Converting Korean natural-language avionics test procedures into an executable XML domain-specific language (DSL) is a labor-intensive bottleneck, yet neither automation extreme is acceptable in this safety-critical setting: end-to-end large language model (LLM) generation is unauditable and emits schema-violating values, while purely deterministic extraction leaves externally-grounded slots unsolved. We ask where, and how sparingly, an LLM should be invoked. From a 64-slot schema we derive a data-computed, five-class evidence-grounding slot taxonomy and build a confidence-gated selective hybrid: a deterministic tier resolves every slot whose evidence is lexical or catalogue-grounded, and an on-premises retrieval-augmented LLM is invoked only for residual open-grounded slots with low retrieval confidence, under a gate threshold selected on a held-out validation split. On a leakage-controlled clean test split of 1,705 unique procedures (drawn from 213,013 industrial pairs after removing 11.4× duplication and 7.4% near-duplicate leakage), the hybrid is statistically equivalent to the pure on-premises LLM (59.4% instance Exact Match; paired 90% CI ±0.7 pp within a ±2 pp margin) while invoking the LLM on only 37% of instances, keeping 90% of emitted slots deterministic and audit-logged, and eliminating closed-vocabulary schema violations by construction (0 vs. 0.89%). Instantiating the same gate with a stronger BM25 backend exceeds the pure LLM (60.4%, McNemar p=0.044) at a 14% call rate. Because strict full automation remains below deployment thresholds, we position the system as an authoring-support tool: on a 1,879-instance candidate-ranking analysis, the correct value appears in the top-five list for 84.5% of grounded slots (mean reciprocal rank 0.953), supporting a human-in-the-loop workflow in which a reviewer controls every emitted action.We conclude that a taxonomy-driven selective hybrid is simultaneously accurate, LLM-frugal, and supportive of DO-330 tool-qualification evidence generation.

Table question answering has been studied using datasets drawn from a variety of tabular sources and task formats. However, most publicly available resources have been created in high-resource languages such as English. For low-resource languages, researchers are often required to construct new datasets or translate existing ones, which incurs substantial time, effort, and financial cost. In contrast to natural language text, table data consists of structured entries whose interpretation is less affected by language-specific syntax or word order. In this work, we present a cost-effective strategy for multilingual table QA that relies on selectively translating only the questions of existing datasets. Leveraging the language-agnostic structure of tables, our approach maintains the original table content while translating queries into multiple target languages. To address possible performance drops caused by using table data in the source language rather than the target language, we apply cross-lingual adaptation techniques using contrastive learning and adversarial training. In addition, to strengthen reasoning ability while avoiding degradation in languages not seen during pre-training, we perform supplementary pre-training of a RoBERTa-based multilingual encoder with SQL-derived table data. Finally, we extend our investigation beyond encoder-based architectures and evaluate decoder-only large language models under the same multilingual table QA setting. The experiments show that LLaMA-3 models exhibit strong cross-lingual generalization even without using translated table context and often achieve competitive performance using only Korean table data. Moreover, the performance gap among training configurations such as translated queries or translated datasets is notably smaller compared to encoder-based models, highlighting the inherent multilingual robustness of modern LLMs. We further evaluate LLaMA-3 models on domain-specific table datasets and observe that domain knowledge acquired from Korean tables transfers effectively across languages even without multilingual supervision, underscoring the potential of LLMs for specialized multilingual table reasoning. These findings demonstrate that LLMs can serve as an effective alternative for multilingual table QA, particularly in low-resource or partially translated environments.

Recent advances in deep learning have highlighted the importance of Grapheme-to-Phoneme (G2P) conversion in natural language processing and speech synthesis. Korean exhibits complex phonological changes such as liaison, initial sound law, and consonant assimilation, making it challenging to handle all exceptional patterns with simple models alone. In this paper, we propose integrating an autoregressive (AR) model with a phonological knowledge base that leverages standard pronunciation rules and a pre-analyzed dictionary, and conduct systematic comparisons with non-autoregressive (NAR) variants to validate the effectiveness of sequential processing for Korean G2P. Experimental results show that a BiLSTM-LSTM AR model based on ELECTRA embeddings achieves a phoneme error rate (PER) of 0.2%, word error rate (WER) of 0.68%, and a sentence accuracy of 95.16%, outperforming both a traditional rule-based approach (24.51% sentence accuracy) and non-autoregressive variants implemented on the same dataset (achieving 81.32%-85.72% sentence accuracy). When syllable constraints are introduced, accuracy significantly improves to 95.41% (p <0.05). Meanwhile, incorporating a pre-analyzed dictionary enhances the handling of neologisms and proper nouns without substantially increasing inference time. Through comprehensive same-dataset comparisons between AR and NAR approaches, our study demonstrates the effectiveness of combining AR models with phonological knowledge bases in Korean G2P, with AR models consistently outperforming NAR variants by 3.12-8.04% in sentence accuracy, while identifying key challenges including rule conflicts and handling multiple standard pronunciations. The methodology and findings provide a generalizable framework for other morpho-phonologically complex languages.

Table question answering datasets have been released in various table sources and diverse tasks. However, most of these table question answering datasets are constructed in resource-rich languages like English. For low-resource languages, it is necessary to create new datasets or translate existing ones to train models on table question answering tasks. This process requires substantial costs and manpower. Table data, unlike natural text, is structured in a way that is less affected by word order and language-specific characteristics. In this paper, we propose a method to minimize the costs associated with translating and applying new table question answering datasets. By leveraging the characteristics of table data, we suggest translating only the question data of the question answering table datasets into multiple languages. To minimize the performance degradation caused by using the native table data of the original dataset and not directly utilizing the table data in the target language, we applied adaptation methods using contrastive learning and adversarial training. To enhance the reasoning capabilities required for table question answering without diminishing performance in languages other than the one used for pre-training, we further pre-trained a RoBERTa-based multilingual encoder on table data using SQL data. The proposed cross-lingual adaptation demonstrated performance improvements in most of the languages evaluated. Additionally, the pre-training using SQL data significantly improved performance across all languages, even though the pre-training table data consisted of English data.

Previous works with auto-regressive pre-trained language model achieved state-of-art results in semantic parsing tasks. On the other hand, these models generate answers directly, it is difficult to get selected cells for generated answers and answers that require numerical reasoning. In this paper, we present a self-training framework to generate logical forms with weakly supervised question answering. Self-training framework composed of the training procedure and pseudo label correction procedure. In the training procedure, we train models to generate column and row positions of answer cells with supervised training data whose tagged answer cells are a subset of the input table cells. In the pseudo label correction procedure, we correct the tagged dataset for training the model. Furthermore, to enhance the correcting accuracy, we employed reinforcement learning, incorporating a reward function that yields a greater reward as the computed correct answer approaches the benchmark correct answer. Ultimately, we observed an improvement in accuracy from 47.1 to 54.06, a 6.89 increase, using the tapex-base model. Furthermore, we managed to outperform the existing performance, achieving 64.17, using the OmniTab model. Thus, our proposed method, SWING, a self-learning table parsing framework, has shown improved performance in numerical reasoning cases, and can predict rationale cells that allow auto-regressive PLM to compute answers without the need for logical form.

In machine reading comprehension, answers to questions may be found in either text or tables. Previous studies have shown that table-specific pre-trained language models perform well when applied to tables; however, applying such models to input data that consists of both tables and text can be challenging. To address this issue, we introduce the hybrid reader model that can manage both tables and text using a modified K-Adapter architecture for effectively encoding the structured information of tables. The training process infuses knowledge for tabular data into a pre-trained model while retaining its original weights from pre-training. Hence, the pre-trained model is able to learn and utilize table information without sacrificing its previous training. Our proposed hybrid reader model achieved comparable or superior performance to that of a specialized model on the Korean MRC dataset, KorQuAD 2.0, using the provided adapters. Furthermore, we conducted experiments on an additional English MRC dataset and confirmed that our proposed model achieves performance comparable to that of the existing model. Our study indicates that employing a single hybrid model instead of two separate models can require fewer computing resources and less time while achieving comparable or superior performance, especially for techniques that apply projection and adapter.

본 논문은 자연어 처리 분야의 중요한 주제인 철자 오류 교정에 초점을 맞춘다. 연구는 Generative Pre-trained Transformer(GPT) 모델을 활용하여 철자 오류 교정 방법을 실험적으로 탐구한다. 특히, GPT-1 및 InstructGPT 모델에서 적용된 지도식 미세조정 방법론을 사용하여 철자 오류 교정의 성능을 향상시키는 것을 목표로 한다. 본 논문에서의 지도식 미세조정은 인간의 직접적인 피드백을 활용하여 모델의 성능을 향상시키는 방법을 활용한다. 그리고 다양한 학습 프롬프트 전략을 도입하여 철자 오류 교정 모델의 성능을 더욱 높이고자 했다. 이러한 전략은 양방향 문맥정보를 활용한 학습 방법으로, 실험 결과 GPT-2 기본 모델 대비 F1 점수에서 상당한 향상을 보였으며, 이전 연구 결과보다도 더 높은 성능을 달성했다. 본 연구는 철자 오류 교정을 위한 GPT 모델의 활용 및 지도식 미세조정 방법론과 학습 프롬프트 전략의 유효성을 입증하며, 이 분야의 연구를 한 단계 발전시킬 가능성을 제시한다.

The objective of this study was to create graph embedding vectors using Korean WordNet (KorLex) and apply them to neural network word-embedding models. Semantic knowledge, especially lexical semantic knowledge in a language, can be represented by word-embedding vectors or graph structures of lexical databases, such as WordNet. Both representations capture common semantics; however, some semantic knowledge is only captured in a specific way or not at all. In a previous study, Path2vec mapped WordNet graphs to graph-embedding vectors using similarity scores between two words. In this study, we propose two main approaches. First, we mapped the knowledge in the Korean lexical database KorLex onto graph-embedding vectors. We then applied these embedding vectors to deep neural network word embeddings to capture additional semantic knowledge in the Korean language. On a custom test set, the proposed approach improved performance by capturing additional semantic knowledge in similarity and analogy analyses. We plan to apply a variant of this to other deep neural embedding models.

The rapid advancement of large language models (LLMs) has opened up new possibilities for various natural language processing tasks. This study explores the potential of LLMs for author profiling in digital text forensics, which involves identifying characteristics such as age and gender from writing style—a crucial task in forensic investigations of anonymous or pseudonymous communications. Experiments were conducted using state-of-the-art LLMs, including Polyglot, EEVE, and Bllossom, to evaluate their performance in author profiling. Different fine-tuning strategies, such as full fine-tuning, Low-Rank Adaptation (LoRA), and Quantized LoRA (QLoRA), were compared to determine the most effective methods for adapting LLMs to the specific needs of this task. The results show that fine-tuned LLMs can effectively predict authors’ age and gender based on their writing styles, with Polyglot-based models generally outperforming EEVE and Bllossom models. Additionally, LoRA and QLoRA strategies significantly reduce computational costs and memory requirements while maintaining performance comparable to full fine-tuning. However, error analysis reveals limitations in the current LLM-based approach, including difficulty in capturing subtle linguistic variations across age groups and potential biases from pre-training data. These challenges are discussed and future research directions to address them are proposed. This study underscores the potential of LLMs in author profiling for digital text forensics, suggesting promising avenues for further exploration and refinement.

Named entity recognition (NER) in natural language processing encompasses three primary types: flat, nested, and discontinuous. While the flat type often garners attention from researchers, nested NER poses a significant challenge. Current approaches to addressing nested NER involve sequence labeling methods with merged label layers, cascaded models, and those rooted in reading comprehension. Among these, sequence labeling with merged label layers stands out for its simplicity and ease of implementation. Yet, highlighted issues persist within this method, prompting our aim to enhance its efficacy. In this study, we propose augmentations to the sequence labeling approach by employing a pipeline model bifurcated into sequence labeling and text classification tasks. Departing from annotating specific entity categories, we amalgamated types into main and sub-categories for a unified treatment. These categories were subsequently embedded as identifiers in the recognition text for the text categorization task. Our choice of resolution involved BERT+BiLSTM+CRF for sequence labeling and the BERT model for text classification. Experiments were conducted across three nested NER datasets: GENIA, CMeEE, and GermEval 2014, featuring annotations varying from four to two levels. Before model training, we conducted separate statistical analyses on nested entities within the medical dataset CMeEE and the everyday life dataset GermEval 2014. Our research unveiled a consistent dominance of a particular entity category within nested entities across both datasets. This observation suggests the potential utility of labeling primary and subsidiary entities for effective category recognition. Model performance was evaluated based on F1 scores, considering correct recognition only when both the complete entity name and category were identified. Results showcased substantial performance enhancement after our proposed modifications compared to the original method. Additionally, our improved model exhibited strong competitiveness against existing models. F1 scores on the GENIA, CMeEE, and GermEval 2014 datasets reached 79.21, 66.71, and 87.81, respectively. Our research highlights that, while preserving the original method’s simplicity and implementation ease, our enhanced model achieves heightened performance and competitive prowess compared to other methodologies.

사전학습 언어모델을 활용하면 다양한 자연어처리 태스크를 수행하기 수월해진다, 일반적으로 태스크별로 미세조정 과정을 통해 추가로 모델을 학습시켜 사용하게 되는데, 사전학습 언어모델 구조의 특성상 다수의 딥러닝 계층이 상호작용하면서 학습이 진행된다. 인접한 계층들은 순차적으로 언어를 분석하며 이해하고 태스크를 학습하여 수행한다. 이때 모델의 언어 능력의 향상을 위해 추가적인 지식 주입을 한다면, 태스크에 대한 모델의 성능 향상에 긍정적인 영향을 줄 수 있다.BR 본 논문에서는 기존 미세조정 방식에 대한 새로운 패러다임을 제안한다. 이전 연구에 따르면, 모델이 데이터를 학습할 시에 모델 내 상위 계층과 하위 계층이 다른 역할을 수행할 수 있음을 확인하였다. 이를 기반으로 하위 계층에 통사적인 지식을 주입한 후에, 주입된 지식을 바탕으로 의미론적인 태스크를 미세조정하는 모델 학습 방식을 제시한다. 본 연구에서는 제안 방법을 세 가지 사전학습 언어 모델인 KLUE-BERT, KLUE-RoBERTa, KoELECTRA 로의 적용을 통해, KLUE benchmark의 MRC, NLI, STS 세 가지 태스크에서 최대 1.2%p의 성능 향상을 확인할 수 있었다. 제시한 방법론은 미세조정 과정에 대해 한층 유연한 통찰을 줄 수 있으며, 주입된 정보를 더욱 효율적으로 다룰 수 있는 기법을 통해 추가적인 성능 향상을 기대할 수 있을 것이다.

표 데이터는 일반적인 텍스트 데이터와 다르게 구조적인 특장점으로 정보를 압축해 표현할 수 있다. 이는 표가 다양한 도메인에서 활용되는 것으로 이어지며, 기계독해 영역에서의 표 기계독해 능력이 차지하는 비중은 점점 커지고 있다. 하지만 도메인마다 표의 구조와 요구되는 지식이 달라 언어 모델을 단일 도메인으로 학습했을 때 다른 도메인에서의 모델의 평가 성능이 하락해 일반화 성능이 낮게 나타날 가능성이 크다. 이를 극복하기 위해서는 다양한 도메인의 데이터셋 구축이 우선이 되어야 하며, 단순 사전학습한 모델이 아닌 다양한 기법을 적용하는 것이 중요하다. 본 연구에서는 도메인 일반화 성능을 높이기 위해 도메인 간 불변하는 언어적 특성(Invariant-feature)을 학습하는 언어 모델을 설계한다. 각 도메인별 평가 데이터셋에서의 성능을 높이기 위해서 적대적 학습을 이용하는 방법과 표 데이터에 특화된 임베딩 레이어와 트랜스포머 레이어를 추가하는 모델의 구조를 변형하는 방법을 적용하였다. 적대적 학습을 적용했을 때는 표와 관련된 특화된 임베딩을 추가하지 않는 구조의 모델에서 성능이 향상되는 것을 확인했으며, 표에 특화된 트랜스포머 레이어를 추가하고 추가된 레이어가 표에 특화된 임베딩을 추가로 입력받도록 했을 때, 모든 도메인의 데이터에서 가장 향상된 성능을 보였다.

소셜 네트워크 서비스(SNS)를 이용한 의사소통이 폭발적으로 증가함에 따라 메신저 기능을 통해 텍스트 데이터가 방대하게 발생하고 있다. 반면 최근 자연어 처리(Natural Language Processing) 분야의 발전으로 감성 분류, 욕설 탐지, 챗봇 등 다양한 애플리케이션이 개발되어 제공되고 있으나, 한국어 구어체 텍스트에서 발화자의 성별, 연령대와 같은 저자의 다양한 특징을 분류하려는 시도는 전무한 상황이다. 본 연구에서는 한국어 구어체를 활용하여 저자 프로파일링을 위한 성별 분류 모델을 제안한다. 발화자의 성별 분류를 위해 카카오톡 대화 데이터를 기반으로, 한국어 댓글로 학습한 KcBERT(Korean Comments BERT)에 일상대화와 유사한 '네이트판(Nate Pan)' 데이터를 추가로 학습하여 Domain Adaptation을 진행한다. 그 후 어휘 외적인 정보를 결합한 모델로 실험한 결과 약 95%의 정확도를 달성하여 성능이 향상 됨을 보였다. 본 연구에서는 Domain Adaptation을 위해 자체 수집한 '네이트판(Nate Pan)' 데이터 세트와 국립국어원 제공 데이터 세트를 활용하고, 모델의 학습과 평가를 위해서 AI HUB의 '한국어 SNS' 데이터 세트를 이용한다.

In this study, we aim to automatically construct a test dataset for testing the performance of spelling error correction systems. The Google Web 1T corpus, which includes data on 10 quadrillion phrases, is used for this purpose. Therefore, error words used in the test dataset use error words generated by real web users. There are seven types of error words. In order to obtain the error word, a word set that appears simultaneously with the surrounding context (3-g range) of the location of the error word generation is searched. In this calculation, we exclude error words with wide edit distances that cause the resolution of original words to become exceedingly difficult. In order to select the final error word from the word set, a word with a high value is selected by calculating the context probability using 3-g. In the experiment, the performance was measured for two systems (grammarly, MS Word) in service and the recently announced spelling error correction system (Neuspell). The highest performance was the F1 score of 56%, which shows the overall performance, indicating the need for research on spelling errors.

본 논문에서는 한국어 수사를 위한 기계 독해 데이터 세트 구축 방법을 제안한다. 구축한 데이터 세트를 BERT 기반 한국어 질의응답 언어모형에 추가로 훈련하여, 수사를 포함한 질의응답의 성능 개선을 가져왔다. 데이터 세트 구축을 위해 대한민국 법령문서를 이용하였고, 해당 문서의 구조적 정보를 반영하기 위한 가공 절차를 거쳤다. 질의응답 생성 절차 마련을 위해 한국어 물음말, 한국어 분류사, 국제단위계 및 범용 단위를 수사와 결합하는 방식을 활용했다. 생성한 데이터 세트의 일반화를 알기 위해 전자신문 기사를 포함한 평가 데이터 세트로 실험을 수행했다. 실험 결과, 해당 데이터 세트를 훈련한 시스템의 성능은 EM 83.4, F1 91.0으로, 기존 시스템보다 EM 14.0, F1 9.6 향상되었다. 데이터 세트 평가를 위해 KorQuAD 검증 데이터 세트 중 수사를 포함한 데이터를 무작위로 추출하여 실험했고, 이 또한 기존 시스템보다 EM 13.0, F1 7.2 향상된 성능을 보였다.

Speech processing technology has great potential in the medical field to provide beneficial solutions for both patients and doctors. Speech interfaces, represented by speech synthesis and speech recognition, can be used to transcribe medical documents, control medical devices, correct speech and hearing impairments, and assist the visually impaired. However, it is essential to predict prosody phrase boundaries for accurate natural speech synthesis. This study proposes a method to build a reliable learning corpus to train prosody boundary prediction models based on deep learning. In addition, we offer a way to generate a rule-based model that can predict the prosody boundary from the constructed corpus and use the result to train a deep learning-based model. As a result, we have built a coherent corpus, even though many workers have participated in its development. The estimated pairwise agreement of corpus annotations is between 0.7477 and 0.7916 and kappa coefficient (K) between 0.7057 and 0.7569. In addition, the deep learning-based model based on the rules obtained from the corpus showed a prediction accuracy of 78.57% for the three-level prosody phrase boundary, 87.33% for the two-level prosody phrase boundary.

Supervised disambiguation using a large amount of corpus data delivers better performance than other word sense disambiguation methods. However, it is not easy to construct large-scale, sense-tagged corpora since this requires high cost and time. On the other hand, implementing unsupervised disambiguation is relatively easy, although most of the efforts have not been satisfactory. A primary reason for the performance degradation of unsupervised disambiguation is that the semantic occurrence probability of ambiguous words is not available. Hence, a data deficiency problem occurs while determining the dependency between words. This paper proposes an unsupervised disambiguation method using a prior probability estimation based on the Korean WordNet. This performs better than supervised disambiguation. In the Korean WordNet, all the words have similar semantic characteristics to their related words. Thus, it is assumed that the dependency between words is the same as the dependency between their related words. This resolves the data deficiency problem by determining the dependency between words by calculating the χ2 statistic between related words. Moreover, in order to have the same effect as using the semantic occurrence probability as prior probability, which is used in supervised disambiguation, semantically related words of ambiguous vocabulary are obtained and utilized as prior probability data. An experiment was conducted with Korean, English, and Chinese to evaluate the performance of our proposed lexical disambiguation method. We found that our proposed method had better performance than supervised disambiguation methods even though our method is based on unsupervised disambiguation (using a knowledge-based approach).

This study aims to solve the context-sensitive spelling error problem for English documents. There are two types of spelling errors in English: non-word spelling errors and context-sensitive spelling errors. Non-word spelling errors are simple to correct because they can only be detected by matching the words in sentences with those in a dictionary; however, context-sensitive spelling errors entail increased difficulty of correction because the relationship between the word to be corrected and the surrounding context must be known. Spelling errors are considered noise in every field that uses text information, and preprocessing via document correction is necessary to minimize this problem. Context-sensitive spelling errors include homophone errors (which arise from the incorrect use of words that sound the same but are spelled differently), typographical errors (caused by striking an incorrect key on a keyboard), grammatical errors (which occur when the user does not know the correct grammatical rules), and cross word boundary errors (which arise from incorrect spacing between words). This study focuses on typographical errors. The context-sensitive spelling error problem is solved using the deep learning method, which is not an existing statistical method. The deep learning language model-based correction approach is divided into four parts, namely, correction based on word embedding information, contextual embedding information, an auto-regressive (AR) language model, and an auto-encoding (AE) language model. In this study, the best correction performance was obtained for the AE language model-based approach, and we verified its performance through a detailed correction test.