HTML5 Webook

15/64

NICT REPORT 13a noun, such as “What can be done to solve global warming?” and “What if” questions such as “What if global warming continues?” However, WISDOM X can also answer “Why” questions to find reasons for things, such as “Why has Japan fallen into deflation?” The re-sponses to such questions were relatively long; however, and not suitable as answers. In FY2017, we developed a deep learning technology that summarizes such long re-sponses, enabling shorter answers to be giv-en, such as “Because society is continuing to age” for the question above. As a computing platform supporting this technical development, we have integrated a deep learning framework into the RaSC mid-dleware. RaSC was developed at this center Fig.2 : WEKDA dialog system architectureand has been used in the past to run large scale software such as WISDOM X on a large scale cluster in parallel and at high speed. This has enabled us to run WEKDA, which is an amalgamation of deep learning technolo-gy, at low cost and high speed. While con-ducting this research, we also developed a new batch scheduler to improve utilization of GPGPUs, and were able to improve efficiency of GPGPU use in the Institute. Image analysis technology R&DWe have developed a method for cluster-ing large volumes of tourism photographs collected from SNS according to structure, as a technology to automatically build a tourism-support image corpus. The method is imple-mented by performing clustering on a graph with vertices for each image, connected by comparing local feature points (a match graph)(Fig.3). Earlier clustering methods had the disadvantage of not being able to detect small clusters, so to solve this problem, we developed a new clustering method using a random walk technique. This method was presented at the IEEE International Confer-ence on Image Processing (ICIP) 2017. In the proposed method, the accuracy of clustering would be negatively affected by making too many steps in the random walk, so we also developed a method to prevent taking too many steps, and conducted tests to confirm that it is effective. User utteranceRoute branchKeyword extractionEmotion analysisDecide route according to type of user utterance“What should we eat in Kyoto?”“It's hot today isn't it?”“It's hot!”Generate response utteranceAnswer rankingGenerate response utterance from the answerResponse utteranceDefaultQuestionShort feedback utterances “iPS cells are amazing, aren't they?”Generate question“Yes, they found a possible treatment for hypertrophic cardiomyopathy using iPS cells.”“Eating boiled tofu in winter in Kyoto is a matter of taste.”Search for the question as-is. Look up responseLook up response“A research group at Keio University found a possible treatment for hypertro-phic cardiomyopathy using iPS cells.”Default routeReturn short sounds/responses indicating comprehensionChange the form of the answer to an appropri-ate response consider-ing current feeling.“What would you like to see about iPS cells?”Add a new module in the future!Rank answers and select an appropriate response.Fig.3 : Technology to build an image corpus to support tourism automaticallyResearch and Development