研究者詳細 - 駒水　孝裕

論文 - 駒水　孝裕

分割表示 100 件中 41 - 60 件目／全件表示 >>

Small Object Detection for Birds with Swin Transformer

Huo, D; Kastner, MA; Liu, T; Kawanishi, Y; Hirayama, T; Komamizu, T; Ide, I

2023 18TH INTERNATIONAL CONFERENCE ON MACHINE VISION AND APPLICATIONS, MVA 頁： 1 - 5 2023年

掲載種別：研究論文（国際会議プロシーディングス）出版者・発行元：Proceedings of MVA 2023 - 18th International Conference on Machine Vision and Applications

Object detection is the task of detecting objects in an image. In this task, the detection of small objects is particularly difficult. Other than the small size, it is also accompanied by difficulties due to blur, occlusion, and so on. Current small object detection methods are tailored to small and dense situations, such as pedestrians in a crowd or far objects in remote sensing scenarios. However, when the target object is small and sparse, there is a lack of objects available for training, making it more difficult to learn effective features. In this paper, we propose a specialized method for detecting a specific category of small objects; birds. Particularly, we improve the features learned by the neck; the sub-network between the backbone and the prediction head, to learn more effective features with a hierarchical design. We employ Swin Transformer to upsample the image features. Moreover, we change the shifted window size for adapting to small objects. Experiments show that the proposed Swin Transformer-based neck combined with CenterNet can lead to good performance by changing the window sizes. We further find that smaller window sizes (default 2) benefit mAPs for small object detection.

DOI： 10.23919/MVA57639.2023.10216093

その他リンク： https://dblp.uni-trier.de/db/conf/mva/mva2023.html#HuoKLKHKI23

Multi-Task Learning-based Text Classification with Subword-Phrase Extraction

Kimura Y., Komamizu T., Hatano K.

ACM International Conference Proceeding Series 頁： 23 - 30 2022年12月

掲載種別：研究論文（国際会議プロシーディングス）出版者・発行元：ACM International Conference Proceeding Series

Text classification using deep learning, which is trained with a tremendous amount of text, has achieved superior performance than traditional methods. In addition to its success, multi-Task learning has become a promising approach for text classification; for instance, a multi-Task learning approach employs named entity recognition as an auxiliary task for text classification. The existing MTL-based text classification methods depend on auxiliary tasks using supervised labels, which require large human and/or financial efforts to create. To reduce these efforts, this paper proposes a multi-Task learning-based text classification framework which reduces the additional efforts on supervised label creation. A basic idea to realize this is that to utilize phrasal expressions consisting of subwords (called subword-phrase). To the best of our knowledge, there has been no text classification approach on top of subword-phrases, because subwords do not always express a coherent set of meanings. The proposed framework is new to add subword-phrase recognition as an auxiliary task, and to utilize subword-phrases for text classification. To realize the low-cost auxiliary recognition task, the framework extracts subword-phrases in an unsupervised manner. The experimental evaluation of the five popular datasets for text classification showcases the effectiveness of the involvement of the subword-phrase recognition as an auxiliary task. It also shows comparative results with the state-of-The-Art method.

DOI： 10.1145/3568562.3568635

Detection of Birds in a 3D Environment Referring to Audio-Visual Information

Kawanishi, Y; Ide, I; Chu, B; Matsuhira, C; Kastner, MA; Komamizu, T; Deguchi, D

2022 18TH IEEE INTERNATIONAL CONFERENCE ON ADVANCED VIDEO AND SIGNAL BASED SURVEILLANCE (AVSS 2022) 頁： 1 - 7 2022年

掲載種別：研究論文（国際会議プロシーディングス）出版者・発行元：AVSS 2022 - 18th IEEE International Conference on Advanced Video and Signal-Based Surveillance

We propose a method to detect birds in a 3D environment referring to both audio information observed from a microphone array and visual information observed from a panorama camera. In general, in panorama images, birds appear relatively too small to be detected accurately even with the state-of-the-art deep learning models. Thus, the proposed method takes a two step approach where the birds are first roughly located referring to audio information by Sound Source Localization (SSL), and then image detection is applied within its vicinity. Through evaluation on a dataset annotated with bounding boxes surrounding the birds, we show that the proposed method improves detection performance of birds that appear in relatively small sizes in the image, in both accuracy and processing speed.

DOI： 10.1109/AVSS56176.2022.9959510

その他リンク： https://dblp.uni-trier.de/db/conf/avss/avss2022.html#KawanishiICMKKD22

Action Semantic Alignment for Image Captioning

Huo D., Kastner M.A., Komamizu T., Ide I.

Proceedings - 5th International Conference on Multimedia Information Processing and Retrieval, MIPR 2022 頁： 194 - 197 2022年

掲載種別：研究論文（国際会議プロシーディングス）出版者・発行元：Proceedings - 5th International Conference on Multimedia Information Processing and Retrieval, MIPR 2022

Image captioning is one of the main goals in vision and language processing, which aims to generate proper descriptions of images. Recently, the attention mechanisms became crucial in captioning tasks, as they can capture global dependencies between modalities. Moreover, some works have used objects detected from the input image as anchor points, so called object tags, to ease such alignments resulting in good performance for this task. In this paper, we newly introduce action information as a prior to further improve this, by adding action tags for training. The action tags can learn alignment at action semantic level and catch the previously ignored dimension of action, that could be very important in image captioning. We found that training with action tags can be used to describe images in a dynamic style. Furthermore, we found it can actually lead to a significant improvement compared with other methods in captioning performance measured by common metrics.

DOI： 10.1109/MIPR54900.2022.00041

その他リンク： https://dblp.uni-trier.de/db/conf/mipr/mipr2022.html#HuoKKI22

Towards Efficient Data Access Through Multiple Relationship in Graph-Structured Digital Archives

Kusu, K; Komamizu, T; Hatano, K

FROM BORN-PHYSICAL TO BORN-VIRTUAL: AUGMENTING INTELLIGENCE IN DIGITAL LIBRARIES, ICADL 2022 13636 巻頁： 377 - 391 2022年

掲載種別：論文集(書籍)内論文出版者・発行元：Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

The research field of digital libraries mainly deals with data with graph structure. Graph database management systems (GDBMSs) are suitable for managing data in the digital library because the data size is large and its structure is complex. However, when performing a non-simple search or analysis on a graph, GDBMSs cannot avoid reaching already-scanned nodes from different starting nodes by repeatedly traversing edges such as property paths pattern in SPARQL. Therefore, when a GDBMS reaches high degree nodes, the number of graph traversals increases in proportion to the number of its adjacent nodes. Consequently, the cost of traversing multiple paths extremely increases affected by nodes connected enormous the number of edges in conventional GDBMSs. In this paper, we propose a data access approach by repeatedly traversing edges belonging to a specific relationship or anything one while distinguishing between high degree nodes and low degree ones. Finally, a result of our experiment indicated our approach can increase the speed of repeat traversals by a factor of a maximum of ten.

DOI： 10.1007/978-3-031-21756-2_29

Intuitive Gait Modeling using Mimetic-Words for Gait Description and Generation

Kato H., Hirayama T., Doman K., Ide I., Kawanishi Y., Komamizu T., Deguchi D., Murase H.

Proceedings - 5th International Conference on Multimedia Information Processing and Retrieval, MIPR 2022 頁： 240 - 245 2022年

掲載種別：研究論文（国際会議プロシーディングス）出版者・発行元：Proceedings - 5th International Conference on Multimedia Information Processing and Retrieval, MIPR 2022

Gait is one of the most familiar action for us, that is why we can distinguish slight difference of human gaits and perceive their impressions. However, the relationship has been never explored because of the absence of intuitive labels for the slight differences. In this paper, to solve this problem, we propose a intuitive gait model using Japanese mimetic-words. A mimetic-word has sound-symbolism, which means that there is an association between linguistic sounds and sensory experiences, and the phonemes of a mimetic-word is strongly related to the visual sensation. Thanks to the sound-symbolism, Japanese mimetic-words have a possibility of modeling gaits intuitively. Thus, we have previously proposed a method which describes gait with a mimetic-word. In this paper, in the opposite direction, we propose a method which generates gait from a mimetic-word, and confirm the effectiveness of the proposed intuitive gait model which consists of the phonetic-vector through evaluations of both the generation task and the description task.

DOI： 10.1109/MIPR54900.2022.00050

その他リンク： https://dblp.uni-trier.de/db/conf/mipr/mipr2022.html#KatoHDIKKDM22

An Ensemble Framework of Multi-ratio Undersampling-based Imbalanced Classification 査読有り

駒水孝裕

Journal of Data Intelligence 2 (1) 巻頁： 30 - 46 2021年

FPX-G: First Person Exploration for Graph

Komamizu T., Ito S., Ogawa Y., Toyama K.

Proceedings - 4th International Conference on Multimedia Information Processing and Retrieval, MIPR 2021 頁： 70 - 76 2021年

掲載種別：研究論文（国際会議プロシーディングス）出版者・発行元：Proceedings - 4th International Conference on Multimedia Information Processing and Retrieval, MIPR 2021

Data exploration is a fundamental user task in the information seeking process. In data exploration, users have ambiguous information needs, and they traverse across the data for gathering information. In this paper, a novel data exploration system, called FPX-G, is proposed that uses virtual reality (VR) technology. VR-based data exploration (or immersive analytics) is a recent trend in data analytics, and the existing work approaches involve aggregated information in an interactive and 3D manner. However, exploration for individual pieces of data scarcely has been approached. Traditional data exploration is done on 2D displays, therefore space is limited, and there is no depth. FPX-G fully utilizes 3D space to make individual piece of data visible in the user’s line of sight. In this paper, the data structure in FPX-G is designed as a graph, and the data exploration process is modeled as graph traversal. To utilize the capability of VR, FPX-G provides a first person view-based interface from which users can look at individual pieces of data and can walk through the data (like walking in a library). In addition to the walking mechanism, to deal with limited physical space in a room, FPX-G introduces eye-tracking technology for traversing data through a graph. A simulation-based evaluation reveals that FPX-G provides a significantly efficient interface for exploring data compared with the traditional 2D interface.

DOI： 10.1109/MIPR51284.2021.00018

その他リンク： https://dblp.uni-trier.de/db/conf/mipr/mipr2021.html#KomamizuIOT21

Evaluation Scheme of Focal Translation for Japanese Partially Amended Statutes

Yamakoshi T., Komamizu T., Ogawa Y., Toyama K.

WAT 2021 - 8th Workshop on Asian Translation, Proceedings of the Workshop 頁： 124 - 132 2021年

出版者・発行元：WAT 2021 - 8th Workshop on Asian Translation, Proceedings of the Workshop

For updating the translations of Japanese statutes based on their amendments, we need to consider the translation “focality;” that is, we should only modify expressions that are relevant to the amendment and retain the others to avoid misconstruing its contents. In this paper, we introduce an evaluation metric and a corpus to improve focality evaluations. Our metric is called an Inclusive Score for DIfferential Translation: (ISDIT). ISDIT consists of two factors: (1) the n-gram recall of expressions unaffected by the amendment and (2) the n-gram precision of the output compared to the reference. This metric supersedes an existing one for focality by simultaneously calculating the translation quality of the changed expressions in addition to that of the unchanged expressions. We also newly compile a corpus for Japanese partially amendment translation that secures the focality of the post-amendment translations, while an existing evaluation corpus does not. With the metric and the corpus, we examine the performance of existing translation methods for Japanese partially amendment translations.

Combining Multi-ratio Undersampling and Metric Learning for Imbalanced Classification.

Takahiro Komamizu

Journal of Data Intelligence 2 巻 ( 4 ) 頁： 462 - 474 2021年

Random walk-based entity representation learning and re-ranking for entity search

Komamizu, T

KNOWLEDGE AND INFORMATION SYSTEMS 62 巻 ( 8 ) 頁： 2989 - 3013 2020年8月

掲載種別：研究論文（学術雑誌）出版者・発行元：Knowledge and Information Systems

Linked Data (LD) has become a valuable source of factual records, and entity search is a fundamental task in LD. The task is, given a query consisting of a set of keywords, to retrieve a set of relevant entities in LD. The state-of-the-art approaches for entity search are based on information retrieval techniques. This paper first examines these approaches with a traditional evaluation metric, recall@k, to reveal their potential for improvement. To obtain evidence for the potentials, an investigation is carried out on the relationship between queries and answer entities in terms of path lengths on a graph of LD. On the basis of the investigation, learning representations of entities are dealt with. The existing methods of entity search are based on heuristics that determine relevant fields (i.e., predicates and related entities) to constitute entity representations. Since the heuristics require burdensome human decisions, this paper is aimed at removing the burden with a graph proximity measurement. To this end, in this paper, RWRDoc is proposed. It is an RWR (random walk with restart)-based representation learning method that learns representations of entities by using weighted combinations of representations of reachable entities w.r.t. RWR. RWRDoc is mainly designed to improve recall scores; therefore, as shown in experiments, it lacks capability in ranking. In order to improve the ranking qualities, this paper proposes a personalized PageRank-based re-ranking method, PPRSD (Personalized PageRank-based Score Distribution), for the retrieved results. PPRSD distributes relevance scores calculated by text-based entity search methods in a personalized PageRank manner. Experimental evaluations showcase that RWRDoc can improve search qualities in terms of recall@1000 and PPRSD can compensate for RWRDoc’s insufficient ranking capability, and the evaluations confirmed this compensation.

DOI： 10.1007/s10115-020-01445-4

ランダムフォレストを用いた法令用語の校正

山腰貴大, 小川泰弘, 駒水孝裕, 外山勝彦

人工知能学会論文誌 35 巻 ( 1 ) 頁： H-J53_1 - 14 2020年1月

記述言語：日本語掲載種別：研究論文（学術雑誌）出版者・発行元：一般社団法人人工知能学会

We propose a method that assists legislation drafters in finding inappropriate use of Japanese legal terms and their corrections from Japanese statutory sentences. In particular, we focus on sets of similar legal terms whose usages are strictly defined in legislation drafting rules that have been established over the years. In this paper, we first define input and output of legal term correction task. We regard it as a special case of sentence completion test with multiple choices. Next, we describe a legal term correction method for Japanese statutory sentences. Our method predicts suitable legal terms using Random Forest classifiers. The classifiers in our method use adjacent words to a target legal term as input features, and are optimized in various parameters including the number of adjacent words to be used for each legal term set. We conduct an experiment using actual statutory sentences from 3,983 existing acts and cabinet orders that consist of approximately 47M words in total. As for legal term sets, we pick 27 sets from legislation drafting manuals. The experimental result shows that our method outperformed existing modern word prediction methods using neural language models and that each Random Forest classifier utilizes characteristics of its corresponding legal term set.

DOI： 10.1527/tjsai.h-j53

事前学習モデルBERTによる法令用語の校正

山腰貴大, 駒水孝裕, 小川泰弘, 外山勝彦

人工知能学会全国大会論文集 2020 巻 ( 0 ) 頁： 4P3OS805 - 4P3OS805 2020年

出版者・発行元：一般社団法人人工知能学会

<p>法令文書には，「者」「物」「もの」や「規定」「規程」のように互いに類似している法令用語が出現する．このような法令用語は，法制執務（法令の起草・制定・改廃など法令文書の作成・管理に関する業務）の慣習や規則によって使用法とともに定義されている．法令において，これらの法令用語はそれに従い，厳密に書き分ける必要がある．契約書や約款などの広義の法令文書においても，誤解を防ぐために，法令に準じて正しく書き分けることが望ましい．そこで，本研究では，与えられた法令文から法令用語を検出し，誤用と思われるものに対してその修正案を出力することにより，法令文書の作成を支援する手法を提案する．本手法では，このタスクを選択肢付き穴埋め問題とみなし，分類器により解決する．分類器は，一般文によって事前学習したBERTモデルから構築する．このとき，（１）法令文によるドメイン適応，（２）訓練データのアンダーサンプリング，（３）分類器の統一の三つの工夫を施すことにより性能向上を図る．実験の結果，ランダムフォレストやニューラル言語モデルによる分類器よりも本手法の方が高い性能を発揮することを明らかにした．</p>

DOI： 10.11517/pjsai.JSAI2020.0_4P3OS805

Japanese mistakable legal term correction using infrequency-aware bert classifier

Yamakoshi T., Komamizu T., Ogawa Y., Toyama K.

Transactions of the Japanese Society for Artificial Intelligence 35 巻 ( 4 ) 頁： 1 - 17 2020年

記述言語：英語出版者・発行元：Transactions of the Japanese Society for Artificial Intelligence

We propose a method to assist legislative drafters that locates inappropriate legal terms in Japanese statutory sentences and suggests corrections. We focus on sets of mistakable legal terms whose usages are defined in legislation drafting rules. Our method predicts suitable legal terms using a classifier based on BERT (Bidirectional Encoder Representations from Transformers). The BERT classifier is pretrained with a huge number of whole sentences; thus, it contains abundant linguistic knowledge. Classifiers for predicting legal terms suffer from two-level infrequency: term-level infrequency and set-level infrequency. The former causes a class imbalance problem and the latter causes an underfitting problem; both degrade classification performance. To overcome these problems, we apply three techniques, namely, preliminary domain adaptation, repetitive soft undersampling, and classifier unification. The preliminary domain adaptation improves overall performance by providing prior knowledge of statutory sentences, the repetitive soft undersampling overcomes term-level infrequency, and the classifier unification overcomes set-level infrequency while saving storage consumption. Our experiments show that our classifier outperforms conventional classifiers using Random Forest or language models, and that all three training techniques improve performance.

DOI： 10.1527/tjsai.E-K25

Exploring Relevant Parts Between Legal Documents Using Substructure Matching 査読有り

Komamizu, T; Fujioka, K; Ogawa, Y; Toyama, K

NEW FRONTIERS IN ARTIFICIAL INTELLIGENCE, JSAI-ISAI 2019 12331 巻頁： 5 - 19 2020年

記述言語：英語掲載種別：研究論文（その他学術会議資料等）出版者・発行元：Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

Legal documents are typically hierarchically structured. This paper focuses on ordinances and rules (OR documents for short) in the local governments, which are designed for social lives under the governments. OR documents are composed of provisions for social lives in various aspects such as healthy development of youths and landscape preservation. OR documents in different local governments share common provisions but also include different provisions depending on their social situations. There is a large demand on helping governmental officers draft OR documents, especially searching “relevant parts” of OR documents. To help drafting OR documents, this paper designs the relevancy of OR documents with two basic measurements; matching ratio and provision commonality. Based on the relevancy, this paper develops a structured document search algorithm for OR documents. Experimental evaluation on real OR documents in Japan demonstrates that the proposed algorithm successfully discovers relevant parts of OR documents.

DOI： 10.1007/978-3-030-58790-1_1

その他リンク： https://dblp.uni-trier.de/db/conf/jsai/jsai2019w.html#KomamizuFOT19

SPARQL with XQuery-based Filtering.

Takahiro Komamizu

CoRR abs/2009.06194 巻 2020年

SPARQL with XQuery-based filtering

Komamizu T.

CEUR Workshop Proceedings 2721 巻頁： 69 - 73 2020年

出版者・発行元：CEUR Workshop Proceedings

Linked Open Data (LOD) has been proliferated over various domains, however, there are still lots of open data in various format other than RDF. Document-centric XML data are such open data that are connected with entities in LOD as supplemental documents for these entities. To utilize document-centric XML data linked from entities in LOD, in this paper, a SPARQL-based seamless access method on RDF and XML data is proposed. In particular, an extension to SPARQL, XQueryFILTER, which enables XQuery as a filter in SPARQL is proposed. For efficient query processing of the combination of SPARQL and XQuery, a query optimization is proposed. Experimental scenarios using real-world data showcase the effectiveness of XQueryFILTER and optimization efficiency.

MUEnsemble: Multi-ratio Undersampling-Based Ensemble Framework for Imbalanced Data

Komamizu, T; Uehara, R; Ogawa, Y; Toyama, K

DATABASE AND EXPERT SYSTEMS APPLICATIONS, DEXA 2020, PT II 12392 巻頁： 213 - 228 2020年

掲載種別：研究論文（国際会議プロシーディングス）出版者・発行元：Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

Class imbalance is commonly observed in real-world data, and it is still problematic in that it hurts classification performance due to biased supervision. Undersampling is one of the effective approaches to the class imbalance. The conventional undersampling-based approaches involve a single fixed sampling ratio. However, different sampling ratios have different preferences toward classes. In this paper, an undersampling-based ensemble framework, MUEnsemble, is proposed. This framework involves weak classifiers of different sampling ratios, and it allows for a flexible design for weighting weak classifiers in different sampling ratios. To demonstrate the principle of the design, in this paper, three quadratic weighting functions and a Gaussian weighting function are presented. To reduce the effort required by users in setting parameters, a grid search-based parameter estimation automates the parameter tuning. An experimental evaluation shows that MUEnsemble outperforms undersampling-based methods and oversampling-based state-of-the-art methods. Also, the evaluation showcases that the Gaussian weighting function is superior to the fundamental weighting functions. In addition, the parameter estimation predicted near-optimal parameters, and MUEnsemble with the estimated parameters outperforms the state-of-the-art methods.

DOI： 10.1007/978-3-030-59051-2_14

その他リンク： https://dblp.uni-trier.de/db/conf/dexa/dexa2020-2.html#KomamizuUOT20

Japanese Mistakable Legal Term Correction using Infrequency-aware BERT Classifier" 査読有り

Takahiro Yamakoshi, Takahiro Komamizu, Yasuhiro Ogawa, Katsuhiko Toyama

the 3rd Annual Workshop on Applications of Artificial Intelligence in the Legal Industry (LegalAI 2019) 頁： 4342-4351 2019年12月

Japanese Mistakable Legal Term Correction using Infrequency-aware BERT Classifier Matching 査読有り

Takahiro Yamakoshi, Takahiro Komamizu, Yasuhiro Ogawa, Katsuhiko Toyama

Proc. 3rd Annual Workshop on Applications of Artificial Intelligence in the Legal Industry 頁： - - 4351 2019年12月

記述言語：英語掲載種別：研究論文（その他学術会議資料等）出版者・発行元：Proceedings - 2019 IEEE International Conference on Big Data, Big Data 2019

We propose a method that assists legislative drafters in locating inappropriate legal terms in Japanese statutory sentences and suggests corrections. We focus on sets of mistakable legal terms whose usages are defined in legislation drafting rules. Our method predicts suitable legal terms using a classifier based on a BERT (Bidirectional Encoder Representations from Transformers) model. We apply three techniques in training the BERT classifier, specifically, preliminary domain adaptation, repetitive soft undersampling, and classifier unification. These techniques cope with two levels of infrequency: legal term-level infrequency that causes class imbalance and legal term set-level infrequency that causes underfitting. Concretely, preliminary domain adaptation improves overall performance by providing prior knowledge of statutory sentences, repetitive soft undersampling improves performance on infrequent legal terms without sacrificing performance on frequent legal terms, and classifier unification improves performance on infrequent legal term sets by sharing common knowledge among legal term sets. Our experiments show that our classifier outperforms conventional classifiers using Random Forest or a language model, and that all three training techniques contribute to performance improvement.

DOI： 10.1109/BigData47090.2019.9006511

その他リンク： https://dblp.uni-trier.de/db/conf/bigdataconf/bigdataconf2019.html#YamakoshiKOT19

　前のページ - 次のページ　

1
2
3
4
5