Tetsuji Kuboyama

Based on the behavior analysis of the co-eating of older adults, this study clarifies the effects of psychological health on meal time conversation. We conducted a dietary experiment in which the same menu was provided in individual and platter formats to four groups consisting of six older adults. From the recorded video, participant's speech behaviors were annotated from the start of eating to 20 minutes. The subjects of analysis were the amount of speech and topics of the participants. The results revealed that during the meal, the speaker could easily secure a situation in which they continued to speak at a certain time, and the participants activated the interaction on the topic of cooking. This study found that co-eating conversation functions as a place to deepen mutual understanding among participants and to give the first meeting an opportunity to talk with each other.

局所性鋭敏ハッシュ（LSH）の一種であるスケッチを用いたk近傍検索について議論する．スケッチを用いるk近傍検索は2段階で行う．第1段階では，質問とのスケッチ間の距離が近いK個の解候補を選択する．ただし，K ≥ kである．第2段階では，K個の解候補に対して実距離計算を行うことでk近傍解を選択する．従来，高い検索精度を保証するためには32ビット以上のスケッチを用いていた．本研究では，スケッチのビット数を16に減らすことにより，バケット法を用いたデータ管理による高速化を可能とし，第1段階の検索コストをほとんど無視できる手法を提案する．16ビットスケッチを用いる検索は，精度を維持するためには，32ビットスケッチを用いる場合より大きな候補数Kを必要とする．データオブジェクトをスケッチの値によってソートしておくことで，第2段階の検索におけるメモリ局所性を向上することで候補数Kの増加による速度低下を低減できる．提案手法を用いると，検索精度を維持しつつ10倍程度の高速化が実現できる． We discuss k nearest neighbor search using sketches, which is a kind of locality sensitive hash (LSH). Search using sketches is processed in two stages. The first stage is to select K solution candidates close in distance between the question and the sketch, where K ≥ k. In the second stage, k nearest neighbor solutions are selected by performing real distance calculation on K candidates. Conventionally, to ensure high search accuracy, sketches of 32-bit or more have been used. In this paper, we reduce the width of sketches to 16-bit for which efficient data management by bucket is applicable. We propose a search method that enables high speed with the first stage of negligible cost. Searches using 16-bit sketches require a larger number of candidates K to maintain accuracy than using 32-bit sketches. However, by sorting the data objects by their sketch values, memory locality in the second stage search is improved and influence by increasing K is canceled. By using the proposed method, about 10 times speedup can be realized while maintaining search accuracy.

application/pdf 研究報告