from almanac art books. We digitized information on these artists such as pieces sold on the art market, information found on the Internet, and valuations, and we used these indicators to visualize the relationship of groups (Japanese oil painters) through a machine classification.

    Download PDF (980K)
  • Jawad HAQBEEN, Takayuki ITO, Rafik HADFI, Sofia SAHAB, Tomohiro NISHID ...
    Session ID: 1C3-OS-6a-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Our paper is especially focused on the international experiment transfer of D-Agree usage and application for wicked city problems of fragile and conflict-affected city like Kabul. The content includes the background, process and formalization of the collaboration and reports on the outline of the Societal experiment of D-Agree in selected district of Kabul Municipality, and outcome-based content analysis of the Kabul experiment compare to Nagoya experiment case. It also mentions the implications of international differences (Nagoya-Kabul) in social systems highlighted from the episodes of the debriefing discussion conducted through utilizing D-Agree.The final objective of our study is to evaluate facilitation of citizens discussion regarding access to city meeting in order to solve urban renewal problems, and evaluate AI-based facilitation of system. Method. Our action research is based on the soft system methodology of Checkland et al. and applies the gaming exercise. We conducted an action research regarding the implementation AI-based facilitation city-citizens discussion within the context of NITech-KM international collaboration by utilizing D-agree and reported its progress.

    Download PDF (925K)
  • the Case of Surrendering Driving License of the Elderly People
    Tomoyuki TATSUMI, Takashi NAKAZAWA, Naoki FUKUTA, Hiroshi YOSHIDA, Min ...
    Session ID: 1C3-OS-6a-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    This paper examined the effects of different facilitation methods in online discussion and showed the possibilities and challenges of automated facilitation. Using an online discussion support system "D-Agree'', we conducted a discussion experiment to compare discussion activities of groups with automated facilitation based on Issue-Based Information System (IBIS) structure, groups with human facilitation, and groups without facilitation. The topic was "surrender of driving license of the elderly people''. The results showed that both automated facilitation and human facilitation seemed to activate the discussion. However, a qualitative evaluation revealed that automated facilitation posted more inappropriate comments than human facilitation, and that facilitation posts, regardless of the methods, were often not responded by participants although the factors behind it were not clearly identified by the data of this experiment.

    Download PDF (345K)
  • Akira KAMIYA, Tokutaka HASEGAWA, Shun SHIRAMATSU
    Session ID: 1C3-OS-6a-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Recently, social issues that threaten sustainable development are increasing in Japanese local societies. To address these issues, public collaboration between citizens, government, and experts is important. To discuss how to address these issues, surveying related activities are important. We aim to develop systems for collecting local social issues and actual solutions from Web articles to support the survey and discussion.In this paper, we tried two approaches: (1) developing an automatic tagging system for social issues on Web articles and (2) developing an information extraction system for related cases from Web articles to address them. For the approach (1), we propose a method for classifying each sentence in Web articles into social issue tags using a BERT pre-trained on Japanese Wikipedia. For the approach (2), we propose a method for extracting cases related to social problems from the Web articles. First, we design and develop a corpus of social issues using the Web Annotation Data Model. Second, we describe a method of extracting related cases on social issues from Web articles using a corpus as training data.

    Download PDF (682K)
  • Ai TOIDA, Satoshi IKADA
    Session ID: 1C3-OS-6a-04
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this paper, we propose a new method for generating overall optimized transportation plans using inter-AI cooperative technology. The proposed method encourages collective cooperative behavior by multi AI agents to achieve more socially overall optimized transportation plans which cannot be achieved by an individual company. In this method, a new overall optimized set of transportation plan is generated from a given set of transportation plans which are planned by each company. The method is formed as a kind of cooperative game played by logistics companies, transportation companies and shippers. This overall optimization process requires adjusting conditions such as pickup/arrival time or price among players. Those processes are very complicate and time consuming for humans. We proposed method adapt to logistics problems possible by multi AI agents which automatically negotiate technique for agreement each other.

    Download PDF (441K)
  • Naoki FUKUTA
    Session ID: 1C3-OS-6a-05
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this paper, we summarize our current approach and concerns on a computational mechanism with a set of software tools for querying online discussions and argumentations with ontology inference on linked open data access over different domains for making a better knowledge repository. The approach is based on, and extended from our previously presented prototype designs which are utilizing Linked Data and Knowledge Graphs structure with query transformation techniques.

    Download PDF (148K)
  • Hirofumi MASUI, Kazuki NAKABAYASHI, Tadahiro TANIGUCHI
    Session ID: 1C4-OS-6b-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Communication-field mechanism design includes rules and incentives to indirectly control the communication of a group of people, e.g., discussion, debate, meeting, and consultation, by introducing constraints to the communication process. Similarly, we hypothesize that such constraints are beneficial for the application of speech recognition technologies based on artificial intelligence. In this paper, we evaluate this hypothesis by using an automatic speech recognition system with Dealing Rights to Speak (DRS) as a proof of concept. Our experimental results show that the introduction of DRS can effectively improve the performance of the speech recognition system.

    Download PDF (286K)
  • Naoko YAMAGUCHI, Takayuki ITO
    Session ID: 1C4-OS-6b-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    We have been studying a large-scale consensus support system with agent technology, and had developed a prototype system named D-Agree which has the automated facilitation function. In order to verify the practical application of D-Agree, we conducted a workshop (AgentCrowd2019) with using D-Agree in the international conference (PRIMA2019). The aim of this workshop was sharing knowledge and future co-creation regarding to the agent studies through the collective dialogue. We used D-agree in the divergence phase for brain storming, and made a roadmap as a deliverable of the workshop. In this paper we show the effect of D-Agree on the workshop with the analysis results both discussion logs and roadmap contents.

    Download PDF (511K)
  • Ryuta ARISAKA, Takayuki ITO
    Session ID: 1C4-OS-6b-04
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    There is an increasing interest in formally modelling many-person-argumentation. Multi-agent argumentation is an active research field within formal argumentation that studies the topic. In this work, we illustrate research challenges of multi-agent argumentation, particularly of that with 3 and more participants, which, compared to 2< scenarios, has not been widely studied. We highlight the recent progress.

    Download PDF (144K)
  • Koshi OTA, Koichi FUJIWARA, Nao INATSU, Toshitaka YAMAKAWA, Takatomo K ...
    Session ID: 1C5-GS-13-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Objective: This study proposes a heat stroke detecting model using heart rate variability (HRV) and an anomaly detecting technique. If heat stroke can be detected at an early stage, its patient can rest before it is too late. Methods: Since it is reported that heat stress influences HRV, we developed the heat stroke detecting model that detects abnormality in HRV due to heat stroke by using multivariate statistical process control (MSPC). Results: We measured the HRV data from 30 employees who worked in multiple steel works, whose total data length was 1,042 hours. The result of applying the developed heat stroke detection model showed that the sensitivity of 40.0% and the false positive rate of 1.9 times per hour. Conclusion: This study constructed an early-stage heat stroke detection model by using HRV analysis and MSPC. Significance: This study illustrated the possibility of detecting heat stroke by using HRV analysis.

    Download PDF (726K)
  • Shunya HANAI, Shohei KATO, Koichi SAKAGUCHI, Takuto SAKUMA, Reiko OHDA ...
    Session ID: 1C5-GS-13-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In recent years, developed countries such as Japan have become a super-aging society, and a further increase in dementia patients is a serious problem. Dementia has different causes and treatments depending on the underlying disease, so it is important to diagnosis the disease correctly. However, some diseases are difficult to diagnosis by a general practitioner, and frontotemporal lobar degeneration (FTLD) is one of them. FTLD is a neurodegenerative disease that causes dementia and is a designated intractable disease in Japan. This disease has fewer cases than other dementia and is difficult to distinguish from Alzheimer's disease (AD). So, patients with suspected FTLD should be diagnosed by a specialist. Therefore, an easy screening is needed to refer patients with suspected FTLD to a specialist. In this study, we attempt to distinguish three groups of FTLD, AD, and healthy control (HC) using speech. We used ensemble learning to resolve the data imbalance, and classified by acoustic features extracted from speech. As a result, the above three groups were classified with 82% accuracy, 0.74 F-measure. Therefore speech analysis-based screening using ensemble learning is effective in classifying target diseases.

    Download PDF (404K)
  • Maho SHIOTANI, Takahiro HIYAMA, Yoshikuni SATO, Jun OZAWA, Yoshiyuki K ...
    Session ID: 1C5-GS-13-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Weaken lower limb muscle strength for elderly people is a risk of fall and gait disability. Therefore, early detection and prevention of weaken muscle is important. In previous study, relationship between fastest gait movement and lower limb muscle strength was researched, though the purposed measurement method can be difficult for elderly subjects. So, other method which use daily gait movement is needed. In addition, gait movement is different by gender, but there is no study researching about estimation method of muscle strength considering difference of gender. This research aimed to construct model estimating lower limb strength by gender, using daily gait movement. Daily gait movement of 44 healthy subjects (22 for male and 22 for female) were measured. Then, 3 types of estimation model for lower limb strength were constructed; using gait movement data from male, female, and all subjects. As a result, both in single gender data model correlation coefficient of estimated value and measured value is over 0.7 and mean absolute error (MAE) was 0.12 N/kg but lower accuracy was shown in model using data from all subjects including both gender which correlation coefficient is 0.55 and MAE is 0.11 N/kg. Therefore, importance of gender information in analyzing relationship between lower limb strength and gait movement was shown.

    Download PDF (883K)
  • Tetsuya SHIRAISHI, Yuriko HIJIKATA, Kazuo TOKUSHIGE, Shouichirou ISHIH ...
    Session ID: 1C5-GS-13-04
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    The aim of this study was to establish a method for estimating functional prognosis in cerebral hemorrhage patients. Supervised automated machine learning using neural network and gradient boosting decision tree algorithm demonstrated that FIM (functional independence measure) gain can be predicted by feature values such as age, sex, location and size of cerebral hematoma and FIM score. For motor FIM gain (total amount of 13-item motor subscale score), features with high contribution rate were listed such as extension site of hematoma, age, size, volume and location of hematoma. These features will be useful to predict patient’s clinical functional recover in rehabilitation conference.

    Download PDF (246K)
  • Taku RI, Tatsuya YAMAZAKI
    Session ID: 1C5-GS-13-05
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this paper, we propose a method for detecting nodules that cause cancer from three-dimensional (3D) computer tomography (CT) images of the lung field. In the proposed method, a 3D image extracted from the lung field of a CT image is input to a model constructed by a 3D convolutional neural network. Then, when the input small area image is determined to be a nodule, a mark is labelled at the corresponding position in the CT image. In this paper, in order to verify the effectiveness of the model in the proposed method, the classification accuracy is verified using a model constructed with two classes, nodules and non-nodules, and then several CT images were used to detect nodules. As a result, the accuracy rate of the model was 94.44%. However, some false-positives were confirmed by nodules detection.

    Download PDF (791K)
  • Masahiro SUZUKI, Hiroki SAKAJI, Kiyoshi IZUMI, Hiroyasu MATSUSHIMA, Ya ...
    Session ID: 1D3-GS-13-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this paper, we propose a methodology of forecasting the change rate of net income which an analyst estimates by applying natural language processing and neural networks in the context of analyst reports. We examine the contents of the reports for useful information on forecasting the direction of revision in analyst estimate earnings. First, our method extracts opinion sentences from the reports while the remaining parts are classified as non-opinion sentences.Second, our method forecasts movements of analyst estimate earnings by inputting the opinion and non-opinion sentences into separate neural networks. In addition to the reports, we input the trend of the net income to the networks. As a result, we found that there were differences between securities firms depending on whether analysts' net income estimates were based on opinions or facts. We also found that the trend of the net income was effective for forecast as well as an analyst report.

    Download PDF (333K)
  • Shogo AKIYAMA, Junichi EGUCHI, Tomoya SUZUKI
    Session ID: 1D3-GS-13-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    ESG investment is getting popular as an investing method that evaluates environmental, social and governance (ESG) efforts of each company. However, although the disclosure of information on ESG is already common, it is difficult to objectively evaluate ESG efforts of companiesdue to the ambiguity in the definition of ESG words: environment, society, and governance. For this reason, we applied Word2Vec to extract similar words to each ESG word, and evaluated ESG efforts of companies from the viewpoint of these extracted words. As a result, we confirmed that the companies with higher ESG score can make a more profitable portfolio than those with lower ESG score, and this ESG score can be useful for factor investing.

    Download PDF (337K)
  • Hiroyuki SAKAI, Hiroki SAKAJI, Kiyoshi IZUMI, Tohgoroh MATSUI, Keitaro ...
    Session ID: 1D3-GS-13-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this research, we propose a method for extracting sentences containing causal information from articles describing the market conditions of the Nikkei Stock Average. The sentences containing causal information are needed to generate market analysis comments. Our method extracts articles describing the market conditions of the Nikkei Stock Average from economic newspaper articles and extracting sentences containing causal information from the extracted articles by deep learning. Here, our method automatically generates the training data necessary to extract the articles describing the market conditions and sentences containing causal information by deep learning and achieved high accuracy. Moreover, our method extracts complementary information of the content described in the causal sentences by using economic causal-chain search.

    Download PDF (391K)
  • XIN ZHANG, YUSUKE MOTOKI, YUYA SONEOKA, YUSUKE IWASAWA, YUTAKA MATSUO
    Session ID: 1D3-GS-13-05
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Natural language processing tasks targeting the SDGs (Sustainable Development Goals), which have started to influence social structures and corporate philosophy, have recently begun. Because of the lack of language resources, efforts in Japanese were difficult. In this study, we collected Japanese SDGs-related data from materials published by universities and created a data set. And the SDGs classification model was constructed. As the augmentation method, 1. a part-of-speech replacement using the BERT MASK model 2. A reverse translation method in which the English translation using Google transfer was translated into Japanese again was used. Classification was performed using a topic model (LDA etc.) which is a classical machine learning method and BERT etc. which is a deep learning model. The results show the results of the augmentation in the minority data task. Produces relatively high accuracy in a small number of data.

    Download PDF (615K)
  • Integration of Intelligent Dialogue Agent and Cognitive Training System
    Sho HIROSE, Daisuke KITAKOSHI, Kentaro SUZUKI, Akihiro YAMASITA, Masat ...
    Session ID: 1D4-GS-13-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In Japan, an advanced super-aged society, preventive care approaches have been paid much attention.These approaches aim to decrease the number of people who require nursing care or support.In this paper, we develop Comprehensive Preventive Care System (CPCS) aiming to realize care prevention activities such as cognitive training and watching over by family/friends through dialogue with older adults. We evaluate the proposed system's impressions and effects by conducting several experiments employing prototype version of the CPCS prototypes.In addition, we apply a Speech Content Coordinating Function (SCCF) using reinforcement learning based on policy gradient method to the IDA that allows the agent to perform natural dialogue. This paper investigates whether the SCCF can appropriately adjust the agent's speech content depending on the users' situation.

    Download PDF (1083K)
  • Hiroaki Sugiyama SUGIYAMA, Kenji NAKAMURA, Yoshihiro HARADA, Tatsuya O ...
    Session ID: 1D4-GS-13-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In recent years, the population is aging rapidly, and many people spend their days in nursing homes. While such services are expected to maintain cognitive functions through communication with others, it is not always possible to get enough communication on a daily basis because the long working hours of facility staff become a social problem. To relax this tough situation, dialogue agents are expected to chat with people. However, at present it is not possible to achieve satisfactory chats with people. In this study, we focus on a simpler form of dialogue, the game of Shiritori. We verify whether a game with a dialogue agent is effective for maintaining cognitive functions through a three-month WoZ experiment. Through our experiment, playing Shiritori with a dialogue agent significantly improves cognitive measures of dementia patients.

    Download PDF (1347K)
  • Jun SONODA, Tomoyuki KIMOTO, Yasushi KANAZAWA
    Session ID: 1D4-GS-13-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In recent years, marine plastic has become a world problem. In this study, we have developed an automatic detection method for the marine plastic in/on the beach by the ground-penetrating radar (GPR) and the unmanned aerial vehicle (UAV) images with the deep learning. We have generated the GPR images for training using a fast finite-difference time-domain (FDTD) simulation with graphics processing units (GPUs). Also, we have made the training images of plastics by UAV images. The training images have been learned by a 5-layers convolutional neural network (CNN) and the YOLOv3. We have shown that unlearned plastics images in/on the beach can be detected with 95% accuracy by using our proposed method.

    Download PDF (317K)
  • Nanako NARIKIYO, Karin KITAMURA, Toya TAKIGUCHI, Hikaru TAKAHASHI, Sak ...
    Session ID: 1D4-GS-13-04
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    The purpose of this study is to help companies to create an environment which workers can work friendly.Recently,there is a problem of mental health for workers.The main cause is the stress which workers feel when they work for a long time and when they make bad human relationships in their job. So,we examined an effective proposal for mental health care with AI.Concretely,we focused on a stress check sheet which is one of mental health care.We hypothesized that information presentation of stress check is an effective method. As a result, information presentation did not influence a boss to change his or her will and action.So,we found that information presentation isn’t an effective way.As a supplementary, we introduced three personality traits. We found that people who can accept a negative opinion would like to improve their team in spite of the information presentation.

    Download PDF (358K)
  • Implementation science from proof of concept to Evidence-Informed Policy Making
    Kota TAKAOKA, Jiro SAKAMOTO, Emiho HASHIMOTO, Daiki HOJO, Yui FURUKAWA ...
    Session ID: 1D4-GS-13-05
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    The number of reported child abuse cases has been dramatically increased in Japan. To support children and practitioners in the field, we developed AI, called “AiCAN” stand for Assistance of Intelligence for Child Abuse and Neglect, and we have been conducting the Proof of Concept (PoC) in a municipality, Japan. The purpose of the presentation is how AI would be useful for practitioners and policy making. AiCAN has two different algorithms such as machine learning and probabilistic modeling for different practitioners’ needs. As a result of the PoC, we found the AI could create the cycle of value-chain in the field. To develop AiCAN continuously, additional studies and agile development are required.

    Download PDF (300K)
  • Kazuhito TAMURA, Kazuo HARA, Ikumi SUZUKI
    Session ID: 1D5-GS-9-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In recent years, neural language models have been reported to have successfully acquired English grammar. In this paper, we investigated whether the neural language model can acquire Japanese grammar. We used the case frame dictionary of Kyoto University to generate Japanese sentences and non-Japanese sentences to examine language models. We compare the performance of neural language model and n-gram models in experiments.

    Download PDF (596K)
  • Hiroya TAKAMURA
    Session ID: 1D5-GS-9-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    This paper provides a brief overview on recent methods for the data-to-text generation task and discusses the detailed specifications of a method to be constructed for a given specific case. The components that are possibly added to the method include the encoder of the input, attention mechanism, copy mechanism, content selection and planning, and content tracking.

    Download PDF (354K)
  • Satoshi HIRADE, Eiichi TANAKA, Takeshi ONISHI
    Session ID: 1D5-GS-9-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this paper, we present a novel method for learning word embeddings. However, several word embedding approaches with extracting subwords from target word have been proposed, those methods have the problem of leaving subwords without meaning associated with the target word. These subwords have negative effects on obtaining better performance of word embeddings. To solve this problem, we adopted switching subword extraction rules based on Japanese character types. With this contrivance, the appearence of the subwords are surpressed. As a result, our method achieved better results on word similarity task than Word2Vec and FastText.

    Download PDF (400K)
  • Kotaro YAMAGUCHI, Natsuki OKA, Tadahiro TANIGUCHI, Ryo OZAKI
    Session ID: 1D5-GS-9-04
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Infants need to segment their native language into phonemes and words at the same time without supervision. Taniguchi, Nagasaka, & Nakashima (2016) showed that Nonparametric Bayesian Double Articulation Analyzer could analyze latent double articulation structure, i.e., hierarchically organized latent words and phonemes, of utterance data consisting of a limited vocabulary in an unsupervised manner by assuming hierarchical Dirichlet process hidden language model (HDP-HLM). In this study, we attempted unsupervised double articulation analysis of natural speech in a video game environment and tried to give meaning to the segmented words. The result of an experiment demonstrated that the utterances were roughly correctly segmented, and the meanings of up, down, left and right were almost correctly learned.

    Download PDF (502K)
  • Soichiro MURAKAMI, Sora TANAKA, Masatsugu HANGYO, Hidetaka KAMIGAITO, ...
    Session ID: 1D5-GS-9-05
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Data-to-Text is the task of generating text that describes the contents from various types of data, such as weather forecast maps and time-series stock data. With the recent advances in neural networks, the data-to-text models have been making remarkable progress in terms of the ability to capture the characteristics of such complicated data and generate text that accurately describes their contents. However, concerning the generation of a document-scale text that includes multiple sentences, the data-to-text model often generates descriptions that mention duplicated contents and lacks the consistency of their topics due to the over-generation problem. In this paper, we focus on generating descriptions having not only accuracy but also consistency regarding their contents to tackle the problem mentioned above, which often causes in the data-to-text task. We propose a method to generate consistent text by predicting a sequence of topics from data and assigning it to the generation model to control the topics and their order. In the experiment, we show that the data-to-text model produces text containing consistent topics by specifying a sequence of topics to be mentioned and that it helps to relieve the over-generation problem.

    Download PDF (579K)
  • Toward Supervised Learning for Analogy Task using Word Vectors
    Shohei HIDAKA
    Session ID: 1E3-GS-9-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Making analogical inference in vector space has become a standard method to test the quality of word vectors. Typically the operator for the analogical inference is manually optimized for the task. In this study, we consider a systematic optimization of the metric-based rank order function for the analogical inference. If we directly evaluate the rank order function, one needs to process a few millions of word vectors every step of optimization. This causes a considerably large computational cost which makes a systematic optimization of such analogical inference intractable. In this study, we propose a theoretical approximation for this rank-order evaluation, and demonstrate an optimization of the analogical inference using the approximated evaluation. Lastly, we discuss about the ``parallelogram'' relationship, which may or may not have a deep connection with the well known ``analogy parallelogram'', revealed by the mathematical analysis of the probability of the distance-based rank order.

    Download PDF (509K)
  • Hiroyuki SHINNOU, Jing BAI, Rui CAO, Wen MA
    Session ID: 1E3-GS-9-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this paper, we point out the problem that BERT is domain dependent, and propose to construct the domain specific pre-training model by using fine-tuning. In particular, parameters of a DistilBERT model are initialized by a trained BERT model, and then they are tuned from the specific domain corpus. As a result, we can efficiently construct the domain specific DistilBERT model. In the experiment, we make the test set for each domain, which is the estimation of a masked word in a sentence. By this test set, we evaluate the domain specific DistilBERT model by comparing with the general BERT model, and show the superiority of our proposed model.

    Download PDF (497K)
  • Katsuhiko UTSUBO
    Session ID: 1E3-GS-9-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In Japanese sentences, the meaning of the context may differ depending on the insertion point of punctuation, so the position of the punctuation is very important. In this research, we create a general method that automatically complements punctuation from text information using deep learning. The proposed method is that the corpus is split using morphological analysis and replaced infrequent words with parts of speech and performs classification of exists of a period or comma using LSTM from word strings before and after the target position. The accuracy of classification has been improved by setting a threshold for the probability output by the model. Furthermore, by limiting the number of input words and replacing them with parts of speech, the calculation time can be reduced without reducing the calculation accuracy. Using this method, experiments using broadcast manuscripts as text corpora have confirmed the effectiveness of this method.

    Download PDF (684K)
  • Maiko ONISHI, Hitomi YANAKA, Koji MINESHIMA, Daisuke BEKKI
    Session ID: 1E3-GS-9-04
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Recognizing temporal relations among events and time expressions has been one of the most challenging tasksin natural language processing. Recent studies mainly focus on deep learning-based models trained with a largetemporal relation corpus. However, it is unclear whether these models can accurately perform complex inferenceswith temporal phenomena. In this paper, we present an inference system to perform inferences over temporalrelations. We use a higher-order inference system based on Combinatory Categorial Grammar (CCG), a systemthat converts input sentences to semantic representations via derivation trees and proves entailment relations viatheorem proving. We show that by adding lexical entries and axioms for temporal relations, the system can performlogical inferences over multiple temporal relations.

    Download PDF (390K)
  • Izumi HARUTA, Koji MINESHIMA, Daisuke BEKKI
    Session ID: 1E3-GS-9-05
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Comparative constructions pose a challenge to Natural Language Inference (NLI), a task of determining whether a text entails a hypothesis. Comparatives are structurally complex in that they interact with other syntactic and semantic phenomena such as ellipsis, numerals, and lexical antonyms. In the context of Formal Semantics, there is a rich body of work on the compositional semantics of comparatives and other gradable expressions on the basis of the notion of degree. However, a computational inference system for comparatives is not developed enough to be used for NLI tasks. In this paper, we present a compositional semantics that maps various comparative constructions in English to semantic representations using current state-of-the-art Combinatory Categorial Grammar (CCG) parsers and introduces an inference system using automated theorem proving which effectively computes complex logical inference with comparatives. We evaluate our system on three NLI datasets that contain complex inferences with comparatives, quantifiers, and numerals. We show that the system achieves better performance, in comparison with recent logic-based and deep-learning-based NLI systems.

    Download PDF (376K)
  • Hayato TOBE, Hiroyuki KANEKO, Kazuhiko MASUMOTO, Takekazu MATSUKAWA
    Session ID: 1E4-GS-9-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In the design of civil engineering structures and buildings such as tunnels, dams, bridges and apartments, it is important to collect geological information on the foundation ground or surrounding ground for safety and appropriate construction. However, there is a large amount of geological literature to be viewed for these designs and constructions, and there is always a shortage of geological engineers who can decipher and summarize the literature. Therefore, we have been developing a text mining system to automatically and quickly analyze and extract important information related to construction from geological literature. As a result of comparing the information automatically extracted by this system with the information compiled by the geotechnical engineers, we found that the former information was comparable to the latter.

    Download PDF (467K)
  • Kenzo KUROTSUCHI, Morimoto YASUTSUGU, Sato MISA, Kohsuke YANAI
    Session ID: 1E4-GS-9-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    For decision support, we developed text mining method to extract information about physical property values related to the decision from a large amount of accumulated technical papers. In numerical expressions, we think that the usefulness will be enhanced by not treating only numerical values, but extracting them from the document by combining the numerical values related to which events (Attribute Value Extraction). We reported a method for extracting pairs of item names and numerical values using StruAP. As a result of the evaluation, we confirmed the effectiveness of this method.

    Download PDF (375K)
  • Kaito TAKANO, Hiroyuki SAKAI, Kei NAKAGAWA
    Session ID: 1E4-GS-9-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    The shareholder convocation notice contains a lot of useful information, such as company profile, major shareholder, and bills to be discussed. The purpose of our research is to automatically extract pages that are likely to affect the stock price from shareholder convocation notices. To this end, we need to tag the pages to automatically extract what information is described on a page-by-page basis. In our research, we propose the following framework: we automatically create training data by a rule-based method and train the deep learning model that extracts important pages. We confirm the effectiveness of our framework for pages that cannot be extracted by the rule-based method.

    Download PDF (585K)
  • Hitoshi SHIMIZU, Tomoharu IWATA
    Session ID: 1E4-GS-9-04
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    The ranking of product search results displayed to users has a significant effect on sales of an E-Commerce website. We propose a method to improve product search using access logs. Using features obtained from products and queries and access logs as learning data, a neural network is trained to display frequently accessed products at high rank. There are multiple types of access logs, such as not only conversions, but also carts and clicks. We experimentally confirmed that training data other than the target type can improve learning to rank.

    Download PDF (509K)
  • Ryoma YAMASHITA, Kensuke HARA, Satoshi TAMURA, Satoru HAYAMIZU
    Session ID: 1E4-GS-9-05
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    At a call center, an operator searches for an appropriate manual based on an inquiry given by the customer.However, it is difficult to choose a correct manual from retrieval results. The purpose of this study is to analyze manuals used in an actual call center in order to build an effective retrieval method such as a two-question-and-two-answer scheme. We applied a clustering technique to the manuals and analyzed data by cluster visualization. We then investigated the potential to employ the approach. We finally compared the proposed method with the baseline method and showed its effectiveness.

    Download PDF (1034K)
  • Ayana NIWA, Kohei WAKIMOTO, Keisuke NISHIGUCHI, Masataka MOURI, Naoaki ...
    Session ID: 1E5-GS-9-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In recent years, slogans, which are sentences that express products and works, have attracted attention due to their importance in marketing. An analysis of the previous slogans has shown that the interesting points of slogans also appear in rhetorical techniques such as metaphor and repetition. Additionally, some experimental results show that slogans with rhetorical techniques improve advertising effects. Therefore, in this study, we focused on slogans that include an antithesis structure, among many rhetorical techniques. More specifically, towards the generation of slogans with antithesis structures, we perform structural analysis using a corpus with antithesis spans annotated. Then, we can extract the knowledge about the structures. Our proposed model introduces end-position detection for each span before antithesis span identification, which enables the model to capture the word correspondences. The proposed model showed better accuracy in a shorter calculation time by refining the span candidates by the end-position detection.

    Download PDF (444K)
  • Tomoya ISOBE, Po-Hsuan HUNG, Shohei IIDA, Yizhen WEI, Takehito UTSURO, ...
    Session ID: 1E5-GS-9-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this paper, we study a concatenation-based multi-source neural machine translation (NMT) model trained with three-language parallel corpus. We show that the concatenation-based multi-source NMT model where a parallel English and Chinese sentences are input to the model as the source sentences improves the BLEU score of the single-source NMT where only English or Chinese source sentence is input to the model. Among major phenomena where the BLEU improves when translating from the source English sentence than from the source Chinese sentence are translation of Katakana loanwords, tense, and particles, etc., while, in the translation of Chinese words when the Chinese and Japanese words share an identical Chinese character, the BLEU improves when translating from the source Chinese sentence than from the source English sentence. We then show that, in the translation by the concatenation-based multi-source NMT model, the BLEU improves the most by correctly incorporating translation of both types of phenomena in a complementary style.

    Download PDF (1064K)
  • Hongyu LI, Tengyang CHEN, Takehito UTSURO, Yasuhide KAWADA
    Session ID: 1E5-GS-9-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In the field of machine comprehension (MC), the task of an MC model is to predict the answer (A) from a question (Q) and related context (C) of the question. However, in this paper, it is discovered that there exist examples that can be correctly answered by an MC model BERT where only the context of the example is given without the question being given, which means that the difficulty of examples of machine comprehension vary. Based on this finding, this paper proposes a method based on BERT which splits the training examples of the MC dataset SQuAD1.1 into “easy to answer” and “hard to answer” ones. Experimental evaluation results of comparing the two models, one of which is trained with the “easy to answer” examples only, while the other of which is trained with the “hard to answer” examples only, show that the latter outperforms the former.

    Download PDF (910K)
  • Hidekazu YANAGIMOTO
    Session ID: 1E5-GS-9-04
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    I propose Gaussian process regression with neural kernel function and apply it to review rating prediction. Neural network is applied to natural language processing and achieves high performance in various tasks. However, neural networks have less interpretability. On the other hand, a probabilistic model has more interpretability. Gaussian process regression is one of probabilistic models and it is important to select appropriate a kernel function. So I combine a recurrent neural network and Gaussian process regression and aim at achieving high performance and high interpretability. In evaluation experiments, the proposed method achieve lower RMSE and MAE.

    Download PDF (253K)
  • Hayato SUGIMOTO, Ayana KUNO, Kohei TANIGUCHI, Rei HAMAKAWA
    Session ID: 1F3-OS-2a-01
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    Avocados, favorite fruits for many consumers, are to detect if they are ripe or not. There are existing studies on the classification of ripening stages of avocados; however, these studies have not been efficient and convenient for consumers to use. Thus, a different approach that is available to the consumers at large is needed. Here, we propose a method for classifying the ripening stages of avocado using deep learning. This system will help consumers detect a ripe avocado when purchasing or cooking with it. The proposed method uses a new approach to detect avocados and classify them into four ripening stages using two deep learning models from the user's avocado image input.

    Download PDF (1030K)
  • Fumiaki SAITOH
    Session ID: 1F3-OS-2a-02
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    The purpose of this study is to analyze the meaning of Sizzle words that induce consumer appetites and their willingness to purchase. Semantic analysis of these words is important because they are widely used in various stages of marketing, such as advertising, packaging, catch copy, and product development. In this study, we adopt “Word2Vec,” which is a tool used for vectorization of the semantic structure of words. A semantic model is constructed using Word2Vec and is adapted to online food reviews (provided by customers) and includes representative Sizzle words. The semantic analysis of the relation between sizzle words and the surrounding words based on Word2Vec may lead to the discovery of novel Sizzle words and their usages.

    Download PDF (664K)
  • Simulation based evaluation of AGVs operation in food service company
    Tokuro SAKURAI, Hiroatsu KANAZAWA, Nobutada FUJII, Takeshi SHIMMURA
    Session ID: 1F3-OS-2a-03
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    This research focuses the restaurant industry which the labor productivity is relatively low. In an action for improving the productivity, AGV (Automated Guided Vehicle) transports food from kitchen to the customer (or area near the customer) in some restaurants. This paper proposes an efficient way of AGV operation. Rules of AGV running and battery charge with the goal of reducing total food retention time in the kitchen has been resolved by computer experiment utilizing multi-agent simulation. It was resulted that total food retention time could be reduced 3.5% for 7 days (2,036 to 1,965 min/day) and a maximum of 7.4% for a certain day (1,981 to 1,834 min/day) by changing battery charging rule which avoid heading for automatic charging while dishes being transported remained. By operating AGV efficiently based on the simulation results, it could be done that (1) reduction of labor quantity, (2) increase in customer service time by working staff, and (3) speeding up of food service. Raising customer satisfaction and eliminating the factors of customer dissatisfaction would be also important initiatives in the service industry.

    Download PDF (457K)
  • Tomomi NONAKA, Takeshi SHIMMURA, Nobutada FUJII
    Session ID: 1F3-OS-2a-04
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this paper, we perform a basic analysis on employee satisfaction and production planning by introducing a service robot which delivers dishes in restaurant service. As an example of a service robot introduced in a Japanese restaurant, we focus on the pantry staff, kitchen staff, and customer service staff, as well as production planning that individual employees implicitly plan and update in their heads. The service robots are used to deliver the dishes prepared at the kitchen to the customer service floor, where the worker specifies the destination and transports the dishes while patroling the restaurant. And the robots also play the role of transporting the finished dishes to the washing place. By analyzing the productivity and employee satisfaction before and after the introduction of the service robots into the categories of serving, cooking area, and customer service, how employees can identify and coordinate work between humans and machines and change process design. It is considered whether the staffs are adapting, and how they understand and respond to their own work and operational improvements with the introduction of robots. From December 2019 to January 2020, an analysis was conducted based on the results of employee questionnaires and interviews conducted at a restaurant.

    Download PDF (213K)
  • Ryutaro HAMA, Azusa SAITO, Yukari TAKAKU, Atsushi HASHIMOTO
    Session ID: 1F3-OS-2a-05
    Published: 2020
    Released: June 19, 2020
    CONFERENCE PROCEEDINGS FREE ACCESS

    This study proposes a method to find cooking activity patterns based on the way of using work space along with the progress of the food preparation. Dividing the workspace into process-related and idle areas with three regularly divided time periods, we obtained total six spatio-temporal sections. Then, applying LDA modeling with Bag-of-Items feature from each section, we extracted features of cooking activities. Further applying clustering based on the extracted features, we found five typical patterns of food preparation.

    Download PDF (1012K)
feedback
Top