ETRI Journal最新文献

The Korean language has written (formal) and spoken (phonetic) forms that differ in their application, which can lead to confusion, especially when dealing with numbers and embedded Western words and phrases. This fact makes it difficult to automate Korean speech recognition models due to the need for a complete transcription training dataset. Because such datasets are frequently constructed using broadcast audio and their accompanying transcriptions, they do not follow a discrete rule-based matching pattern. Furthermore, these mismatches are exacerbated over time due to changing tacit policies. To mitigate this problem, we introduce a data-driven Korean spoken-to-written transcription conversion technique that enhances the automatic conversion of numbers and Western phrases to improve automatic translation model performance.

We focus on open-domain question-answering tasks that involve a chain-of-reasoning, which are primarily implemented using large language models. With an emphasis on cost-effectiveness, we designed EffiChainQA, an architecture centered on the use of small language models. We employed a retrieval-based language model to address the limitations of large language models, such as the hallucination issue and the lack of updated knowledge. To enhance reasoning capabilities, we introduced a question decomposer that leverages a generative language model and serves as a key component in the chain-of-reasoning process. To generate training data for our question decomposer, we leveraged ChatGPT, which is known for its data augmentation ability. Comprehensive experiments were conducted using the HotpotQA dataset. Our method outperformed several established approaches, including the Chain-of-Thoughts approach, which is based on large language models. Moreover, our results are on par with those of state-of-the-art Retrieve-then-Read methods that utilize large language models.

Recent advances in deep learning for speech and visual recognition have accelerated the development of multimodal speech recognition, yielding many innovative results. We introduce a Korean audiovisual speech recognition corpus. This dataset comprises approximately 150 h of manually transcribed and annotated audiovisual data supplemented with additional 2000 h of untranscribed videos collected from YouTube under the Creative Commons License. The dataset is intended to be freely accessible for unrestricted research purposes. Along with the corpus, we propose an open-source framework for automatic speech recognition (ASR) and audiovisual speech recognition (AVSR). We validate the effectiveness of the corpus with evaluations using state-of-the-art ASR and AVSR techniques, capitalizing on both pretrained models and fine-tuning processes. After fine-tuning, ASR and AVSR achieve character error rates of 11.1% and 18.9%, respectively. This error difference highlights the need for improvement in AVSR techniques. We expect that our corpus will be an instrumental resource to support improvements in AVSR.

As automated essay scoring (AES) has progressed from handcrafted techniques to deep learning, holistic scoring capabilities have merged. However, specific trait assessment remains a challenge because of the limited depth of earlier methods in modeling dual assessments for holistic and multi-trait tasks. To overcome this challenge, we explore providing comprehensive feedback while modeling the interconnections between holistic and trait representations. We introduce the DualBERT-Trans-CNN model, which combines transformer-based representations with a novel dual-scale bidirectional encoder representations from transformers (BERT) encoding approach at the document-level. By explicitly leveraging multi-trait representations in a multi-task learning (MTL) framework, our DualBERT-Trans-CNN emphasizes the interrelation between holistic and trait-based score predictions, aiming for improved accuracy. For validation, we conducted extensive tests on the ASAP++ and TOEFL11 datasets. Against models of the same MTL setting, ours showed a 2.0% increase in its holistic score. Additionally, compared with single-task learning (STL) models, ours demonstrated a 3.6% enhancement in average multi-trait performance on the ASAP++ dataset.

This study introduces a new approach in Korean morphological analysis combining dictionary-based techniques with Transformer-based deep learning models. The key innovation is the use of a BERT-based reranking system, significantly enhancing the accuracy of traditional morphological analysis. The method generates multiple suboptimal paths, then employs BERT models for reranking, leveraging their advanced language comprehension. Results show remarkable performance improvements, with the first-stage reranking achieving over 20% improvement in error reduction rate compared with existing models. The second stage, using another BERT variant, further increases this improvement to over 30%. This indicates a significant leap in accuracy, validating the effectiveness of merging dictionary-based analysis with contemporary deep learning. The study suggests future exploration in refined integrations of dictionary and deep learning methods as well as using probabilistic models for enhanced morphological analysis. This hybrid approach sets a new benchmark in the field and offers insights for similar challenges in language processing applications.

Cyclic interleaved frequency division multiplexing (CIFDM), a variant of IFDM, has recently been proposed. While CIFDM employs cyclic interleaving at the transmitter to make multipath components resolvable at the receiver, the current approach of matched filtering followed by multipath combining does not fully exploit the diversity available. This is primarily because the correlation residues among the codes have a significant impact on multipath resolution. As a solution, we introduce a novel multipath successive interference cancellation (SIC) technique for CIFDM, which replaces the conventional matched filtering approach. We have examined the performance of this proposed CIFDM-SIC technique and compared it with the conventional CIFDM-matched filter bank and IFDM schemes. Our simulation results clearly demonstrate the superiority of the proposed scheme over the existing ones.

We obtain the air quality index (AQI) for a descriptive system aimed to communicate pollution risks to the population. The AQI is calculated based on major air pollutants including O₃, CO, SO₂, NO, NO₂, benzene, and particulate matter PM2.5 that should be continuously balanced in clean air. Air pollution is a major limitation for urbanization and population growth in developing countries. Hence, automated AQI prediction by a deep learning method applied to time series may be advantageous. We use a seasonal autoregressive integrated moving average (SARIMA) model for predicting values reflecting past trends considered as seasonal patterns. In addition, a transductive long short-term memory (TLSTM) model learns dependencies through recurring memory blocks, thus learning long-term dependencies for AQI prediction. Further, the TLSTM increases the accuracy close to test points, which constitute a validation group. AQI prediction results confirm that the proposed SARIMA–TLSTM model achieves a higher accuracy (93%) than an existing convolutional neural network (87.98%), least absolute shrinkage and selection operator model (78%), and generative adversarial network (89.4%).

Medical signal processing requires noise and interference-free inputs for precise segregation and classification operations. However, sensing and transmitting wireless media/devices generate noise that results in signal tampering in feature extractions. To address these issues, this article introduces a finite impulse response design based on a two-level transpose Vedic multiplier. The proposed architecture identifies the zero-noise impulse across the varying sensing intervals. In this process, the first level is the process of transpose array operations with equalization implemented to achieve zero noise at any sensed interval. This transpose occurs between successive array representations of the input with continuity. If the continuity is unavailable, then the noise interruption is considerable and results in signal tampering. The second level of the Vedic multiplier is to optimize the transpose speed for zero-noise segregation. This is performed independently for the zero- and nonzero-noise intervals. Finally, the finite impulse response is estimated as the sum of zero- and nonzero-noise inputs at any finite classification.