Recent advancements in Large Vision-Language Models (LVLMs) have significantly expanded their utility in tasks like image captioning and visual question answering. However, they still struggle with object hallucination, where models generate descriptions that inaccurately reflect the visual content by including nonexistent objects or misrepresenting existing ones. While previous methods, such as data augmentation and training-free approaches, strive to tackle this issue, they still encounter scalability challenges and often depend on additional external modules. In this work, we propose Ensemble Decoding (ED), a novel strategy that splits the input image into sub-images and combines logit distributions by assigning weights through the attention map. Furthermore, we introduce ED adaptive plausibility constraint to calibrate logit distribution and FastED, a variant designed for speed-critical applications. Extensive experiments across hallucination benchmarks demonstrate that our proposed method achieves state-of-the-art performance, validating the effectiveness of our approach.

Hyunjong Kim, Suyeon Lee, Yeongjae Cho, Eunseo Ryu, Yohan Jo, Suran Seong, Sungzoon Cho. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Sangyeop Kim, Sohhyung Park, Jaewon Jung, Jinseok Kim, Sungzoon Cho. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Attempts to express information from various documents in graph form are rapidly increasing. The speed and volume in which these documents are being generated call for an automated process, based on machine learning techniques, for cost-effective and timely analysis. Past studies responded to such needs by building knowledge graphs or technology trees from the bibliographic information of documents, or by relying on text mining techniques in order to extract keywords and/or phrases. While these approaches provide an intuitive glance into the technological hotspots or the key features of the select field, there still is room for improvement, especially in terms of recognizing the same entities appearing in different forms so as to interconnect closely related technological concepts properly. In this paper, we propose to build a patent knowledge network using the United States Patent and Trademark Office (USPTO) patent filings for the semiconductor device sector by fine-tuning Huggingface’s named entity recognition (NER) model with our novel edge weight updating neural network. For the named entity normalization, we employ edge weight updating neural network with positive and negative candidates that are chosen by substring matching techniques. Experiment results show that our proposed approach performs very competitively against the conventional keyword extraction models frequently employed in patent analysis, especially for the named entity normalization (NEN) and document retrieval tasks. By grouping entities with named entity normalization model, the resulting knowledge graph achieves higher scores in retrieval tasks. We also show that our model is robust to the out-of-vocabulary problem by employing the fine-tuned BERT NER model.

Detecting anomalies in real-world multivariate time series data is challenging due to complex temporal dependencies and inter-variable correlations. Recently, reconstruction-based deep models have been widely used to solve the problem. However, these methods still suffer from an over-generalization issue and fail to deliver consistently high performance. To address this issue, we propose the MEMTO, a memory-guided Transformer using a reconstruction-based approach. It is designed to incorporate a novel memory module that can learn the degree to which each memory item should be updated in response to the input data. To stabilize the training procedure, we use a two-phase training paradigm which involves using K-means clustering for initializing memory items. Additionally, we introduce a bi-dimensional deviation-based detection criterion that calculates anomaly scores considering both input space and latent space. We evaluate our proposed method on five real-world datasets from diverse domains, and it achieves an average anomaly detection F1-score of 95.74%, significantly outperforming the previous state-of-the-art methods. We also conduct extensive experiments to empirically validate the effectiveness of our proposed model's key components.

Dialogue state tracking for multi-domain dialogues is challenging because the model should be able to track dialogue states across multiple domains and slots. As using pre-trained language models is the de facto standard for natural language processing tasks, many recent studies use them to encode the dialogue context for predicting the dialogue states. Model architectures that have certain inductive biases for modeling the relationship among different domain-slot pairs are also emerging. Our work is based on these research approaches on multi-domain dialogue state tracking. We propose a model architecture that effectively models the relationship among domain-slot pairs using a pre-trained language encoder. Inspired by the way the special [CLS] token in BERT is used to aggregate the information of the whole sequence, we use multiple special tokens for each domain-slot pair that encodes information corresponding to its domain and slot. The special tokens are run together with the dialogue context through the pre-trained language encoder, which effectively models the relationship among different domain-slot pairs. Our experimental results on the datasets MultiWOZ-2.0 and MultiWOZ-2.1 show that our model outperforms other models with the same setting. Our ablation studies incorporate three main parts. The first component shows the effectiveness of our approach exploiting the relationship modeling. The second component compares the effect of using different pre-trained language encoders. The final component involves comparing different initialization methods that could be used for the special tokens. Qualitative analysis of the attention map of the pre-trained language encoder shows that our special tokens encode relevant information through the encoding process by attending to each other.

In this paper, we propose an advanced method to detect abnormal usage patterns for identifying the fine-grained levels of user needs. Most previous studies investigated user need identification based on users textual reviews. Thus, they focused only on the explicit needs of the product levels, and not on the implied needs of the fine-grained levels. Although in a few recent studies the authors attempted to identify user needs based on abnormality detection, they considered only limited elements of the usage sequence, such as touching buttons, and did not consider the important elements, such as the dragging interaction and the pop-up and notification components. Thus, in this study, we considered all the elements of the usage sequence to identify abnormal usage sequences for recognizing user needs at the fine-grained level. Moreover, we utilized the recurrent neural network semantic variational autoencoder (RNN-SVAE) architecture, which is a state-of-the-art architecture for sentence embedding, to represent the usage sequences effectively. In detail, we calculate the vector representation of the entire usage sequence utilizing the RNN-SVAE architecture based on heterogeneous embedding to apply the abnormality detection method for determining abnormal sequences corresponding to user needs. The experimental results verify that our proposed method extracts meaningful abnormal usage patterns that previous approaches do not identify. Additionally, our proposed method shows a higher correlation of the coefficient score between the abnormality score and the importance score of the extracted sequences than do previous approaches.

There has recently been a significant movement toward aiding graphic design tasks based on artificial intelligence or machine learning. In addition, colorization plays an important role within the topic of GUI design. Previous studies regarding automatic colorization have focused on a consideration of the realistic aspects of an image without consideration of the design semantics or usability, which are critical aspects for a practical GUI design. We, therefore, propose an end-to-end network for a generative combination of color sets for a GUI design based on the design semantics, while utilizing thousands of actual GUI design datasets acquired from LG Electronics to train the network. By utilizing the GUI design dataset, our network effectively generates color sets for a GUI design by considering various design aspects, such as the usability factors. In detail, we concatenate the textual design concept, characteristics of the application, and usage frequency for the elements of the design semantics. We then construct a conditional generative adversarial net processing of the design semantics as a condition to generate suitable color sets and construct the GUI design based on these sets. The experiments indicate that our proposed method effectively generates color sets for a GUI design based on the design semantics. In addition, our proposed method shows a better score than other methods on a user test conducted to verify the practicality, perception, recognition, diversity, and esthetic features. Moreover, experimental results prove that users can effectively grasp the intended design concept of our generated GUI design with higher top-1, top-2, and top-3 levels of accuracy.

In automated material handling systems (AMHS), delivery time is an important issue directly associated with the production cost and the quality of the product. In this paper, we propose a dynamic routing strategy to shorten delivery time and delay. We set the target of control by analyzing traffic flows and selecting the region with the highest flow rate and congestion frequency. Then, we impose a routing cost in order to dynamically reflect the real-time changes of traffic states. Our deep reinforcement learning model consists of a Q-learning step and a recurrent neural network, through which traffic states and action values are predicted. Experiment results show that the proposed method decreases manufacturing costs while increasing productivity. Additionally, we find evidence the reinforcement learning structure proposed in this study can autonomously and dynamically adjust to the changes in traffic patterns.

Many researches have exploited textual data, such as news, online blogs, and financial reports, in order to predict stock price movements effectively. Previous studies formed the task as a classification problem predicting upward or downward movement of stock prices from text documents. Such an app roach, however, may be deemed inappropriate when combined with sentiment analysis. In financial documents, same words may convey different sentiments across different sectors; if documents from multiple sectors are learned simultaneously, performance can deteriorate. Therefore, we conducted sentiment analysis of 8-K financial reports of firms sector by sector. In particular, we also employed distributed representation for predicting stock price movements. Experiment results show that our approach improves prediction performance by 25.4% over the baseline model, and that the direction of post-announcement stock price movements shifts accordingly with the polarity of the sentiment of reports. Not only does our model improve predictability, but also provides visualizations, which may assist agents actively trading in the field with understanding the drivers for the observed stock movements. The two main aspects of our model, predictability and interpretability, will provide meaningful insights to help decision-makers in the industry with time-split trading decisions or data-driven detection of promising companies.

The classification of customer-voice data is an important matter in real business since it is necessary for customer-voice data to be delivered to relevant departments and responsible individuals. Additionally, customer-voice data typically includes several novel words, such as typo’s, informal ter ms, or exceedingly general words to discriminate between categories of customer-voice data. Furthermore, noisy data often has a negative effect on the classification task. In this study, advanced novelty detection method is proposed to utilize class vector that possessed high cosine similarity with words to effectively discriminate between classes. The class vector is considered as the centroid or the mean of each word vector distribution as derived from the neural embedding model, and the novelty score of each word is calculated and novel words are effectively detected. Each novelty score is calculated by improvements of GMM and KMC methods utilizing a class vector. The experiments verify the propriety of the proposed method with qualitative observations, and the application of the proposed method with quantitative experiments verifies the representational effectiveness and classification performance of customer-voice data. The experiment results indicate that the performance of a classification of customer-voice data improved with the application of the newly proposed novelty detection method in this study.

Industry classification has served as an important tool for sector analysis in capital market research. Most of the existing classification schemes that are commonly used in academia or industry research require human input in the process of its design and construction. Because compilation and maintenance of such schemes demand comprehensive domain knowledge, it is highly costly and exhaustively time-consuming to update them to correctly reflect the fast-changing trends of the market. In addition, due to its subjective nature, interpretability is limited. As a remedy to such shortcomings, this paper adds to our earlier work, Business Text Industry Classification (BTIC) and proposes a new classification scheme, Business Text Industry Network (BTIN), namely, that automatically produces industry groupings based on the textual information from the corporate disclosures and news articles. BTIN exploits the business section of the Form 10-Ks, in which firms describe their business identities in a rich context, as well as the contents of news articles, which represent market perception of the subject firm. We employ Doc2vec for document embedding and a novel news curation algorithm to represent the relationship among firms in a graph structure, and then apply Louvain method to detect communities ??? or, industries ??? from the network and categorize firms into BTIN groups. Evaluation results using market ratios show that BTIN performs quite competitively against SIC, GICS, and its predecessor, BTIC, in terms of clustering fitness, inter-industry heterogeneity, and intra-industry homogeneity.

Industry classification has been rigorously utilized in academic research and business analytics. The existing classification schemes, however, have been constructed and maintained manually by domain experts, which require exhaustive time and human effort while vulnerable to subjectivity. Hence, the existing classification systems do not properly reflect the fast-changing trends of the firms and the capital market. As a remedy to such shortcomings, this paper proposes a new classification scheme, Business Text Industry Classification (BTIC), namely, that automatically clusters securities based on the textual information from the corporate disclosures. BTIC exploits the business section of the Form 10-Ks, in which firms provide their self-identities in a rich context. We employ doc2vec for document embedding and apply Ward's hierarchical clustering method to categorize securities into BTIC groups. Evaluation results using 12 financial ratios commonly found in financial research show that BTIC performs just as good as SIC and GICS in terms of inter- and intra-industry homogeneity, especially for the higher level of clustering. Given that, we claim that BTIC outperforms SIC and GICS in four aspects: process automation, objectivity, clustering flexibility, and result interpretability.

The existing multi-class classification algorithms based on support vector machine (SVM) generally decompose the original problem into smaller subproblems. However, the decomposition approach raises the problems of unreliable and unbalanced training of individual classifiers. The aim of this work is to alleviate these problems. In this paper, a novel multi-class classification algorithm based on one-class SVM is presented. The distinguishing point of our proposed algorithm is that the algorithm solves a one-class classification problem rather than decomposing the problem into several smaller subproblems. The proposed algorithm solves a multi-class classification problem by expanding the training input vector with the class label and constructing a one-class SVM with the expanded input vectors. In the test phase, test input vectors are assessed by fit to the decision boundary under the assumption of belonging to a class. We conducted experiments on several real-world benchmark datasets. Experimental results showed that the proposed algorithm outperforms other SVM-based multi-class classification algorithms.

We proposed a kernel-based binary classification algorithm, named support vector class description (SVCD), which is an extended version of support vector domain description (SVDD) for one-class classification. SVCD constructs two compact hyperspheres

The support vector machine (SVM) has been spotlighted in the machine learning community because of its theoretical soundness and practical performance. When applied to a large data set, however, it requires a large memory and a long time for training. To cope with the practical difficulty, we propose a pattern selection algorithm based on neighborhood properties. The idea is to select only the patterns that are likely to be located near the decision boundary. Those patterns are expected to be more informative than the randomly selected patterns. The experimental results provide promising evidence that it is possible to successfully employ the proposed algorithm ahead of SVM training.

Outliers refer to "minority" data that are different from most other data. They usually disturb data mining process. But, sometimes they provide valuable information. Thus, it is important to identify outliers in a given data set. In this paper, we propose a novel approach which scores "outlierness" based on the distance from majority data. First, prototype data are identified. Second, those prototypes that are distant from others are eliminated. Finally, the outlierness of each data point is computed as the distance from the remaining prototypes. Experiments involving various data sets show that the proposed approach performs well in terms of accuracy, robustness and versatility.

Support Vector Machine has been well received in machine learning community with its theoretical as well as practical value. However, since its training time complexity is cubic, its use is limited in data mining involving problems with a huge pattern set with a cubic time complexity of its training time. In this paper, we propose a pattern selection method for Support Vector Regression (SVR), using notions of sparseness, variability and uniqueness. Two versions of algorithms, deterministic and stochastic, are presented, which are then applied to an artificial data set and two well known real world data sets. Preliminary results justify further investigation. The proposed method should work well with non-SVM function approximators such as neural networks.

Empirical bankruptcy prediction models have been proposed and widely used in the last decades or so. Historic solvent and default firm data are collected and labeled appropriately. Statistical and neural network models are then "trained" to fit these data. A major problem is the imbalance of data, i.e. much more solvent data than default data. We propose an auto-associative neural network (AANN) that learns the identity mapping of input. By training the network with only solvent data, we built a bankruptcy predictor with better accuracies than conventional 2-class neural network based predictor.

This paper reports the results of a new neural network based trend detector. An auto-associative neural network was trained with the "trend" data obtained from the intra-day KOSPI 200 future price. It was then used to predict a trend. Simple investment strategies based on the detector achieved a one-year return of 31.2 points with no leverage.

Training support vector classifiers (SVC) requires large memory and long cpu time when the pattern set is large. To alleviate the computational burden in SVC training, we previously proposed a preprocessing algorithm which selects only the patterns in the overlap region around the decision boundary, based on neighborhood properties. The k-nearest neighbors' class label entropy for each pattern was used to estimate the pattern's proximity to the decision boundary. The value of parameter k is critical, yet has been determined by a rather ad-hoc fashion. We propose in this paper a systematic procedure to determine k and show its effectiveness through experiments.

Password is the most widely used identity verification method in computer security domain. However, due to its simplicity, it is vulnerable to imposter attacks. Keystroke dynamics adds a shield to password. Password typing patterns or timing vectors of a user are measured and used to train a novelty detector model. However, without manual pre-processing to remove noises and outliers resulting from typing inconsistencies, a poor detection accuracy results. Thus, in this paper, we propose an automatic feature subset selection process that can automatically selects a relevant subset of features and ignores the rest, thus producing a better accuracy. Genetic algorithm is employed to implement a randomized search and SVM, an excellent novelty detector with fast learning speed, is employed as a base learner. Preliminary experiments show a promising result.

This paper proposes the use of multilayer perceptron for brain dysfunction diagnosis. The performance of MLP was better than that of Discriminant Analysis and Decision Tree classifiers, with an 85% accuracy rate in an experimental test involving 332 subjects. In addition, the neural network employing Bayesian learning was able to identify the most important input variable. These two results demonstrate that the neural network can be effectively used in the diagnosis of children with brain dysfunction.

In novelty detection, one tries to discriminate abnormal patterns from normal patterns. As a two class pattern classification problem, novelty detection is quite difficult since in practice only normal patterns are available for training. Novel or abnormal patterns are very few or not available at all. Recently, an auto-associative MLP (AaMLP) has been shown to give a good performance. In this paper, we analyze the output characteristics of trained AaMLPs and show that the AaMLP is indeed a reliable solution for novelty detection. In particular, we prove why nonlinearity in the hidden layer is necessary for novelty detection.

In this paper, we propose an ensemble learning algorithm, which we call the observational learning algorithm (OLA), motivated from the observational learning theory suggested by Bandura (1971). According to the theory, in a group of children who have different and insufficient knowledge for a task, each child can lean how to do the task by observing other children. In the OLA, a neural network ensemble is regarded as a group of children. Each network is trained with the virtual data that are generated from observing other networks as well as the bootstrapping data from the original data set. The virtual data function as both temporal hints having the auxiliary information about the target function and a regularization penalty for making networks smooth. From numerical experiments involving both regression and classification problems, the OLA is shown to give better generalization performance than simple committee and bagging approaches when insufficient and noisy data are given.

In this paper, we report our investigation on the behavior of some Bayesian belief network learning methods, when applied to realistic case data that do not accurately reflect the probability distribution of the underlying domain. The investigation focuses on two types of such data of practical interest, namely data excluding null or normal cases and data contaminated by noise, and their effects on the performance of two learning methods: the K2 algorithm, a representative method of Bayesian approach, and the extended Hebbian learning (EHL) method, representing those of neural network approach. Both analytical and experimental results show that when null cases are removed from the case database, the EHL method is able to accurately construct the underlying causal structure; whereas K2 algorithm may fail to properly handle root nodes. On the other hand, the EHL method is noise sensitive while K2 has certain inherent noise-resistance capability. Suggestions to extend both K2 and EHL to better cope with these problems are also made, and their effectiveness tested through a systematic computer simulation experiment. ©1999 John Wiley & Sons, Inc.

It has been claimed that rotation of neuronal population vectors in the primary motor cortex during tasks involving transformation of movement directions provides direct evidence that a mental rotation algorithm is being used. An alternative explanation offered here, the 'asynchronous superposition hypothesis', asserts that the apparent rotation can result from the summation of two stationary vectors, one in the stimulus direction and one in the target movement direction, which vary in length over time. Computer simulations demonstrate the surprising result that the asynchronous superposition hypothesis can produce activation of motor corex cells with intermediate preferred directions, something previously assumed to confirm a mental rotation algorithm. Simulations also demonstrate that the asynchronous superposition hypothesis accounts for some aspects of the data more naturally than does an explanation based on a mental rotation algorithm, and suggest experimental tests that can distinguish between these two hypotheses.

The cold rolling mill process in steel works uses stands of rolls to flatten a strip to a desired thickness. The accurate prediction of roll force is essential for product quality. Currently, a suboptimal mathematical model is used. We trained two multilayer perceptrons, one to directly predict the roll force and the other to compute a corrective coefficient to be multiplied to the prediction made by the mathematical model. Both networks were shown to improve the accuracy by 30-50%. Combining the two networks and the mathematical model results in systems with an improved reliability.

Current understanding of feature maps in proprioceptive cortex is quite limited. To complement experimental studies, we developed a computational model of map formation in proprioceptive cortex. Muscle length and tension from six muscle groups controlling the position of a model arm in three-dimensional space served as input to the simulated cortex. The resultant feature map consisted of regularly spaced clusters of cortical columns representing individual muscle lengths and tensions. Cortical units became tuned to plausible combinations of tension and length, and multiple representations of each muscle group were present. The map was organized such that compact regions within which all muscle group lengths and tensions are represented could be identified. Most striking was the observation that, although not explicitly present in the input, the cortical map developed a representation of the three-dimensional space in which the arm moved. These findings represent testable predictions about proprioceptive cortex, and may also help clarify some organizational issues concerning primary motor cortex.

Competitive activation mechanisms introduce competitive or inhibitory interactions between units through functional mechanisms instead of inhibitory connections. A unit receives input from another unit proportional to its own activation as well as to that of the sending unit and the connection strength between the two. This, plus the finite output from any unit, induces competition among units that receive activation from the same unit. Here we present a backpropagation learning rule for use with competitive activation mechanisms and show empirically how this learning rule successfully trains networks to perform an exclusive-OR task and a diagnosis task. In particular, networks trained by this learning rule are found to outperform standard backpropagation networks with novel patterns in the diagnosis problem. The ability of competitive networks to bring about context-sensitive competition and cooperation among a set of units proved to be crucial in diagnosing multiple disorders.