深澤 佑介

本研究では、大規模自然言語モデルを用いてSNS投稿からメンタルヘルスの状態を推定する手法を提案する。短文投稿SNSサービスXからメンタルヘルス関連の単語を含む日本語の投稿を収集した。投稿者のメンタルヘルスの不調の有無のアノテーションを3名で行いその多数決によりラベルを決定した。ラベル付きの投稿データを教師データとして日本語BERTのファインチューニングを行った。機械学習による分類結果に比べ、BERTは正解率が10％程度向上した。この結果から、大規模テキストデータによって事前学習したモデルがメンタルヘルスの不調の推定タスクの精度向上に寄与する可能性を示した。

本研究では、メンタルヘルスの不調に関するSNS投稿いわゆる「病みツイート」に特有のオノマトペを抽出し、メンタルヘルス文脈で使われる単語との関係性を解明する。まず、既知のオノマトペ（「イライラ」「ドキドキ」など）に加え、メンタルヘルスの文脈で新たに登場したオノマトペ（「ヘラヘラ」「ガタガタ」など）も抽出した。次に、メンタルヘルス関連の単語を複数手法により抽出し、その単語と共起するオノマトペをLDAによりトピック分類した。トピック分類結果からオノマトペはメンタルヘルス不調に伴う身体感覚や心的経験に関する単語と共起しており、オノマトペからメンタルヘルスの不調を推定できる可能性があることを示した。

本研究では、大規模自然言語モデルを用いて遭難発生時の連絡を受けた際にわかる状況から最終的に判明する被害の状況を予測する手法を提案する。遭難の状況として、日時・天候・山域・山名・住居・性別・年齢・態様・パーティ人数・遭難の状況を結合したテキストデータを準備する。ラベルとして、死亡あるいは生存の2つのラベルを準備する。ラベル付きの遭難状況を教師データとして日本語BERTのファインチューニングを行った。機械学習による分類結果に比べ、BERTは正解率が2％程度向上した。この結果から、大規模テキストデータによって事前学習したモデルが遭難時の被害状況推定タスクの精度向上に寄与する可能性を示した。

In this research, we propose a method of forecasting application demand for qualification exams. Due to the examination dates and venues are predetermined, prospective candidates consider the application deadline and available seats when applying for the exam. Therefore, last-minute applications and the impact of remaining seats is essential. In this paper, we forecast future application demand by multivariate time series data analysis using LSTM. The evaluation reports the results of the comparative experiment of predictive accuracy using the application data of the Eiken S-CBT.

本研究では、大規模自然言語モデルを用いてクマ類と遭遇したときの状況から最終的に判明する被害の程度を予測する手法を提案する。クマ類と遭遇したときの状況として、日時・被害発生場所・被害者の性別・被害者の年齢・クマ類と遭遇時の人数・被害発生状況を結合したテキストデータを準備する。さらに、クマ類による人身被害に関する過去の知見のテキスト化を行った。ラベルとして、死亡、重傷、軽傷、目撃の4つのラベルを準備する。ラベル付きの人身被害発生状況を学習データとして日本語BERTのファインチューニングを行った。2021年～2023年の北海道及び本州の一部地域のクマ類による人身被害データで検証を行い、機械学習手法に比べて、提案手法の精度向上を確認した。

we propose a method of predicting human's activity, including the location and purpose, by using Twitter posts with location information. The proposed method predicts target users' activities based on the location transition and tweet of users in the database. Concretely, we adopt both the similarity of current location and interest, and the similarity of long term interest and location to select the base user and tweet. And then, we can utilize these two baselines to predict target users' activities. We evaluate the proposed method by the following two points: one is the error range of the distance, and the other is the similarity of tweet contents. We used three months of Twitter data with location information (almost 40 mil.) as the database. The experiment results demonstrate that the prediction accuracy of the proposed method is superior to the two control groups which only consider one of the similarity of current location and interest and the similarity of long term interest and location.we propose a method of predicting human's activity, including the location and purpose, by using Twitter posts with location information. The proposed method predicts target users' activities based on the location transition and tweet of users in the database. Concretely, we adopt both the similarity of current location and interest, and the similarity of long term interest and location to select the base user and tweet. And then, we can utilize these two baselines to predict target users' activities. We evaluate the proposed method by the following two points: one is the error range of the distance, and the other is the similarity of tweet contents. We used three months of Twitter data with location information (almost 40 mil.) as the database. The experiment results demonstrate that the prediction accuracy of the proposed method is superior to the two control groups which only consider one of the similarity of current location and interest and the similarity of long term interest and location.

The web is thriving with websites containing How-To articles and DIY guides. In websites like that information about activities and guides how to accomplish them are located. An activity is defined as any action that a person can accomplish. Our research focuses on Retrieving activities from the Web related to a user's search query. To accomplish that we have created a system that performs meta-search in multiple How-To websites using semantic query expansion. An overview of the activities retrieval system is presented along with the explanation of the query expansion algorithm which utilizes ConceptNet. Finally experimental results gathered from a user study are given in order to evaluate the performance of our system.The web is thriving with websites containing How-To articles and DIY guides. In websites like that information about activities and guides how to accomplish them are located. An activity is defined as any action that a person can accomplish. Our research focuses on Retrieving activities from the Web related to a user's search query. To accomplish that we have created a system that performs meta-search in multiple How-To websites using semantic query expansion. An overview of the activities retrieval system is presented along with the explanation of the query expansion algorithm which utilizes ConceptNet. Finally experimental results gathered from a user study are given in order to evaluate the performance of our system.

This paper proposes a content-based recommendation method tuned for multiple content domains. Most content-based recommender systems use content attributes (e.g. description of content and category of content) to represent items and user profiles. This raises the problem that when recommending content from new content domain, the recommendation quality will be poor until the item profile representation is updated through the input of a sufficient quantity of contents in the new content domain. In this paper, to develop an item profile representation that is independent of the contents, we propose a method that uses common categories to represent item profiles it focuses on the real world activities (tasks) that attract the user's interest. Concretely, we present a light-weight task-model that covers a wide variety of tasks and yields item profiles. We also create a recommendation algorithm by incorporating the proposed profile representation into a statistical SVM(Support Vector Machine) based recommendation algorithm. Finally, we conduct a user test and the results of this test show 9% higher user evaluation scores of content recommendations compared to an existing content-based recommendation algorithm using term-based profile representation. This shows the effectiveness of our user-centered profile representation approach. © 2011 - IOS Press and the authors. All rights reserved.

本稿では，ユーザの実世界行動モデルの構築を軽減するため，クラスタリング手法の適用可能性を検証する．検証実験の結果，マニュアルで構築したモデルと 50％ 程度重複するモデルを構築することができた．現状，実世界行動モデルをマニュアルで構築するには 1 ヶ月程度かかることから，クラスタリングとマニュアルを組み合わせることで構築に要する時間を劇的に削減できると考えられる．To construct user's real world activity model in an automatic way, in this paper, the possibility to use the clustering method is verified. As the experimental result, the taxonomic overlap between model constructed with clustering and manually constructed reference model becomes almost 50%. Considering that we need one month to construct the real world activity model buy manual, this result indicates that the time required constructing the model will be reduced dramatically by using the clustering with manual way of constructing the model.