Harksoo Kim

Hongjin Kim, Jeonghyun Kang, Harksoo Kim

This study addresses critical gaps in Automated Essay Scoring (AES) systems and Large Language Models (LLMs) with regard to their ability to effectively identify and score harmful essays. Despite advancements in AES technology, current models often overlook ethically and morally problematic elements within essays, erroneously assigning high scores to essays that may propagate harmful opinions. In this study, we introduce the Harmful Essay Detection (HED) benchmark, which includes essays integrating sensitive topics such as racism and gender bias, to test the efficacy of various LLMs in recognizing and scoring harmful content. Our findings reveal that: (1) LLMs require further enhancement to accurately distinguish between harmful and argumentative essays, and (2) both current AES models and LLMs fail to consider the ethical dimensions of content during scoring. The study underscores the need for developing more robust AES systems that are sensitive to the ethical implications of the content they are scoring.

Jeonghyun Kang, Hongjin Kim, Harksoo Kim

In this work, we introduce the Keep Emotional and Essential Memory (KEEM) dataset, a novel generation-based dataset designed to enhance memory updates in long-term conversational systems. Unlike existing approaches that rely on simple accumulation or operation-based methods, which often result in information conflicts and difficulties in accurately tracking a user's current state, KEEM dynamically generates integrative memories. This process not only preserves essential factual information but also incorporates emotional context and causal relationships, enabling a more nuanced understanding of user interactions. By seamlessly updating a system's memory with both emotional and essential data, our approach promotes deeper empathy and enhances the system's ability to respond meaningfully in open-domain conversations.

Damrin Kim, NamHyeok Kim, Chanjun Park et al.

This paper presents a novel approach of leveraging Inter-Annotator Agreement (IAA), traditionally used for assessing labeling consistency, to optimize Data Management Operations (DMOps). We advocate for the use of IAA in predicting the labeling quality of individual annotators, leading to cost and time efficiency in data production. Additionally, our work highlights the potential of IAA in forecasting document difficulty, thereby boosting the data construction process's overall efficiency. This research underscores IAA's broader application potential in data-driven research optimization and holds significant implications for large-scale data projects prioritizing efficiency, cost reduction, and high-quality data.

Geunyeong Jeong, Juoh Sun, Seonghee Lee et al.

Large Language Models store extensive factual knowledge acquired during large-scale pre-training. However, this knowledge is inherently static, reflecting only the state of the world at the time of training. Knowledge editing has emerged as a promising solution for updating outdated or incorrect facts without full retraining. However, most existing locate-and-edit methods primarily focus on token-level likelihood optimization without addressing semantic coherence. Our analysis reveals that such edited knowledge is often encoded as isolated residual streams in the model's latent space, distinct from pre-existing knowledge and bypassing natural reasoning process. To address this, we propose \textsc{Steam}, a semantic-level knowledge editing framework that enhances integration of updated knowledge into the model's knowledge structure. \textsc{Steam} first identifies target representations as semantic anchors for the updated factual association, then guides the internal representation of the edited fact towards these anchors through an alignment loss during optimization. Experimental results demonstrate that \textsc{Steam} improves model's ability to reason with edited knowledge and enhances semantic coherence, underscoring the importance of latent-space alignment for reliable and coherent knowledge editing. The code is available at https://github.com/GY-Jeong/STEAM.

Taewoong Yoon, Geunyeong Jeong, Geon Park et al.

Self-Consistency (SC) is an effective decoding strategy that improves the reasoning performance of Large Language Models (LLMs) by generating multiple chain-of-thought reasoning paths and selecting the final answer via majority voting. However, it suffers from substantial inference costs because it requires a large number of samples. To mitigate this issue, Difficulty-Adaptive Self-Consistency (DSC) was proposed to reduce unnecessary token usage for easy problems by adjusting the number of samples according to problem difficulty. However, DSC requires additional model calls and pre-sampling to estimate difficulty, and this process is repeated when applying to each dataset, leading to significant computational overhead. In this work, we propose Activation-Informed Difficulty-Aware Self-Consistency (ACTSC) to address these limitations. ACTSC leverages internal difficulty signals reflected in the feed-forward network neuron activations to construct a lightweight difficulty estimation probe, without any additional token generation or model calls. The probe dynamically adjusts the number of samples for SC and can be applied to new datasets without requiring pre-sampling for difficulty estimation. To validate its effectiveness, we conduct experiments on five benchmarks. Experimental results show that ACTSC effectively reduces inference costs while maintaining accuracy relative to existing methods.

Kyubeen Han, Junseo Jang, Hongjin Kim et al.

Instruction-tuning enhances the ability of large language models (LLMs) to follow user instructions more accurately, improving usability while reducing harmful outputs. However, this process may increase the model's dependence on user input, potentially leading to the unfiltered acceptance of misinformation and the generation of hallucinations. Existing studies primarily highlight that LLMs are receptive to external information that contradict their parametric knowledge, but little research has been conducted on the direct impact of instruction-tuning on this phenomenon. In our study, we investigate the impact of instruction-tuning on LLM's susceptibility to misinformation. Our analysis reveals that instruction-tuned LLMs are significantly more likely to accept misinformation when it is presented by the user. A comparison with base models shows that instruction-tuning increases reliance on user-provided information, shifting susceptibility from the assistant role to the user role. Furthermore, we explore additional factors influencing misinformation susceptibility, such as the role of the user in prompt structure, misinformation length, and the presence of warnings in the system prompt. Our findings underscore the need for systematic approaches to mitigate unintended consequences of instruction-tuning and enhance the reliability of LLMs in real-world applications.

Seongsik Park, Harksoo Kim

Sentence-level relation extraction mainly aims to classify the relation between two entities in a sentence. The sentence-level relation extraction corpus often contains data that are difficult for the model to infer or noise data. In this paper, we propose a curriculum learning-based relation extraction model that splits data by difficulty and utilizes them for learning. In the experiments with the representative sentence-level relation extraction datasets, TACRED and Re-TACRED, the proposed method obtained an F1-score of 75.0% and 91.4% respectively, which are the state-of-the-art performance.

Shinhyeok Oh, Dongyub Lee, Taesun Whang et al.

Existing works for aspect-based sentiment analysis (ABSA) have adopted a unified approach, which allows the interactive relations among subtasks. However, we observe that these methods tend to predict polarities based on the literal meaning of aspect and opinion terms and mainly consider relations implicitly among subtasks at the word level. In addition, identifying multiple aspect-opinion pairs with their polarities is much more challenging. Therefore, a comprehensive understanding of contextual information w.r.t. the aspect and opinion are further required in ABSA. In this paper, we propose Deep Contextualized Relation-Aware Network (DCRAN), which allows interactive relations among subtasks with deep contextual information based on two modules (i.e., Aspect and Opinion Propagation and Explicit Self-Supervised Strategies). Especially, we design novel self-supervised strategies for ABSA, which have strengths in dealing with multiple aspects. Experimental results show that DCRAN significantly outperforms previous state-of-the-art methods by large margins on three widely used benchmarks.

Dojun Park, Youngjin Jang, Harksoo Kim

Natural Language Generation (NLG) refers to the operation of expressing the calculation results of a system in human language. Since the quality of generated sentences from an NLG model cannot be fully represented using only quantitative evaluation, they are evaluated using qualitative evaluation by humans in which the meaning or grammar of a sentence is scored according to a subjective criterion. Nevertheless, the existing evaluation methods have a problem as a large score deviation occurs depending on the criteria of evaluators. In this paper, we propose Grammar Accuracy Evaluation (GAE) that can provide the specific evaluating criteria. As a result of analyzing the quality of machine translation by BLEU and GAE, it was confirmed that the BLEU score does not represent the absolute performance of machine translation models and GAE compensates for the shortcomings of BLEU with flexible evaluation of alternative synonyms and changes in sentence structure.

Dojun Park, Youngjin Jang, Harksoo Kim

This work was conducted to find out how tokenization methods affect the training results of machine translation models. In this work, alphabet tokenization, morpheme tokenization, and BPE tokenization were applied to Korean as the source language and English as the target language respectively, and the comparison experiment was conducted by repeating 50,000 epochs of each 9 models using the Transformer neural network. As a result of measuring the BLEU scores of the experimental models, the model that applied BPE tokenization to Korean and morpheme tokenization to English recorded 35.73, showing the best performance.

Gihyeon Choi, Shinhyeok Oh, Harksoo Kim

Previous researchers have considered sentiment analysis as a document classification task, in which input documents are classified into predefined sentiment classes. Although there are sentences in a document that support important evidences for sentiment analysis and sentences that do not, they have treated the document as a bag of sentences. In other words, they have not considered the importance of each sentence in the document. To effectively determine polarity of a document, each sentence in the document should be dealt with different degrees of importance. To address this problem, we propose a document-level sentence classification model based on deep neural networks, in which the importance degrees of sentences in documents are automatically determined through gate mechanisms. To verify our new sentiment analysis model, we conducted experiments using the sentiment datasets in the four different domains such as movie reviews, hotel reviews, restaurant reviews, and music reviews. In the experiments, the proposed model outperformed previous state-of-the-art models that do not consider importance differences of sentences in a document. The experimental results show that the importance of sentences should be considered in a document-level sentiment classification task.

Seongsik Park, Harksoo Kim

Relation extraction is a type of information extraction task that recognizes semantic relationships between entities in a sentence. Many previous studies have focused on extracting only one semantic relation between two entities in a single sentence. However, multiple entities in a sentence are associated through various relations. To address this issue, we propose a relation extraction model based on a dual pointer network with a multi-head attention mechanism. The proposed model finds n-to-1 subject-object relations using a forward object decoder. Then, it finds 1-to-n subject-object relations using a backward subject decoder. Our experiments confirmed that the proposed model outperformed previous models, with an F1-score of 80.8% for the ACE-2005 corpus and an F1-score of 78.3% for the NYT corpus.

Cheoneum Park, Juae Kim, Hyeon-gu Lee et al.

This paper describes our system, Joint Encoders for Stable Suggestion Inference (JESSI), for the SemEval 2019 Task 9: Suggestion Mining from Online Reviews and Forums. JESSI is a combination of two sentence encoders: (a) one using multiple pre-trained word embeddings learned from log-bilinear regression (GloVe) and translation (CoVe) models, and (b) one on top of word encodings from a pre-trained deep bidirectional transformer (BERT). We include a domain adversarial training module when training for out-of-domain samples. Our experiments show that while BERT performs exceptionally well for in-domain samples, several runs of the model show that it is unstable for out-of-domain samples. The problem is mitigated tremendously by (1) combining BERT with a non-BERT encoder, and (2) using an RNN-based classifier on top of BERT. Our final models obtained second place with 77.78\% F-Score on Subtask A (i.e. in-domain) and achieved an F-Score of 79.59\% on Subtask B (i.e. out-of-domain), even without using any additional external data.