81690 research outputs found
Sort by
人口減少時代における中国農村地域と宗教団体 : ソーシャル・キャピタルの視点から [全文の要約]
No full textこの博士論文全文の閲覧方法については、以下のサイトをご参照ください。https://www.lib.hokudai.ac.jp/dissertations/copy-guides
Study on propaganda detection in Polish online news
No full textIn recent years, the quick and easy access to online news sources has significantly influenced public opinion, especially through biased reporting and the spread of misinformation. This trend has been particularly problematic in under-resourced languages, where access to reliable information is limited, and detecting propaganda is more challenging. This study is a step-by-step attempt to address these issues by analyzing sentiment with different approaches and leveraging Large Language Models (LLMs) for the detection of political and establishment biases, as well as propaganda in Polish online news articles.
The first phase of this research focused on creation of a dataset called Polish Online News Corpus (PONC), which includes more than 200,000 articles from two major Polish news providers, TVP Info and TVN24, published between 2019 and 2021 and was later utilized in all experiments. The PONC dataset was constructed through a multi-stage process with the use of BeautifulSoup library to ensure quality and reproducibility. The classification experiment aimed to determine whether the source of a news article could be predicted based on textual features, indirectly assessing stylistic and content differences between TVP Info and TVN24. The following machine learning models were applied: Logistic Regression, Random Forest, Support Vector Machine (SVM), Naive Bayes Classifier. Each model was tested using two feature extraction techniques: Bag of Words (BoW) and TF-IDF. The best performing model was the Random Forest classifier with TF-IDF vectorization, achieving an accuracy of 87.45% and an F1-score of 88.29%. These results confirmed that textual features alone could distinguish between the two news outlets, indicating underlying differences in language use and content framing.
The following task was an analysis of the emotional load in these articles using both lexicon-based and statistical approaches to sentiment analysis. The study aimed to determine whether supposedly objective online news contains emotive content and how different methodologies affect the detection of such biases. Initial experiments involved lexicon-based sentiment analysis using the Nencki Affective Word List (NAWL), a Polish affective lexicon assigning scores across six emotional categories: anger, disgust, fear, happiness, sadness, and neutrality, and zero-shot classification with the XLM-RoBERTa model fine-tuned on the XNLI dataset. The lexicon-based analysis revealed no statistically significant differences in emotional content between the two news outlets, suggesting a limited sensitivity to contextual sentiment. However, the XLM-RoBERTa model provided a more nuanced detection of emotive content, highlighting notable disparities between the sources. Specifically, TVP Info articles exhibited eight times more words classified as disgust compared to NAWL’s baseline, while words associated with fear and happiness were reduced by eleven and five times, respectively. Topic-specific emotional patterns were also observed. For Abortion, articles from TVN24 contained three times more words classified as happy compared to TVP Info. For Church, TVP Info articles displayed twice as many words classified as sad and half as many words classified as happy as TVN24. These results demonstrate that while lexicon-based methods may underrepresent emotional variation, statistical models like XLM-RoBERTa can capture more subtle differences in emotive content, particularly in sensitive topics.
The study also highlighted the challenges and limitations associated with using LLMs for political and establishment bias detection, including the models’ inability to fully understand non-literal language. Despite the challenges, the research demonstrated the potential of LLMs to automate the annotation process, thus saving time and resources compared to manual annotation. Subsequent analyses showed that LLMs, such as gpt-3.5-turbo and gpt-4, could annotate political leanings and establishment biases, albeit with varying degrees of accuracy. Establishment bias proved more challenging to detect, with accuracy rates 10–12% lower compared to political bias detection tasks. This result indicates that political alignment (e.g., left- or right-wing tendencies) was easier for the models to identify than stance towards the ruling government or institutions. The models struggled with non-literal language and subtle political messaging, especially when bias was embedded in metaphors, irony, or indirect references. For instance, phrases with implicit sarcasm or historical references led to inconsistent classifications.
In addition to detecting biases, the study explored the potential of LLMs to identify propaganda techniques used in Polish news articles. The experiments on propaganda detection using Large Language Models (LLMs), specifically gpt-3.5-turbo and gpt-4, focused on analyzing articles from the Polish Online News Corpus (PONC) to identify rhetorical strategies aimed at influencing public opinion. The analysis was conducted on high-emotional charge articles, selected for their controversial topics where propaganda techniques are more likely to appear. gpt-4 achieved a classification accuracy of 74% for Binary Propaganda Detection task and 69% for Technique Classification on PONC subset. Most commonly detected techniques included Appeal to Fear/Prejudice and Loaded Language, whereas Name-calling, Slogans, and Bandwagon were less frequently. More nuanced propaganda techniques such as repetition and whataboutism were rarely detected, suggesting a gap in the models’ capacity to identify subtle manipulative patterns. Additionally, despite controlled prompts and temperature settings, variation in results was noted across trials. While gpt-4 showed promising results in detecting propaganda techniques like loaded language, its ability to identify subtle rhetorical techniques remained limited. These findings emphasize the importance of using LLMs as assistive tools for propaganda detection, but currently not able to fully replace human evaluators.
Finally, it is clarified that utilizing LLMs is a promising approach for detecting propaganda or political and establishment biases in online news, but their effectiveness varies depending on the specific task and model used. It is also noted that further research and improvements in these models are necessary to enhance their accuracy and reliability, particularly in under-resourced languages like Polish. This study contributes to the ongoing efforts to ensure media fairness and balance by providing insights into the capabilities and limitations of LLMs in bias detection.近年、オンラインニュースソースへの迅速かつ容易なアクセスは、公的意見に大きな影響を与えており、特に偏向報道や誤情報の拡散を通じて顕著である。この傾向は、情報リソースの限られた言語圏において特に問題視されており、信頼性の高い情報へのアクセスが制限されているため、プロパガンダの検出がより困難となっている。本研究は、これらの問題に対処するため、異なる手法で感情分析を行い、大規模言語モデル（LLM)を活用してポーランド語のオンラインニュース記事における政治的および体制的バイアス、およびプロパガンダの検出を試みた段階的研究である。
本研究の第一段階では、「Polish Online News Corpus (PONC)」と呼ばれるデータセットの作成に焦点を当てた。このデータセットは、２０１９年から２０２１年にかけてのTVP InfoおよびTVN24の２つの主要ポーランドニュースプロバイダーから取得した２０万件以上の記事で構成されており、その後のすべての実験で利用された。PONCデータセットは、ＢeautifulSoupライブラリを用いた多段階のデータ収集・前処理プロセスを経て作成され、その品質と再現性を確保した。
次に、ニュース記事の出典をテキスト特徴に基づいて分類する実験を実施し、TVP InfoとTVN24の文体および内容の違いを間接的に評価した。使用した機械学習モデルは、ロジスティック回帰、ランダムフォレスト、サポートベクターマシン(SVM)、ナイーブベイズ分類器の４つである。また、特徴抽出手法として、Bag of Words (BoW)とTF-IDFの２種類を適用した。最高の精度を達成したのはTF-IDFを用いたランダムフォレスト分類器であり、性格率87.45％およびＦ1スコア88.29％を記録した。この結果は、テキスト特徴のみで両ニュースプロバイダーの違いを識別できることを示し、言語使用やコンテンツ構成に根本的な差異が存在することを示唆している。
次の分析では、これらの記事の感情的負荷（エモーショナルロード）の分析に焦点を当て、語彙ベースおよび統計的手法の両方を用いて感情分析を行った。本研究の目的は、表面的には客観的であると見なされるオンラインニュースに感情的な要素が含まれているかを評価し、異なる分析手法がバイアス検出にどのような影響を与えるかを明らかにすることであった。
初期実験では、ポーランド語の感情辞書であるNencki Affective Word List (NAWL)を使用し、感情を怒り、嫌悪、恐怖、幸福、悲しみ、中立の６つのカテゴリーに分類した語彙ベースの感情分析を行った。また、XLM-RoBERTaモデル（XNLIデータセットでファインチューニング済み）によるゼロショット分類を使用した。結果として、語彙ベースの分析では両ニュースプロバイダー間に統計的に有意な差は認められず、文脈的感情の検出能力が限定的であることが示された。一方で、XLM-RoBERTaモデルはより繊細な感情の差異を捉えることができ、TVP Infoの記事では、NAWL辞書のベースラインと比較して、「嫌悪」に分類される単語の出現回数が8倍多く、「恐怖」と「幸福」の出現回数はそれぞれ11倍、5倍少ないことが確認された。さらに、特定のテーマに関して感情的パターンの違いが明確に示された。たとえば、中絶に関する記事では、TVN24の記事が TVP Infoより3倍多く「幸福」の表現を含んでいた。一方、教会に関する記事では、TVP Infoの記事がTVN24の記事より2倍多く「悲しみ」を表す単語を含み、「幸福」の語数は半分であった。これらの結果は、語彙ベース手法では感情の多様性を過小評価する一方、統計モデル（XLM-RoBERTa）はより微細な感情の差異を捉えられることを示している。
また、本研究は、LLMsを用いた政治的および体制的バイアスの検出に関する課題と制限も明らかにした。特に、比喩表現や皮肉的表現などの非直接的な言語の理解が不十分であることが課題として挙げられる。GPT-4およびGPT-3.5-turboは、政治的傾向や体制的バイアスの自動注釈付けにおいて時間とリソースの削減に貢献する可能性を示したものの、体制的バイアスの検出精度は、政治的傾向の検出よりも10-12％低下した。これにより、明示的な政治的傾向（例：左右の政治的立場）の方が、政府や制度への支持といった間接的なバイアスよりも識別が容易であることが示された。
最後に、プロパガンダ技術の検出に関しても、LLMの有用性を検証した。GPT-4は、2つのタスク（二値分類および技術分類）で評価され、74％の精度で二値分類、69％の精度で技術分類を達成した。しかし、繰り返しや論点すり替えなどの微妙なプロパガンダ手法の検出精度は低く、複雑な表現の解釈には限界があった。総括として、本研究は、ポーランド語メディアの偏向検出においてLLMｓが補助的ツールとして有望であることを示すと同時に、言語リソースの制限やモデルの解釈力の限界が存在することを強調した。さらなるデータ拡張とモデルのファインチューニングの必要性が示唆されており、本研究の成果は、偏向報道の分析とメディアの公正性の促進に貢献するものである
Implementation of Commonsense Morality in Artificial Intelligence and Its Critical Evaluation from the Anti-speciesist Perspective
No full textStudy on optimal design of electromagnetic devices and acceleration by machine learning
No full text情報化社会の進展に伴って通信トラフィックが急速に増大しており，ネットワーク機器の高性能化，小型化が要求されていることから，無線通信に使われるマイクロ波回路の新たな設計手法が必要とされている．電磁デバイスの設計では，目標の特性と電磁界シミュレーションで計算した特性の差を目的関数として定義し，これを最小化する形状を求めることで設計を行う最適設計のアプローチがよく用いられる．最適設計の一手法であるトポロジー最適化（TO: topology optimization）は高い自由度の設計が可能であり，革新的な構造を見出すことができる．ただし，マイクロ波回路の基本要素であるマイクロストリップライン（MSL: microstrip line）により構成される回路への適用例は少なく，その有効性は不明である．また進化計算を用いた TO は大域的な探索が可能な代わりにその計算コストが問題となっている．電気電子機器の小型化のためにインダクタやモータなどのデバイスが高周波化しており，周波数に比例して増加するヒステリシス損失（ヒス損）を正確に考慮した設計が重要となっている．正確なヒス損解析には近似式を用いた簡便な解析手法よりも多くの計算コストが必要となり，多数のヒス損解析を行うインダクタやモータなどの最適設計では，その計算コストの増大が問題となる．上記の最適設計における計算コストの問題に対処するために，時間のかかる計算を機械学習により置き換える代替モデル法が知られている．計算を置き換えるだけの典型的な代替モデル法では，代替モデルによって特性を精度よく予測できることが前提となっている．一方で MSLで構成されるフィルタ回路は高精度な代替モデルの作成が難しく，精度が不十分な代替モデルで最適設計を実行すると，予測誤差に起因して最適設計が失敗してしまう問題がある．本研究では電磁デバイス，特にマイクロ波回路と電気機器を対象にその最適設計の高度化（TO 法の開発および高速化）を目的とし，以下の３つの課題を考える．1. マイクロ波回路へのトポロジー最適化の適用例が少なく，その有効性は不明である2. マイクロ波のフィルタ回路はその特性の正確な予測が難しく，電磁界解析を置き換えるだけの典型的な代替モデル法が MSL 回路の TO に適用できない3. ヒステリシス特性を正確に考慮できる既存のヒス損解析手法は実行に時間を要するため，最適設計などの多数のヒス損解析が必要な際に問題となる第 3 章では 1. の課題を解決するために，MSL で構成されるフィルタ回路および電力分配回路を TO により設計し，その有効性を検証する．本研究ではガウス基底関数法（NGnet 法）および幾何射影法の二つの TO 手法を検討し，従来の設計手法もあわせて比較する．最適設計を実施した結果，フィルタ回路の設計については NGnet 法の方が従来法よりも回路長を短くできることが分かった．また二つの TO を比較した結果，スタブ構造を活用できる，扁平なメッシュが生成されづらいという二つの理由から，幾何射影法の方が NGnet 法よりも MSL 回路の設計に適していることが明らかになった．第 4 章では 2. の課題を解決するために，代替モデルの精度を予測する手法およびそれを用いた最適設計手法を提案する．提案法では畳み込みニューラルネットワークにより代替モデルの精度を予測し，その予測結果に基づいて電磁界解析と代替モデルを切り替えて最適設計を行う．提案法により MSL からなるフィルタ回路を設計した結果，代替モデルの予測誤差に起因する最適化の失敗（誤誘導）を回避しつつ，最適化時間を 9 倍高速化できた．提案法を用いることで，高精度な代替モデルの作成が難しい MSL のフィルタ回路であっても代替モデルを利用した TO が可能であることを示した．第 5 章では 3. の課題を解決するために，機械学習により電磁デバイスのヒス損解析を高精度かつ高速に実行する手法を提案する．ヒステリシス特性は入力波形の極値によって特徴づけられるという性質を利用し，提案法では入力波形を極値で分割した区間ごとに代替モデルを適用する．インダクタのヒス損解析に対して提案法を適用した結果，従来の高精度なヒス損解析手法と同様の解析結果を得つつ，計算時間を約 2,000 倍高速化できた．さらに提案法を用いてヒス損を最小化するインダクタの最適設計を実行した結果，代替モデルの作成に要する時間を含めて最適設計の所要時間を 2.5 倍高速化できた．The rapid growth of communication traffic in the information society has led to demands for higher performance and smaller network equipment. This has created a need for new design methods for microwave circuits used in wireless communications.
In the design of electromagnetic devices, the optimal design approach is often used, where the difference between the target characteristics and the characteristics computed by electromagnetic simulation is defined as the objective function, and the shape that minimizes this difference is found. Topology optimization (TO), one of the optimal design methods, allows a high degree of design freedom and can find innovative structures. However, TO has not been sufficiently applied to circuits composed of microstrip lines (MSLs), which are the basic elements of microwave circuits, and its effectiveness is not yet known. In addition, although TO using evolutionary computation can perform global searches, its computational cost is a problem.
Devices such as inductors and motors are being used at higher frequencies to miniaturize electrical and electronic equipment, and it has become important to accurately evaluate the hysteresis loss (hysloss), which increases with frequency. Accurate hysloss analysis requires more computational cost than simple analysis methods using approximation formulas, and the increase in computational cost is a problem in the optimal design of inductors and motors, where a large number of hysloss analyses are performed.
To address the above problem of computational cost in optimal design, surrogate model methods are known to replace time-consuming computation with machine learning. The typical surrogate model method that only replaces computation assumes that the surrogate model can accurately predict the characteristics. On the other hand, it is difficult to create a surrogate model with high accuracy for filter circuits composed of MSLs, and if an inaccurate surrogate model is used for optimal design, the optimal design will fail due to prediction errors.
The purpose of this study is to improve (develop and accelerate the TO method) the optimal design of electromagnetic devices, especially microwave circuits and electrical devices. The following three problems are considered.
1.The effectiveness of topology optimization for microwave circuits is unknown because there are few examples of its application to microwave circuits.
2.The typical surrogate model method, which only replaces the electromagnetic computation, cannot be applied to the TO of MSL circuits because it is difficult to accurately predict the characteristics of microwave filter circuits.
3.Existing accurate hysloss analysis methods are time-consuming to perform, which is a problem when many hysloss analyses are required, such as for optimal design.
In Chapter 3, to solve Problem 1, we design a filter circuit and a power divider circuit composed of MSLs using TO and verify the effectiveness of the TO method. Two TO methods, the Gaussian basis function method (NGnet method) and the geometry projection method, are considered in this study, and conventional design methods are also compared. The optimal design results show that the NGnet method can shorten the circuit length for filter circuit design better than the conventional method. The comparison of the two TO methods shows that the geometry projection method is more suitable for designing MSL circuits than the NGnet method for two reasons: it can utilize stub structures, and it is less likely to generate flat meshes.
In Chapter 4, to solve Problem 2, we propose a method for predicting the accuracy of surrogate model and an optimal design method based on this method. The proposed method uses a convolutional neural network to predict the accuracy of the surrogate model, and switches between electromagnetic field analysis and surrogate model based on the prediction results to perform optimal design. The proposed method can avoid optimization failures (misleading) caused by surrogate model prediction errors and accelerate the optimization time by a factor of 9. The proposed method enables TO using surrogate model even for MSL filter circuits, for which it is difficult to create a highly accurate surrogate model.
In Chapter 5, to solve Problem 3, we propose a method to perform hysloss analysis of electromagnetic devices with high accuracy and speed using machine learning. The proposed method applies a surrogate model to each section of the input waveform divided by its extreme values, taking advantage of the fact that hysteresis characteristics are characterized by the extreme values of the input waveform. As a result of applying the proposed method to the hysloss analysis of an inductor, the computational time was reduced by 2,000 times while obtaining the same analysis results as those obtained by conventional high-precision hysloss analysis methods. Furthermore, the optimal design of inductors that minimize hysloss using the proposed method resulted in a 2.5-fold reduction in the time required, including the time required to create surrogate models
Stationary Reuse of Electric Vehicle Batteries and Their Utilization in Regional Microgrid [an abstract of dissertation and a summary of dissertation review]
No full textResearch on the Integration of Operational Planning for Mobility as a Service and Micro Grids [an abstract of dissertation and a summary of dissertation review]
No full text
Virtual reality in undergraduate and postgraduate nursing education : a scoping review protocol integrating data mining for topic discovery
No full textVirtual Reality (VR) encompasses a range of computer-based technologies that simulate complex scenarios, offering immersive, experiential learning in a controlled virtual environment. In nursing education, VR has the potential to enhance both technical and non-technical competencies. However, the existing literature on VR in nursing education is fragmented, making it challenging to fully grasp its scope, applications, and emerging trends. This scoping review protocol outlines a systematic approach to mapping the existing literature on the use of VR in undergraduate and postgraduate nursing education. Following Joanna Briggs Institute guidelines and the PRISMA-ScR 2020 framework, the review will include studies published in English, Spanish, or Italian, as well as those with an accessible HTML version to enable accurate automated translation and eligibility assessment. A comprehensive search will be conducted across PubMed, Scopus, CINAHL, and EMBASE, with no time restrictions. Two independent reviewers will assess study eligibility and extract data using a standardized form. Additionally, data mining techniques, including Latent Dirichlet Allocation enhanced by Bayesian optimization, will be employed to identify trends and emerging topics in the field, providing valuable insights for educators, researchers, and policymakers.
•This scoping review protocol outlines the methodology for systematically mapping the existing literature on VR in nursing education, providing a comprehensive overview of its applications at undergraduate and postgraduate levels.
•Advanced data mining techniques will be applied to uncover emerging trends and key topics, enhancing the understanding of VR’s evolving role in nursing education.
•Findings will offer methodological and practical insights, supporting educators, researchers, and policymakers in optimizing and expanding VR-based learning strategies in nursing