Recently, many researches for node representation learning have focused on contrastive learning. Node-level contrastive learning aims to distinguish the same node representations (positive pairs) in two different views from other node representations (negative pairs). However, since negative pairs are sampled regardless of graph structure, most constrastive methods make no consideration of the graph homophily that similar nodes may be more likely to attach to each other than dissimilar ones. In this study, we propose two ideas that take into account the structure information of graph. 1) Edge Reconstruction Loss. It uses the representations of the proximity nodes as positive pairs. 2) Average Edge Reconstruction Loss. It uses each node representation and average of representations of the proximity nodes as positive pairs. We perform experiments with datasets which have various properties such as citation and co-selling. Experimental results show that our method improves the conventional baseline study.

In recent years, there have been various studies to estimate the causal relations of systems from observational samples. DirectLiNGAM is one of the popular causal discovery methods and aims at efficient and stable estimation by iteratively identifying and removing the variables at the top of the underlying causal relation. In real-world scenarios, sufficient samples are often not available due to technical, ethical, or cost problems. In such cases, DirectLiNGAM removes information necessary for estimating top variables as the iterations proceed, resulting in the deterioration of causal discovery performance. This paper proposes a new approach to address the problem by estimating the partial causal structures based on independent relations among variables and preserving the necessary information. We show that the proposed approach can reduce the estimation error by more than 80% compared to DirectLiNGAM on randomly generated causal discovery problems, especially under small samples.

Domain adaptation (DA) is a method for learning a model with good performance for target data, given labeled source data and unlabeled target data from different domains, by learning a domain-invariant feature space. Heterogeneous domain adaptation (HDA) is a type of DA that can be applied when the feature space differs between the source and target data. Conventional HDA assumes that all labels exist in the source data, but in reality there are cases where only positive examples exist. In this paper, we propose a HDA method in such a setting, i.e., in a PU learning setting where only positive source data and unlabeled target data exist. By using adversarial learning, the proposed method simultaneously achieves binary classification of positive and negative examples in the target data and learning of domain-invariant feature space. We experimentally show that the proposed method outperforms the performance of various baseline methods.

Recent theoretical and empirical studies show that the existence of strong lottery tickets (SLT), i.e., sparse subnetworks that achieve high performance \textit{without any weights updates}, in randomly initialized over-parameterized networks. However, little is known about how these SLT are discovered by the de-facto edge-popup algorithm (EP), making it difficult to improve its performance. In this paper, we first show that EP suffers from \textit{the dying edge problem}, i.e., most weights are \textit{never} used during the entire search process, suggesting that the edge-pop algorithm only searches around the randomly selected initial subnetworks. We then propose a \textit{iterative edge-pop (iEP)}, which repeats the EP while gradually increasing the pruning rate and rewinding the learning rate after each iteration. To validate the effectiveness of the proposed method, we conducted experiments using ImageNet, CIFAR-10, and CIFAR-100 datasets. As a result, we achieved a performance of 76.0\% with approximately 20 million parameters using iEP, while the regular weight had 22 million parameters with 73.3\% accuracy and Edge-Pop had about 20 million parameters with 73.3\% accuracy on ImageNet. Our results also provide new insight into why iterative pruning often helps to find good sparse networks.

In the stock market, market beta, which indicates the linkage between the stock index and the individual stock price, is one of the index values that attracts attention from investors. The price of stocks with high market beta (high beta) may strongly reflect investors' sentiment toward the market outlook. On the other hand, the recent trend of the individual stock does not necessarily coincide with the overall market trend, as it also depends on its recent performance. Therefore, the bidding for stocks with extreme trends may strongly reflect investors' expectations for the future relative to the market and the individual stocks. In this study, we propose a method of anomaly detection based on the price movements of stocks in the stock market. We divide the constituent stocks into regions based on market beta and trend, which are expected to reflect investor sentiment toward the market outlook. By applying the Graph Based Entropy method to the price movements of each region, we attempted to detect anomalies such as a strong downtrend in a stock index. We performed the tests on three equity indices, TOPIX 500, S&P 500, and STOXX® Europe 600, and succeeded in detecting several strong downtrends.

ROCKET (RandOM Convolutional KERrnel Transform) gains remarkable success to achive comparable performance with other SOTA algorithm in a fraction of the time. ROCKET, a feature-generating algorithm for uni- and multi-variate time series is used in combination with linear classifiers, such as Ridge model. Apart from achieving good classification result, to know the driving factor of the classification is equally important in practice. This corresponds to know which time series variable contribute to the classification result. Usually magnitude of weight coefficients for each feature is used as the measure. But ROCKET `convolves' the original time series variables, and makes it difficult to extract such an information from the classification result. In this paper we propose a method to elicit importance of each original time series variable from weight coefficients of linear classifier with ROCKET features. We show our method can successfully reconstruct the importance of original variables of multivariate time series from the weight coefficients of Ridge classifier applied on the ROCKET-generated features.

Hierarchical clustering is one of the most popular methods used to extract cluster structures in a dataset. However, if the hierarchical clustering algorithm is sensitive to a small perturbation to the dataset, then the credibility of the output hierarchical clustering are compromised. To address this issue, we consider the average sensitivity of hierarchical clustering algorithms, which measures the change in the output hierarchical clustering upon deletion of a random data point from the dataset. Then, we propose a divisive hierarchical clustering algorithm with which we can tune the average sensitivity. Experimental results on benchmark and real-world datasets confirm that the proposed method is stable against the deletion of a few data points, while existing algorithms are not.

To promote social implementation of Society 5.0, it is essential to detect various kinds of information in the real world. With the complexity of the real world, the obtained data inevitably become hyper-multi-dimensional. In this study, we propose a new method of dimensionality reduction that effectively exploits the latent wave properties of many time series data. Specifically, the first step is to cluster the multidimensional time series data into a specific number of clusters based on similarity. Then, assuming that data belonging to the same cluster exist in the same wavelength band, the synthetic wave principle is applied. Based on the physical fact that waves after superimposition of waves of different wavelengths can be represented by the harmonic mean of each wave, a dimensionality reduction is performed that preserves the information in the original multidimensional data. in this way, we propose dimensionality reduction method that can compress each variable with less information loss than conventional methods.

We introduce Cross3DVG, a new task for cross-dataset visual grounding in 3D scenes, revealing the shortcomings of current 3D visual grounding models developed in the limited datasets and hence easy to overfit specific scene sets. For Cross3DVG, we have created a new large-scale 3D visual grounding dataset that contains over 63k diverse linguistic annotations to 3D objects in 1,380 RGB-D indoor scans from the 3RScan dataset with human annotation. This is corresponding to the existing 52k descriptions on the ScanNet-based 3D visual grounding dataset of ScanRefer. We perform cross 3D visual grounding experiments in that we train a 3D visual grounding model with the source 3D visual grounding dataset and then evaluate it on the target 3D visual grounding dataset without target labels (i.e., zero-shot setting.) Extensive experiments using well-established visual grounding models as well as a CLIP-based 2D-3D integration method show that (i) cross 3d visual grounding has significantly lower performance than learning and evaluation in a single dataset (ii) better detectors and transformer-based headers for 3D grounding are useful, and (iii) fusing 2D-3D data using CLIP can further improve performance.

Recently, research on graph neural networks has attracted attention. In particular, graph convolution is applied to the analysis of human-to-human interaction in SNS and road traffic network to predict future interaction and traffic volume. However, these graph data are usually time-series graphs. It is difficult to apply analysis methods to static graphs. In this paper, we apply graph autoencoders to time-series graphs.

In this work, we propose a defense method to disable the effect of adversarial examples patch attack against object detection by painting bounding boxes which have a score value lower than detection threshold. We show our method is effective by evaluation over INRIA Person Dataset.

Online coupon distribution is a significant marketing measure that leads to increased sales. However, distributing coupons blindly risks lowering a company's profit ratio. It is, therefore, essential to estimate the coupon effect. In addition, users' potential purchase intention is thought to make a difference in the coupon effect. For example, users with low purchase intentions are likely to increase their gross profit through coupons. In contrast, users with high purchase intentions will likely decrease their gross profit through coupons. Therefore, it is possible to conduct highly effective targeting by analyzing the relationship between potential purchase intention and the coupon effect. In this study, we propose a framework containing an experimental design and a verification method based on machine learning to analyze the relationship between the coupon effect and the user's potential purchase intention. Finally, we demonstrate the effectiveness of the proposed framework by applying it to real-world data.

In general, Machine Learning does not ensure the proper prediction if the statistical structures of the features between training data and prediction data are different, but it is sometimes required to apply a prediction model to a target which may have the different structure from its train data. In recent years, the studies to address this challenge have been actively conducted, and one of them is Adversarial Discriminative Domain Adaptation(ADDA), which is the classification model with adversarial training of Generative Adversarial Networks(GAN). ADDA has a defect that it uses all data from mini-batch, which includes bad data for training. In this study, we propose the improved method of ADDA with the application of GAN's related study, Top-k training. This application would enable ADDA to select useful data based on internal outputs, and the prediction accuracy is expected to increase. The result of the experiment shows that proposed method has significances of the accuracy and the length of training time.

Despite the successes of the Gaussian process in modeling highly dimensional complex dynamics, describing fluctuations of financial time series is still challenging. The problem arises from non-Gaussian, in particular, the asymmetric and fat-tail nature of the financial time series. In this paper, we propose a generalized hyperbolic process (GHP) as an alternative to the Gaussian process and Student's t-process to incorporate asymmetric non-Gaussian distribution into the Bayesian kernel model. The GHP is realized by the marginalization of a mixture Gaussian process with the generalized inverse Gaussian distribution. For prediction, we analyticaly derive the conditional distribution of the GHP. To estimate the parameters of the GHP, we present an expectation-maximization algorithm. In addition, we present parameter estimation results of the GHP for synthetic and empirical market datasets.

In recent studies, Deep learning (DL) models demonstrate high performances on various datasets. FT-Transformer (FTT), which applies the Transformer model to tabular data, has been proposed as an effective DL model for tabular data. FTT performed higher on several datasets than gradient boosting models, the current mainstream for tabular data. Initially proposed for unstructured data, Transformer shows high performance by sensitively considering the relationships between all features (e.g., words and patch images) through the attention mechanism. However, the relationships between features in tabular data can be considered less complex than those in unstructured data such as documents and images. Therefore, we propose an improved FTT suitable for tabular data that does not excessively consider the unnecessary relationship between features in the Transformer's attention mechanism and improves performance and computational efficiency. We perform evaluation experiments on regression, binary classification, and multi-level classification tasks and show our model's effectiveness.

Japan is an earthquake-prone country, with 1,000 to 2,000 sensible earthquakes observed per year. Seismological research is also active, and the Japan Meteorological Agency, the National Research Institute for Earth Science and Disaster Prevention, and local governments have established seismic observation networks. In recent years, various studies have been conducted to detect, classify, and predict the intensity of earthquakes using machine learning techniques based on the large amount of observed seismic waveform data. Therefore, it is necessary for researchers to set and collect the location of the hypocenter, time of occurrence, and target observation points in order to create data for training purposes. In this paper, we construct an earthquake ontology and assign URIs to earthquakes based on observed waveforms and hypocenters to investigate the availability and distribution of earthquake catalogs that can be used as learning data.

In the field of earthquake engineering, an empirical equation called Ground Motion Prediction Equation (GMPE) is often used to predict the intensity of ground motions caused by earthquakes. In previous studies, the relationship between object variables and explanatory variables has been determined empirically and subjectively based on simplified physical models. Constructing GMPEs that better reproduce the data will not only improve the accuracy of earthquake motion prediction, but also lead to the acquisition of knowledge in seismology and earthquake engineering. Recently, AI Feynman, a physics-inspired method for symbolic regression that combines neural network fitting with a suite of physics-inspired techniques, has been proposed and is superior to conventional methods. In this study, we attempted to apply AI Feynman to the construction of GMPEs. Synthetic tests showed that the symbolic regression was successfully achieved by normalizing the values of explanatory variables, while the symbolic regression of GMPEs was difficult in the noisy case.

Recent advances in machine learning technologies have enabled the automatic detection of earthquakes from waveform data. In particular, various state-of-the-art deep-learning methods have been applied to this endeavor. In this study, we proposed and tested a novel phase detection method employing deep learning, which is based on a standard convolutional neural network in a new framework. The novelty of the proposed method is its separate explicit learning strategy for global and local representations of waveforms, which enhances its robustness and flexibility. Prior to modelling the proposed method, we identified local representations of the waveform by the multiple clustering of waveforms, in which the data points were optimally partitioned. Based on this result, we considered a global representation and two local representations of the waveform. Subsequently, different phase detection models were trained for each global and local representation. The overall phase probability was evaluated as a product of the phase probabilities of each model. This additional information on local representations makes the proposed method robust to noise, which is demonstrated by its application to the test data. Furthermore, an application to seismic swarm data demonstrated the robust performance of the proposed method compared with those of other deep learning methods.

In former data markets, there is a lack of diversity and quantity of data. Even if there was a plenty of data sets, it was often difficult to obtain the desired data. Instead of relying on data experts to supply data, this paper propose a methodology that can be applied to data collection and data transactions by specifying ambiguous requirements through conversational AI model. The goal of this methodology is to increase the amount and diversity of data in the data market and to spread data-based labor through innovation in data collection and transaction. The method proposed here helps to structure issues by breaking them down into data variables, which makes it easier to collect diverse data. Therefore, data providers are able to offer a wider range of data within the marketplace.

Online learning is advantageous for its efficiency and effectiveness in handling ever-growing data. Most existing methods assume that the features are fixed, but they can keep varying in such a way that old ones vanish and new ones emerge. To address these capricious features, this study proposes a subspace learning method. Specifically, a devised subspace estimator maps heterogeneous feature instances to a low-dimensional subspace and then a classifier is learned in this latent subspace. The estimator and the classifier are obtained recursively via alternating updating to sketch data in an online fashion. Under some mild assumptions, we provide its theoretical performance guarantee. The experimental results on several datasets corroborate the rationality of the theoretical analysis and the effectiveness of this novel scheme.

We propose a new method for training data generation in self-play deep reinforcement learning, which are widely used in Game-AI like AlphaGoZero, AlphaZero, and so on. Generally, such self-play learning has not utilized most of search results that are generated in self-play. Currently, few researches try to make use of them. The proposed method converts the search result to training data by estimating final win/lose rewards and policy for it. The experimental investigation with various hyperparameters for the training suggests that the proposed method will help learning the policy effectively and stabilize the training.

Long-tailed recognition, where the per-class sample size is highly skewed, has recently gained in importance and is challenging because the accuracy of data belonging to classes with a few samples deteriorates in naive training. Two-stage training with weight decay and weight clipping has been proposed to improve the accuracy of long-tailed data. However, this method requires tuning many hyperparameters in the second stage of training, and why it is effective for long-tailed data is unknown. We analyzed the algorithm and found that it can be decomposed into the increase in the FDR of the feature extractor by weight decay and logit adjustment by weight decay and weight clipping. On the basis of this analysis, we propose a training algorithm without the second stage that results in both improved accuracy and simplification.

CNN (Convolutional Neural Network), which is used for image recognition and object detection and has recently been implemented in the System on Chip (SoC) environment, has a large amount of computation in the convolution layer, so performance may be degraded in the SoC environment. In this research, as an investigation of speeding up Convolution calculations for GPUs on SoC, an algorithm that solves Convolution operation by matrix multiplication was implemented with OpenCL, and the processing time of the object detection algorithm YOLO-Nano was measured on Intel and Qualcomm SoCs. As a result, compared to the Tensorflow Lite CPU, Qualcomm and Intel CPUs achieved speed improvement effects of 1.04 times and 6.37 times, respectively, and GPUs achieved speed improvement effects of 1.21 times and 1.52 times, respectively.

With the recent development of machine learning technology, machine learning-based decision-making has been used in various situations. However, some researchers have pointed out the unfairness of machine learning. Most existing studies on fairness assume that all the data have sensitive attributes. However, in many situations, we cannot observe the sensitive attributes because laws and regulations may prevent the collection of such data, and also the annotation cost is high. In this study, we develop fair classification algorithms for situations where only a small number of sensitive attributes are available. Our algorithm creates a sensitive attribute classifier using a semi-supervised learning method and assigns the output of a classifier to unlabelled data as a pseudo-sensitive attribute. As a result, we can execute a existing fairness method using pseudo-sensitive attributes. Experimental results show that the proposed method can achieve a competitive trade-off between accuracy and fairness compared to the ideal case where all the data have sensitive attributes, even though our method can access only a few labels of sensitive attributes. In addition, we found that the accuracy of the pseudo-sensitive attributes is essential to achieve fairness. Also, we found that filtering by confidence could negatively affect the accuracy.

In this study, we develop a computational model in which an agent without a lexicon discovers words and their meanings by extending the model for cross-situational learning with unsupervised word segmentation. A computational model for cross-situational learning was proposed that learns the word's meaning by estimating its attributes and categories. However, this model did not include word segmentation and did not assume the ungrounded words, i.e., words that are not associated with sensory information. The proposed model simultaneously infers the words contained in sentences, the attributes and categories corresponding to those words, and ungrounded words or not. Experimental results show that our model, which considers sensory information, improves segmentation performance by 2.1\% and clustering performance by accounting for ungrounded words.

One feature of e-testing for educational assessment is an automated test assembly of parallel test forms, for which each form has equivalent measurement accuracy but with a different set of items. An important task for automated test assembly is to assemble as many tests as possible. Although many automatic uniform test assembly methods exist, the maximum clique using the integer programming method is known to assemble the greatest number of uniform tests with the highest measurement accuracy. However, the automated test assembly often causes a bias of item exposure. This bias problem decreases the reliability of items and tests. To solve this problem, this study formulates the test assembly problem as the objective function of integer programming with two logistic item exposure penalties. The first penalty is a deterministic penalty of logistic item exposure. The second penalty is a stochastic penalty with logistic item exposure based on the Big-M method, a standard technique in mathematical programming. Numerical experiments demonstrate that the proposed methods reduce the bias of item exposure without decreasing the number of tests.

To detect and prescribe for individuals potentially failing to maintain their weight after weight loss, it is essential to predict how much weight individuals will maintain in the future (weight maintenance rate) and provide interpretable insights about factors behind the predicted failure of weight maintenance. To this end, this paper proposes a weight maintenance rate prediction method with high interpretability and accuracy. Using statistical causal discovery DirectLiNGAM, this method captures causalities among variables to make prediction results interpretable and selects direct causal variables as prediction features to estimate weight maintenance rate accurately. We evaluated our method on the observational data of our weight loss study for 140 subjects over 8 weeks collecting weight, physical activity, sleep, and food logs on a daily basis. As a result, our proposed method identified 4 direct causal features from 103 variables and showed the highest prediction performance with the causal features. Based on the results, we discussed important features in predicting weight maintenance rate and behaviors that should be monitored to avoid weight maintenance failure.

High accuracy has been achieved in various Japanese language processing tasks by fine-tuning pre-trained Japanese BERT. Input text for Japanese BERT needs to be tokenized into words and subwords, but there are various word dictionaries and subwordization methods. In this study, we create Japanese BERT models with different tokenizers and examine their effects on the masked language model, a pre-training task, and on downstream tasks. It is found that differences in tokenizers cause accuracy differences in masked language models and downstream tasks, and that the performance of masked language models and downstream tasks are not necessarily dependent on each other.

We have verified that the sentence vectors output by Sentence-BERT capture the meaning of sentences using k-means and UMAP. As a result, we confirmed that the sentence vectors generated by Sentence-BERT capture the meaning of sentences very well. In this study, we examine the properties and characteristics of the sentence vectors that are considered to capture the meaning of sentences. We visualize the sentence vectors by imaging the sentences, and examine the output results when changes are made to the sentence vectors. As a result, we confirmed that there is a difference in the information expressed in each dimension as a feature of the sentence vector, although the roles are not completely divided.

Providing a means of communication to support daily living for patients with severe motor dysfunction is crucial. We propose a brain-computer interface system based on near-infrared spectroscopy (NIRS) using ensemble learning to utilize physiological signals for communication. We used the OEG-SpO₂ to measure NIRS signals in three patients with neurological disorders. Brain function was measured using a block design consisting of 30 seconds of rest and task each, with a mental arithmetic task and a music recall task. The classifier was a random forest with feature selection and dimensionality compression. We evaluated the model performance integrating predictions obtained from datasets generated by applying different preprocessing methods. The results showed an accuracy of 85%, 79%, and 67% in participants A, B, and C, respectively. We plan to improve the discrimination performance and validate the robustness of the model against non-stationarity of the NIRS signal by long-term measurements.

The standard blood pressure monitor measures systolic blood pressure (SBP) and diastolic blood pressure (DBP) by using oscillometric method. The method measures blood pressure (BP) precisely by the variabilities in waveforms amplitude but sometimes vulnerable to noise. Recently, the novel method has been developed by employing the neural network based on the features of beat-by-beat waveforms. It classifies each beat into the one of the three different phases, presystolic, between systolic and diastolic, and after diastolic. It could use more information amount than the oscillometric method but still requires some improvements on its performance. In the study, we improved the novel method based on the phase estimation by employing the new features and the new architecture of the neural network. The experiments were conducted to investigate the performance of the developed method. According to the experimental results, the method improves the standard deviation of errors on both SBP and DBP.

Delirium is a type of cognitive dysfunction. Patients suffering from delirium require appropriate treatment, including early detection and careful medication, due to increased medical costs and prolonged hospitalization. However, since the diagnosis of delirium is generally costly in terms of manpower and time, the establishment of early automated diagnostic techniques is desirable. Therefore, this study aims to realize a multimodal early prediction model of delirium using height and weight, heart rate, blood glucose level, and externally captured video images of the patient and so on. In this paper, we followed the model proposed in a previous study, appropriately cleansed patient information obtained from the public dataset MIMIC-III, and validated it to evaluate the method proposed in the previous study. The results confirm the suitability of emergency medicine data for use in multimodal delirium early prediction models.

The sleep patterns of patients with dementia have been observed to be affected. This study aims to investigate the feasibility of developing a machine learning model that can classify scores of dementia screening tests based on sleep activity data that can be collected with minimal burden on participants. Data on sleep activity was collected from 124 elderly patients with varying levels of cognitive ability. The Mini Mental State Estimation (MMSE) cognitive test scores were used to determine the cognitive states of the patients. To classify the dementia scale and identify individuals with low-MMSE, we employed an efficient sequence model to capture time-series changes in sleep activity. Using LSTM models, a maximum macro F1 score of 0.67 was achieved in the bina

Objective: The purpose of this study was to create artificial intelligence (AI) models for detecting COVID-19 infection on chest X-ray images and to evaluate the performance of the AI models. Methods: In this study, we used chest X-ray images taken at our institution, and PCR results were used as correct labels. We trained models using convolutional neural network (CNN) and Transformer models and evaluated the performance of the models. We also created transfer learning models using publicly available datasets. Results: There were 214 COVID-19 positive and 153 COVID-19 negative cases. The ages ranged from 15 to 98 years old (mean 66.0 years old), and there were 208 males and 159 females. The accuracy was 60.8% and the area under the curve was 0.664 with the CNN model using transfer learning. There was no significant difference in the performance of the AI models between CNN and Transformer, and transfer learning did not significantly improve the performance of the AI models. Conclusion: The performance of the AI models using CNN and Transformer for detecting COVID-19 infection on chest X-ray images taken at our institution was not satisfactory, even with the use of transfer learning.

Code search is a task to find programming codes that semantically match the given natural language queries. Even though some of the existing datasets for this task are multilingual on the programming language side, their query data are only in English. In this research, we create a multilingual code search dataset in four natural and four programming languages using a neural machine translation model. Using our dataset, we pre-train and fine-tune the transformer-based models, and then evaluate them on multiple code search test sets. Our results showed that the model pre-trained with all natural and programming language data has achieved the best performance in most cases. Exceptionally, the model pre-trained only with Python for programming language data performed better when tested on Python data.

Deep learning models have achieved high accuracy in various natural language processing tasks, but it is controversial whether these models encode the structural information of sentences. In this context, Recurrent Neural Network Grammars (RNNGs) were proposed as a model considering syntactic structures. In this study, we implemented RNN-CCGs, language models that substitute CFG, the underlying grammar of RNNGs, with Combinatory Categorial Grammar (CCG). Compared to CFG, CCG provides more appropriate syntactic structures for natural language and provides paths of semantic composition. Since RNNGs do not consider part-of-speech tags, we implemented a model that predicts POS tags necessary for semantic composition. We compared RNN-CCGs with RNNGs with/without POS tags and evaluated their behaviours.

The accuracy of POI (Point of Interest) categories is becoming increasingly important since numerous users use services that rely on POI categories nowadays. Machine learning models are widely used to infer POI categories from various information. Recently, it has been reported that multimodal deep models show high performance in many tasks. In this paper, we propose a multimodal deep model for POI category prediction using both linguistic and image information. In order to use image information effectively, the proposed model (1) introduces a loss against prediction based only on linguistic information and (2) introduces pooling to input multiple images for each POI. Using Yahoo! Japan's POI database, we confirmed that the proposed method improves the performance of POI category prediction compared to the baseline that uses only linguistic or image information.

Deep learning for natural language processing (NLP) outperforms traditional approaches in many tasks. High-performing deep learning models are realized by proficiently combining techniques in model architecture such as attention mechanisms. Open-access large scale pre-trained models and easier pipeline construction based on End-to-End learning have lowered barriers to develop such models. The practice of academia to share fundamental language resources such as morphological analysis tools and linguistic datasets as well as the relaxation of copyright on automatic collection of text data also encourage research and development of models for NLP. In real businesses, ethical considerations are required to ensure that models do not output harmful expressions. However, such consideration suitable for everyone is difficult to achieve because there are no universal norms in ethics. In addition, the performance of deep learning models has uncertainty in principle. Furthermore, the security risks specific to machine learning models should also be noted.

Prompts have attracted attention as a way to exploit the performance of pre-trained language models, one of which is the chain-of-thought prompt. Chain-of-thought prompts are prompts that encourage the explicit expression of intermediate thoughts in order to derive a final answer, and have attracted attention for their ability to improve multi-stage reasoning. On the other hand, it remains unclear how chain-of-thought prompts affect models and enable multi-step reasoning. In this paper, we examine how neurons in models are internally influenced in multi-step reasoning tasks, against the background of existing studies that interpret task performance based on the activation of neurons in language models. The results revealed that there are neurons that are commonly activated in multiple chain-of-thought prompts in multi-step reasoning. We also found that suppressing the activation of these neurons worsened reasoning performance. These results have implications for the mechanisms by which models acquire reasoning ability.

In recent years, generative models using diffusion process have achieved the state-of-the-art performance in the continuous domain and have been actively studied in discrete data generation. In this study, we propose caption generation using a language model and a classifier based on diffusion process. To improve the performance of caption generation, we examine the difference in accuracy with and without a pre-trained language model in the classifier, and investigate under what conditions appropriate captions can be generated for each image. Although the accuracy of our method using diffusion process was not good, we have confirmed that natural language generation could be controlled by the performance of a classifier in the sampling process.

In this paper, we study the effectiveness of several prompting techniques for controlling the formality level of machine translation (MT) using former existing pre-trained Large Language Models (LLM), including GPT-3 and ChatGPT. Our experimental setting includes a selection of state-of-the-art LLMs and uses an En-Ja parallel corpus specifically designed to test formality control in machine translation, and we propose an approach based on machine learning for evaluating the control capabilities of MT models. Overall, our results provide empirical evidence suggesting that our classification-based evaluation works well in practice and that prompting is a viable approach to control the formality level of En-Ja machine translation using LLMs.

Most data-to-text studies use numerical data to generate natural language sentences to describe events in the target domain, but few natural language sentence generation methods have been developed to capture the analytical meaning of numerical data or relationships among multiple numerical data. In this study, we propose a natural language sentence generation method that can capture the relationship between two time-series data sets and various relationships regarding their trends. We created two time-series data sets and experimented with them, and found that the evaluation data could be reproduced with considerable accuracy in the generated natural language sentences.

In machine learning, large amounts of data are needed to improve model performance.However, collecting them is costly, so a technique called data augmentation is used to generate new data from existing data.In natural language processing, there is a text data augmentation technique called round-trip translation,which translates text data into another language and then translates it back into the original language to generate a paraphrase of the original text.However, the round-trip translation is computationally expensive and time-consuming because it requires twice translations for one text. In this paper, we propose a faster text augmentation method using a model trained to make the round-trip translation.The dataset of this training consists of original texts and the results of their round-trip translation.Experimental results show that the proposed method, using the Text-To-Text Transfer Transformer (T5),can augment data at most about 1.6 times faster than round-trip translation.Furthermore, T5 can generate paraphrases not included in the training data based on the knowledge acquired through pretraining.

We study synthetic corpus-based approaches for language models (LMs) to acquire logical deductive reasoning ability. The previous studies trained LMs on synthetically generated examples of deductive reasoning, which have been effective to an extent. However, it has not yet been studied on what aspect of deductive reasoning ability deduction corpora have enhanced LMs. This investigation is essential to discuss the future directions of deductive reasoning. We investigate this by generating and using a comprehensive set of ``ablation corpora'', where one corpus emphasizes a specific aspect different from those emphasized by the other corpora. Finally, on the basis of these results, we discuss the future directions for applying deduction corpora or other approaches for each aspect.

Te-form subordinate clauses are common expressions in Japanese that show multiple usages; thus, it is an important task in natural language processing to automatically determine the usage of these clauses. In this study, we designed an annotation guideline and manually classified the usage of te-form subordinate clauses. Various classifications have been proposed regarding te-form subordinate clauses. However, in attempts to create annotation guidelines, it tends to be difficult for non-linguist to make consistent judgments. Therefore, we design an annotation guideline using "linguistic tests." Linguistic tests include operations such as determining whether a target expression can be paraphrased, which we claim to reduce the variations of judgments. Moreover, we implement and train a neural classifier based on the BERT language model using the annotated corpus, which automatically classify the usage of te-forms.

Natural language sentences have "dynamic" aspects in that their interpretation depends on contexts (e.g., anaphora). In formal semantics, it has been suggested that dependent type theory provides a natural explanation for dynamic phenomena. Since tense also involves context dependence, such type-theoretical formal semantics is a promising approach to temporal interpretation. In this study, we present a type-theoretical analysis of the temporal relations in Japanese relative clauses. We use Dependent Type Semantics, a compositional semantics based on dependent type theory. We demonstrate that our proposed theory explains various examples by dynamically determining the temporal interpretation of relative clauses in the same way as anaphora.

Is there a semantic theory that can robustly predict and explain the accumulated variety of linguistic facts, and is there a way to evaluate whether such a theory is successful? This study addresses these questions by combining Dependent Type Semantics (DTS), a proof-theoretic framework of natural language semantics based on dependent type theory, and the methodology of automatic verification of semantic theories through implementation. This is achieved by combining the technology of 1) a Japanese CCG parser, 2) an automatic theorem prover for dependent type theory, and 3) the inference test set JSeM describing Japanese semantic phenomena. We give an overview of the empirical semantic studies conducted so far under the framework of DTS, and discuss the prospects and challenges in applying the above methodology to those studies.

Dependent type theory has been used to formulate natural language semantics. Natural language semantics based on it provides several methods for natural language inference, making dependent type theory an attractive alternative in natural language semantics. Dependent Type Semantics (DTS) is a natural language semantics based on dependent type theory, and anaphora is one of major semantic phenomena that DTS can account for. Recently, underspecified types have been introduced in DTS. Underspecified types enable to treat anaphora resolution as proof search in dependent type theory, while a method of anaphora resolution is not implemented yet. We aim to implement the anaphora resolution process via underspecified types, by using the proof assistant Coq. We formulate this process by means of Coq's refine tactic. Furthermore, some examples of natural language inference involving anaphora resolution crucially are discussed: we prove these inferences as theorems in Coq.

Constructing linguistically valid CCG treebanks is necessary since CCG parsing often uses CCG treebanks as training and evaluation data. However, it is known that the current Japanese CCG treebank, CCGbank, incorrectly analyzes Japanese syntactic structures, including passive and causative constructions. The ABCTreebank, a treebank for ABC grammar, has made many improvements, such as argument structures. However, it does not describe the detailed syntactic features of Japanese CCG. Meanwhile, the output of the Japanese CCG parser, lightblue, successfully provides the syntactic structures with detailed syntactic features but faces the challenge of capturing the argument structures correctly. In this study, we propose a method to generate a Japanese treebank with more linguistically valid and detailed information by combining the advantages of the ABCTreebank with lightblue. We develop an algorithm to filter lightblue's lexical items using ABCTreebank and construct a linguistically valid CCG treebank by transforming the output of lightblue.

Improving throughput and quality control in the wafer process steps are important elements in semiconductor manufacturing. Wafer transfer schedule optimization has a significant impact on these factors. Wafer process steps are becoming more complex, so the importance of flexible and efficient wafer transfer schedule optimization methods is increasing. Therefore, we have developed a cluster tool simulator that has fewer constraints than real equipment and can be used to verify wafer transfer schedule optimization algorithms. This paper reports the results of the verification of the processing sequence optimization method by the Monte Carlo method with the simulator. The Monte Carlo method is applicable to complex wafer process steps because it does not require evaluation during the transfer process and can be applied if only the final state is evaluated. In addition, we proposed and verified a method to reuse the transfer schedule of some wafers when applying the Monte Carlo method. This enables the creation of schedules with short optimization times in sudden changes, such as chamber failure, or when processing large numbers of wafers of the same type.