研究者詳細 - 嶋田　和孝

論文 - 嶋田　和孝

分割表示146 件中 21 - 40 件目／全件表示 >>

Dataset Construction and Opinion Holder Detection Using Pre-trained Models 査読有り国際誌

Mahmud Al-, Shimada Kazutaka

International Journal of Service and Knowledge Management ( 特定非営利活動法人国際応用情報学研究機構 ) 7 ( 2 ) 1 2023年01月

Rating Prediction of Multi-Aspect Reviews Using Simultaneous Learning 査読有り

Takeo M., Kawasaki S., Shimada K.

Proceedings of 2023 International Conference on Asian Language Processing, IALP 2023 358 - 363 2023年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

This paper proposes a method for rating prediction tasks in review documents. Existing models for predicting the rating scores assigned to each aspect were often developed separately. However, there is a relationship among aspects. Therefore, we apply a multi-task learning framework to our prediction models. We evaluate our model with a review dataset about a game software domain written in Japanese. Each review document contains six-level rating scores for seven aspects of the game software. We utilize BERT as the backbone model for the prediction. Our model simultaneously learns the parameters of seven BERTs for seven aspects. Hence, we name this method "Simultaneous Learning."It leads to the improvement of the performance of the prediction. In addition, we compare two types of input data for the rating prediction task: all sentences and selected sentences. Experimental results show the effectiveness of our simultaneous learning model in the multi-aspect rating prediction. Our simultaneous learning method obtained 0.713 on RMSE. It demonstrated a 0.046 improvement in RMSE, as compared with non-simultaneous learning. Further, we confirmed that our method based on simultaneous learning was effective in the case that the number of sentences related to an aspect is small.

DOI： 10.1109/IALP61005.2023.10337276

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85181775477&origin=inward

Multilingual Symptom Prediction by Simultaneous Learning Using BERT 査読有り

Zhu C., Pahari N., Shimada K.

Proceedings of 2023 International Conference on Asian Language Processing, IALP 2023 100 - 105 2023年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Natural language processing (NLP) techniques have become increasingly significant in the medical domain. However, the amount of relevant medical text data remains limited. In this work, we propose a BERT-based multilingual simultaneous learning (MSL) model for reducing the problem of scarce data. We evaluate the benefit of MSL on the NTCIR-13 MedWeb multi- label symptoms classification task. The results indicate that the MSL model performs slightly better than Single-Task Learning (STL) models. Additionally, it shows that the similarity between languages has an impact on the performance of the MSL model.

DOI： 10.1109/IALP61005.2023.10337285

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85181777849&origin=inward

Language Preference for Expression of Sentiment for Nepali-English Bilingual Speakers on Social Media 査読有り

Pahari N., Shimada K.

CALCS 2023 - Computational Approaches to Linguistic Code-Switching, Proceedings of the Workshop 23 - 32 2023年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Nepali-English code-switching (CS) has been a growing phenomenon in Nepalese society, especially in social media. The code-switching text can be leveraged to understand the sociolinguistic behaviours of the multilingual speakers. Existing studies have attempted to identify the language preference of the multilingual speakers for expressing different emotions using text in different language pairs. In this work, we aim to study the language preference of multilingual Nepali-English CS speakers while expressing sentiment in social media. We create a novel dataset for sentiment analysis using the public Nepali-English code-switched comments in YouTube. After performing the statistical study on the dataset, we find that the proportion of use of Nepali language is higher in negative comments when compared with positive comments, hence concluding the preference for using native language while expressing negative sentiment. Machine learning and transformer-based models are used as the baseline models for the dataset for sentiment classification. The dataset is released publicly.

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85185000192&origin=inward

Handling Data Scarcity in Code-Mixing Using Learning-Based Approaches 査読有り

Pahari N., Shimada K.

Proceedings - 2023 14th IIAI International Congress on Advanced Applied Informatics, IIAI-AAI 2023 7 - 12 2023年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

With the rise of multilingual populations, the use of structures and vocabularies from multiple languages in a single communication act is increasing. The act of communication can be written or verbal. This phenomenon of mixing components from multiple languages is called code-mixing or code-switching. The penetration of social networks has increased the occurrence of code-mixing. Even with the abundance of data, the scarcity of annotated data has limited the research on code-mixing in the natural language processing (NLP) community. Here, we aim to analyze learning-based approaches to overcome the scarcity of annotated data by fine-tuning various multilingual transformer-based pre-trained models by conducting experiments on Nepaliand Hindi-English language identification tasks. Multitask learning, which can learn from the related task and improve the main task, and self-supervised learning, which utilizes the unannotated data in the same domain, are studied in this paper. The results demonstrate that these approaches are beneficial in improving the performance of the transformer-based models for the scarcity of annotated data.

DOI： 10.1109/IIAI-AAI59060.2023.00012

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85183462866&origin=inward

Disfluency detection with context information from real utterances and generative utterances 査読有り

Nakashima H., Shimada K.

Proceedings - 2023 14th IIAI International Congress on Advanced Applied Informatics, IIAI-AAI 2023 462 - 467 2023年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Disfluency is one of the linguistic phenomenons that often occurs in speech. Recently, automatic speech recognition has improved the accuracy of its transcript. The perfect transcript contains filler, self-repairs, distluency, and so on, because they often occurs. We focus on distluency in speech recognition outputs. The distluency in the transcript can be noise for the downstream tasks that use the transcript. In this paper, we propose a BERT-based distluency detection model for dialogue utterance. BERT handles contextual information of a target utterance to detect distluency. However, dialogue utterances can be interrupted by other speakers in conversation. Hence, utterances in dialogue tend to be short and have less contextual information. Therefore, the context of the target utterances should be complemented. In this paper, we propose two types of complementing approaches. The complement is preprocessing for the input of BERT. The first complement approach utilizes real utterances around the target utterance. In this approach, we need to select sentences from real utterances, based on some conditions: e.g., the number of utterances and the same speaker or not. The second complement approach utilizes utterances generated by a language model. The language model is trained to generate continuous utterances from the input. The generated utterance is utilized for expanding the context of the input data. In experiments, utterances from the speaker of target utterances improve the distluency detection performance. However, generated context could not outperform the real context.

DOI： 10.1109/IIAI-AAI59060.2023.00095

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85183462855&origin=inward

Detecting speech recognition errors using topic information and BERT 査読有り

Yokoyama M., Shimada K.

Proceedings - 2023 14th IIAI International Congress on Advanced Applied Informatics, IIAI-AAI 2023 400 - 405 2023年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Detecting errors in speech recognition is important because these errors negatively affect the accuracy of downstream tasks. To detect recognition errors, we focus on context, label sequence, and topics. Words in a sentence contain the context. Context information is useful in detecting errors from speech recognition because the errors must be outside the scope of the context. Furthermore, recognition error detection is a sequential problem, namely a sequential labeling task. Therefore, handling a label sequence is useful for error detection. In addition, a target speech contains topics, such as a talk about music, sports, and so on. A word that does not match the topic is considered an error. Therefore, the topic information is also useful. In this paper, we propose a method based on BERT, CRF, and LDA from the viewpoints BERT is used for context recognition. Conditional Random Fields (CRF) are used for the sequence labeling task. Latent Dirichlet Allocation (LDA) is to capture topic information. In this study, we compared several combinations of them: BERT only, BERT-CRF, and BERT-LDA-CRF models. We evaluate the models with the Corpus of Spontaneous Japanese (CSJ) corpus. The experimental result shows the effectiveness of the integration of CRF and LDA into BERT.

DOI： 10.1109/IIAI-AAI59060.2023.00084

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85183463970&origin=inward

Combination and Knowledge Extension of Pre-trained Language Model for Offensive Language Detection 査読有り

Li Z., Shimada K.

Proceedings - 2023 14th IIAI International Congress on Advanced Applied Informatics, IIAI-AAI 2023 82 - 87 2023年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Nowadays, more and more offensive comments are posted on social media. Those offensive comments can seriously cause mental damage to other people. Therefore, those toxic and harmful comments should be detected timely and accurate. In this paper, we exploit the powerful pre-train language models (PLM) in detecting offensive language. We consider two PLMs in our paper: BERT and DeepMoji. We finetune their combination and evaluate the performance with two datasets: Ask FM and Curious Cat. We found that DeepMoji outperforms BERT. We analyze the result from two aspects and conclude that the task and data of pre-training are important to PLM. Seeing that BERT is less effective than DeepMoji, it is possible to improve the performance of BERT. We then propose a 'Knowledge Extension' method to improve the performance of the BERT model. We find that the performance of PLM with high-quality extensional knowledge can be improved significantly.

DOI： 10.1109/IIAI-AAI59060.2023.00026

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85183459713&origin=inward

Demonstration of Effectiveness of Nativeness in Stepwise Learning by Performing Sentiment Analysis 査読有り国際誌

Al-Mahmud , Shimada K.

2023 International Conference on Next-Generation Computing, IoT and Machine Learning, NCIM 2023 2023年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Nowadays, more people and entities express opinions on online platforms. With this increasing trend of online opinion mass information, it is impossible to analyze them manually. For this reason, an automatic approach to opinion and sentiment analysis is essential and useful in the real world. Several approaches based on machine learning techniques in sentiment analysis have been proposed. However, machine learning essentially needs massive data to generate a robust model. On the other hand, collecting the training data is costly, especially in low-resource language. To solve this problem, we introduce stepwise learning. Stepwise learning is a technique that improves the model capacity for a target task (i.e., main task) by exploiting auxiliary information from source tasks. It is effective when the data in the target tasks are inadequate. The target task of this paper is sentiment analysis for reviews in the Bangla language. Our proposed method consists of two stages: source and target tasks. For the source task, we train a transformers-based model, such as BanglaBERT, by a large dataset translated into the Bangla language; the original dataset is written in English. For the target task, we retrain and evaluate the model with a relatively small target dataset. We further investigate the effect of opinion expression style of writers between source and target data in stepwise learning. We call it 'nativeness.' The experimental result shows that nativeness has an important role in improving performance.

DOI： 10.1109/NCIM59001.2023.10212970

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85170528820&origin=inward

Combining Pre-Trained Language Models and Features for Offensive Language Detection 査読有り国際誌

Li Z., Shimada K.

Proceedings - 2022 13th International Congress on Advanced Applied Informatics Winter, IIAI-AAI-Winter 2022 5 - 10 2022年12月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Nowadays, people often express their abusive and offensive thoughts to others on social media easier. The abusive and toxic comments hurt others seriously. Therefore those abusive and toxic comments should be detected properly through natural language processing. In this paper, we focus on two types of features in offensive language: word-level and sentence-level fea-tures. We use lexicon-based and standard bag-of-words features as the word level. We introduce BERT-based and DeepMoji-based features as the sentence level. We apply the four features to a machine learning approach: support vector machines. We evaluate the method using the combinations of four features with a dataset, Curious Cat. The best F1 score was generated by the method with all features. This result shows the effectiveness of our proposed method. In addition, the experimental result indicates that DeepMoji generated from Twitter data is better than BERT which is generated from written language, for an offensive language detection task about social media data.

DOI： 10.1109/IIAI-AAI-Winter58034.2022.00012

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85160906880&origin=inward

Annotation and multi-modal methods for quality assessment of multi-party discussion 査読有り国際誌

Tsukasa Shiota, Kazutaka Shimada

Proceedings of the 36th Pacific Asia Conference on Language, Information and Computation 175 - 182 2022年10月

Multi-Task Learning Using BERT With Soft Parameter Sharing Between Layers 査読有り国際誌

Pahari N., Shimada K.

2022 Joint 12th International Conference on Soft Computing and Intelligent Systems and 23rd International Symposium on Advanced Intelligent Systems, SCIS and ISIS 2022 2022年10月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

The ability to learn from one task and pass the knowledge to another task is a remarkable job that is possible because of multi-task learning (MTL). It is helpful in the scenario where data is scarce. In this paper, we study the use of MTL in natural language processing (NLP) using BERT for sentiment classification, category classification, and aspect-opinion sequence classification. The Restaurant-ACOS dataset is split into two subsets, namely the sentiment-category data subset, for the first two tasks and the sequence classification data subset, for the third task. The BERT layers are grouped into top, middle, and bottom layers. These layers are either frozen, shared, or kept unshared for individual tasks. Experiments are conducted to find the best combination for the overall layers of BERT. The experiments demonstrate that MTL can improve the performance on these tasks when compared as a baseline method that uses single task learning. One of the best configurations was based on freezing on the bottom layers, sharing on the middle layers, and keeping on the top layers, respectively. In addition, we investigate the role of each layer via the size of layers. From the result, the individual layers capture the task-specific patterns whereas shared layers capture the domain knowledge.

DOI： 10.1109/SCISISIS55246.2022.10001943

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85146643824&origin=inward

Dataset Construction and Classification Based on Pre-trained Models for Opinion Holder Detection 査読有り

Al-Mahmud A., Shimada K.

Proceedings - 2022 12th International Congress on Advanced Applied Informatics, IIAI-AAI 2022 65 - 70 2022年07月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Nowadays, it is getting increased in the massive amount of internet users. People express subjective thinking (i.e., opinion) implicitly and explicitly on online platforms such as Facebook, Twitter, Amazon product reviews, etc. Opinion holders are the people or entities who express opinions implicitly and explicitly on the online platform. With the increasing trends of online opinion mass information, it is impossible to detect them manually. For this reason, an automatic approach for opinion holder detection is essential. Opinion holder detection is useful to detect specific person's/entity's concerns about a particular topic, product, or problem. Opinion holder detection consists of two steps: the presence of opinion holders in text and identification of opinion holders. In this paper, we focus on the first step, namely the presence of opinion holders in text. We handle this task as a binary classification problem: INSIDE or OUTSIDE. At first, we prepare a new English dataset for this task. Then, we apply two types of pre-trained models, BERT and DistilBERT, to the INSIDE/OUTSIDE classification task. BERT is a transformer-based pre-trained language model. DistilBERT is a small, fast, cheap, and light model based on knowledge distillation from the BERT architecture. As to the binary classification task, we employ a logistic regression model on the top layer of the pre-trained models. We compare the language models employed in our experiment in terms of the F1 score and accuracy. The experimental result shows that DistilBERT obtained superior performance among the models: 0.901 on the F1 score and 0.924 on the accuracy.

DOI： 10.1109/IIAIAAI55812.2022.00023

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85139566304&origin=inward

Pseudo data acquisition using machine translation and simile identification 査読有り

Jimi J., Shimada K.

Proceedings - 2022 12th International Congress on Advanced Applied Informatics, IIAI-AAI 2022 391 - 396 2022年07月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

The simile is a kind of figurative language. It expresses the target of the figurative language by using some typical phrases such as "like". It is important to distinguish whether the sentence is a simile or a literal for understanding a sentence. However, a large amount of data is required to generate a classifier by machine learning. Moreover, creating the dataset is costly. In this paper, we propose a pseudo dataset acquisition method for simile identification. We first construct a dataset of simile and literal sentences using machine translation. This process automatically generates pseudo simile and literal instances from three types of corpora. Then, we apply some machine learning approaches to the simile identification task. We compare Support Vector Machine, Naive Bayes, and BERT in the experiment. The experimental result shows the validity of the pseudo dataset as compared with a simple baseline. For the fine-tuning of BERT, our large pseudo training data were more effective than a small manual training data.

DOI： 10.1109/IIAIAAI55812.2022.00084

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85139568402&origin=inward

A Compression-Based Multiple Subword Segmentation for Neural Machine Translation 査読有り国際誌

Keita Nonaka, Kazutaka Yamanouchi, Tomohiro I, Tsuyoshi Okita, Kazutaka Shimada, Hiroshi Sakamoto

Electronics 11 ( 7 ) 2022年03月

記述言語：英語掲載種別：研究論文（学術雑誌）

In this study, we propose a simple and effective preprocessing method for subword segmentation based on a data compression algorithm. Compression-based subword segmentation has recently attracted significant attention as a preprocessing method for training data in neural machine translation. Among them, BPE/BPE-dropout is one of the fastest and most effective methods compared to conventional approaches; however, compression-based approaches have a drawback in that generating multiple segmentations is difficult due to the determinism. To overcome this difficulty, we focus on a stochastic string algorithm, called locally consistent parsing (LCP), that has been applied to achieve optimum compression. Employing the stochastic parsing mechanism of LCP, we propose LCP-dropout for multiple subword segmentation that improves BPE/BPE-dropout, and we show that it outperforms various baselines in learning from especially small training data.

DOI： 10.3390/electronics11071014

Scopus

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85126910361&origin=inward

テキストアナリティクスとその先へ

嶋田和孝

情報・システムソサイエティ誌 ( 一般社団法人電子情報通信学会 ) 26 ( 4 ) 4 - 5 2022年02月

Recognizing a participant behavior in a multi-party conversation 査読有り

Shimada Kazutaka, Yonemitsu Shunsuke

Information Engineering Express ( 特定非営利活動法人国際応用情報学研究機構 ) 8 ( 1 ) 1 2022年01月

Corpus construction for topic-based summarization of multi-party conversation 査読有り国際誌

Yuri Nakayama, Tsukasa Shiota, Kazutaka Shimada

Proceedings of the International Conference on Asian Language Processing (IALP) 229 - 234 2021年12月

Sarcasm Detection with External Entity Information 査読有り

Xufei Xu, Kazutaka Shimada

Proceedings of the 10th International Conference on Smart Computing and Artificial Intelligence 2021年12月

Discussion Structure Prediction Based on Two-step Method 査読有り国際誌

Takumi Himeno, Kazutaka Shimada

Proceedings of Recent Advances in Natural Language Processing, RANLP 2021 543 - 551 2021年09月

　前のページ - 次のページ　

<嶋田 和孝>

論文 - 嶋田 和孝

<嶋田　和孝>

論文 - 嶋田　和孝