81690 research outputs found
Sort by
二種類の分散表現の融合に関する研究
No full textNowadays, connectionism is the mainstream theory that prevails in Artificial Intelligence (AI) domain. The theory is implemented by deep learning of neural networks and big data. Accompanied by the enhancement of computational resources, with any well-designed training objective, a neural network with billions of parameters is trainable for addressing tasks in the sub-domains of Computer Vision (CV), Audio Signal Processing (ASP), and Natural Language Processing (NLP). In the NLP domain that I am dedicated to, the transformer-based architecture wields a significant impact on the existing powerful language models such as auto-encoding models like BERT and autoregressive models like GPT. These neural models are well known as Large Language Models (LLMs), influencing not only academic research directions, but also our daily life. The semantic representations output from LLMs are distributed representations, which are the same definition for non-transformer-based neural embedding models like Word2Vec, ELMo, and so on. Distributed representations are dense vector representations (commonly referred to as embeddings), where each dimension of all vectors is the semantic axis distributed with numerical values for different words or tokens. These pieces of semantic information are distributed across the entire vector for each single word or token. On the other hand, distributional representations are sparse count vectors, where the frequency of occurrence of each word in any given document or sentence is used as a feature. The implementation of distributional semantics is known as Vector Space Models (VSMs). The classical one is term-document matrix, where term indicates all individual words comprising different documents. VSMs are commonly used for information retrieval and document classification via the measurement of semantic similarity using frequency distribution in the term-document matrix. There are two common grounds between distributed representations and distributional representations. The first is that distributional representations are somewhat equivalent to distributed representations once the term-document matrix of the former is factorized into dense vector representations. The second is the distributional hypothesis — you shall know a word by the company it keeps. This hypothesis has contributed to some effective training objectives for contemporary neural models in NLP. For example, the hypothesis is adopted by the Masked Language Model (MLM) pre-training objective of BERT and the Continuous Bag of Words (CBOW) pre-training objective of word embeddings like Word2Vec. Meanwhile, distributional representations are also based on the hypothesis. The similarity of semantics can be measured by cosine similarity of distributional representations in their vector space; e.g., words having similar document distributions tend to achieve high cosine similarity. During the period of my PhD study, my research focus is to investigate whether the fusion of distributional semantics and distributed semantics can contribute to NLP development in terms of real-word applications and semantic representations. I come up with two fusion approaches: mapping and structural connection. Regarding the mapping approach, I propose a novel method by mapping weighed distributed representations to distributional representations via two trainable projection matrices. The trained matrices are frozen during inference. As a result, any sentence pair from different task datasets can obtain their distributional semantic representations via the projection from distributed representations, and the obtained representations are used as input to different classifiers for paraphrase identification or other NLP sentence-pair tasks. In this case, only the parameters of classifiers need to be optimized for different tasks. This approach contributes to power-efficient continual learning of sentence-pair tasks. I write a journal article for this approach titled “Parameter-efficient feature-based transfer for paraphrase identification”, which is accepted by Natural Language Engineering journal. Regarding the approach of structural connection, I propose DG embeddings also derived from distributional representations, which are created based on the inspiration of a real-world scenario called headword search. DG embeddings are used to gloss bidirectional context words for target-word predictions in the MLM pre-training objective of BERT. As a result, better acquisition of vocabulary is attainable for BERT after pre-training. Also, this work leads to a computational inexpensive future expansion owing to the above-mentioned property of distributional semantics. I write a journal article for this approach titled “DG Embeddings: The unsupervised definition embeddings learned from dictionary and glossary to gloss context words of Cloze task”, which is accepted by Knowledge-Based Systems journal
RNAi依存的ヘテロクロマチン形成における転写活性化因子としてのMoc3に関する研究
No full textTranscription from pericentromeric heterochromatin and its transcripts have crucial roles in diverse cellular processes such as heterochromatin establishment, heterochromatin maintenance, genome stability, and early embryo development. In higher eukaryotes, the misregulation of transcription from heterochromatin results in cancers and other harmful diseases. In fission yeast Schizosaccharomyces pombe, heterochromatin is established and maintained primarily by RNA interference (RNAi) dependent system. A core process of RNAi-dependent heterochromatin formation is widely conserved in various eukaryotes; the production of small interfering RNA (siRNA) and the targeting of siRNA-bound Argonaute to chromatin using its sequence complementarity. In addition, transcription from heterochromatin by RNA polymerase II is critical for providing substrates for siRNA synthesis and a platform for assembling RNAi factors. However, the detailed mechanism of transcriptional regulation and the transcription factors are still unclear. To date, various heterochromatin factors have been identified by genetic screening using fission yeast. Here, we attempted to find novel genes whose products are involved in the transcriptional regulation of pericentromeric heterochromatin by reverse genetic screening. In the screening, we found that a zinc finger protein Moc3 localizes pericentromeric heterochromatin through its zinc finger domain. It activates strand-specific transcription when heterochromatin structure or heterochromatin-dependent silencing is compromised. Although Moc3 is not essential for heterochromatin maintenance, the absence of Moc3 significantly reduces the efficient establishment of heterochromatin. These results indicate that Moc3 acts as a transcriptional activator of pericentromeric heterochromatin to induce RNAi-dependent heterochromatin establishment
模擬大気腐食環境下における高強度低合金鋼の水素侵入挙動に関する研究
No full textStress corrosion cracking (SCC) is a crucial safety concern of steel structures. The coupled effects of tensile stress and corrosive environments on steel can cause crack growth and lead to unexpected and sudden failure. Currently, a potential issue of SCC is to address the hydrogen-induced problems due to a positive correlation between the strength of a metal and its susceptibility to hydrogen embrittlement. For this reason, the wide utilisation of high-strength low-alloy steels in the construction industry, especially in storage and distribution systems for hydrogen, makes the hydrogen-assistant failure of these components a critical concern. Atmospheric corrosion is an environmental degradation phenomenon of steel structures and is commonly accompanied by hydrogen evolution as a side reaction. A considerable amount of hydrogen is supposed to be produced and subsequently absorbed by the steel, presenting challenges to the material durability of steel infrastructures. It is, therefore, necessary to consider the potential problems of corrosion and understand the mechanisms of hydrogen-assistant cracking of steel during corrosion in atmospheric environments. The present thesis concentrates on the impact of environmental factors such as relative humidity (RH), temperature, and salt aerosols on corrosion performance and the corresponding hydrogen uptake during atmospheric corrosion of high-strength low-alloy steels. Chapter 1 provides a fundamental overview of the state of knowledge with respect to atmospheric corrosion of steel and hydrogen-assistant cracking, which provides the information required to understand the background and importance of hydrogen-related issues in the scope of atmospheric corrosion. The main focus of Chapter 2 lies in determining the hydrogen absorption behaviour of steel during corrosion under different RH conditions. RH is a factor of crucial importance since it is responsible for the presence of electrolytes during atmospheric corrosion. When RH reaches a specific value known as deliquescence relative humidity (DRH), the deposited salts on the steel surfaces will absorb moisture from the air to form aqueous electrolyte layers for corrosion to occur. Thus, the effect of RH on the corrosion and hydrogen absorption of steel under NaCl deposit was investigated by controlling the RH at different levels below the DRH of NaCl during wet-dry cycles. The results suggest that hydrogen absorption into steel can occur even if the surrounding RH is below the DRH of NaCl. The amount of permeated hydrogen increases with higher RH in the dry stage of the wet-dry cycles, which is attributed to the prolonged period of the presence of liquid phase inside the corrosion products. Chapter 3 assesses the applicability of the Hard and Soft Acids and Bases (HSAB) concept in explaining the effect of metal cations of dissolved salts on corrosion kinetics under salt droplets. An analysis of corrosion and hydrogen permeation of steel under aqueous NaCl, MgCl2, ZnCl2, and AlCl3 solution droplets was conducted to investigate the role of different hygroscopic slats and their metal cations in the cyclic wet-dry corrosion. When corrosion occurs under salt droplets containing Mg2+, Zn2+, and Al3+, cation-containing layers are supposed to form on the steel surfaces and present corrosion resistance. Based on the HSAB concept, the metal cations with higher cation hardness are easier to combine with the hard bases such as H2O and OH- of the passive films, forming stable protective passive films against corrosion attack under salt droplets. Chapter 4 verifies whether the concept of the time of wetness (TOW) is practicable to understand the association between hygroscopic salts and the hydrogen absorption behaviour of steel during wet-dry cycles. After water evaporates and steel surfaces become apparently dry, the hygroscopic property of the salts deposited within the porous rust layers plays a dominant role in the corrosion and hydrogen absorption processes. The salts with low deliquescence values retain the presence of thin electrolyte layers with high chloride concentrations within the porous rust structures, affording a prolonged TOW for corrosion and hydrogen absorption into steel. Chapter 5 summarises the findings and gives the overall conclusion of the research work. Based on the experimental results and analysis, the TOW and HSAB concepts are applicable in explaining the role of salts in the corrosion and hydrogen absorption activity of high-strength low-alloy steel during atmospheric corrosion. This study suggests that localised pitting and acidification under rust layers are responsible for accelerated hydrogen entry into steel during wet-dry cycles
A Study on formation of MAX-phase layer on carbon steels by powder pack method [an abstract of dissertation and a summary of dissertation review]
No full text札幌市の下水処理水および河川水中の希土類元素―新興微量汚染物質として [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
