Hwiyeol Jo

Kang Min Yoo, Junyeob Kim, Hyuhng Joon Kim et al.

Despite recent explosion of interests in in-context learning, the underlying mechanism and the precise impact of the quality of demonstrations remain elusive. Intuitively, ground-truth labels should have as much impact in in-context learning (ICL) as supervised learning, but recent work reported that the input-label correspondence is significantly less important than previously thought. Intrigued by this counter-intuitive observation, we re-examine the importance of ground-truth labels in in-context learning. With the introduction of two novel metrics, namely Label-Correctness Sensitivity and Ground-truth Label Effect Ratio (GLER), we were able to conduct quantifiable analysis on the impact of ground-truth label demonstrations. Through extensive analyses, we find that the correct input-label mappings can have varying impacts on the downstream in-context learning performances, depending on the experimental configuration. Through additional studies, we identify key components, such as the verbosity of prompt templates and the language model size, as the controlling factor to achieve more noise-resilient ICL.

Seunghee Kim, Bumkyu Park, Kyudan Jung et al.

Most testbeds for omni-modal models assess multimodal understanding via textual outputs, leaving it unclear whether these models can properly speak their answers. To study this, we introduce OmniACBench, a benchmark for evaluating context-grounded acoustic control in omni-modal models. Given a spoken instruction, a text script, and an image, a model must read the script aloud with an appropriate tone and manner. OmniACBench comprises 3,559 verified instances covering six acoustic features: speech rate, phonation, pronunciation, emotion, global accent, and timbre. Extensive experiments on eight models reveal their limitations in the proposed setting, despite their strong performance on prior textual-output evaluations. Our analyses show that the main bottleneck lies not in processing individual modalities, but in integrating multimodal context for faithful speech generation. Moreover, we identify three common failure modes-weak direct control, failed implicit inference, and failed multimodal grounding-providing insights for developing models that can verbalize responses effectively.

Choonghyun Park, Hyuhng Joon Kim, Junyeob Kim et al.

AI Generated Text (AIGT) detectors are developed with texts from humans and LLMs of common tasks. Despite the diversity of plausible prompt choices, these datasets are generally constructed with a limited number of prompts. The lack of prompt variation can introduce prompt-specific shortcut features that exist in data collected with the chosen prompt, but do not generalize to others. In this paper, we analyze the impact of such shortcuts in AIGT detection. We propose Feedback-based Adversarial Instruction List Optimization (FAILOpt), an attack that searches for instructions deceptive to AIGT detectors exploiting prompt-specific shortcuts. FAILOpt effectively drops the detection performance of the target detector, comparable to other attacks based on adversarial in-context examples. We also utilize our method to enhance the robustness of the detector by mitigating the shortcuts. Based on the findings, we further train the classifier with the dataset augmented by FAILOpt prompt. The augmented classifier exhibits improvements across generation models, tasks, and attacks. Our code will be available at https://github.com/zxcvvxcz/FAILOpt.

Joosung Lee, Hwiyeol Jo, Donghyeon Ko et al.

While large language models (LLMs) demonstrate strong capabilities across diverse user queries, they still suffer from hallucinations, often arising from knowledge misalignment between pre-training and fine-tuning. To address this misalignment, we reliably estimate a fine-grained, instance-level knowledge score via multi-sampled inference. Using the knowledge score, we scale the learning signal according to the model's existing knowledge, while encouraging explicit "I don't know" responses for out-of-scope queries. Experimental results show that this approach allows the model to explicitly express uncertainty when it lacks knowledge, while maintaining accuracy on questions it can answer. Furthermore, we propose evaluation metrics for uncertainty, showing that accurate discrimination between known and unknown instances consistently improves performance.

Hwiyeol Jo, Joosung Lee, Jaehone Lee et al.

Evaluating generative models, such as large language models (LLMs), commonly involves question-answering tasks where the final answer is selected based on probability of answer choices. On the other hand, for models requiring reasoning, the method of answer extraction plays a critical role. Our research reveals that the performance of reasoning models and their final answer distributions are highly sensitive to the answer extraction algorithm employed. In order to mitigate this, we propose a basic framework: Answer Regeneration. The method uses an additional model inference, providing the prior input and output prefaced by the prompt "Answer:". The final answer is then selected or extracted from the regenerated output. We show that this extraction-rule-agnostic approach exhibits improved performance and enhanced robustness. Furthermore, we have applied this framework to general math problems and open-ended question answering tasks. Our analysis and this framework could offer a more reliable results for model evaluation.

Joosung Lee, Cheonbok Park, Hwiyeol Jo et al.

Large Language Models (LLMs) are widely used to generate plausible text on online platforms, without revealing the generation process. As users increasingly encounter such black-box outputs, detecting hallucinations has become a critical challenge. To address this challenge, we focus on developing a hallucination detection framework for black-box generators. Motivated by the observation that hallucinations, once introduced, tend to persist, we sample future contexts. The sampled future contexts provide valuable clues for hallucination detection and can be effectively integrated with various sampling-based methods. We extensively demonstrate performance improvements across multiple methods using our proposed sampling approach.

Hwiyeol Jo, Hyunwoo Lee, Kang Min Yoo et al.

The advancements in large language models (LLMs) have brought significant progress in NLP tasks. However, if a task cannot be fully described in prompts, the models could fail to carry out the task. In this paper, we propose a simple yet effective method to contextualize a task toward a LLM. The method utilizes (1) open-ended zero-shot inference from the entire dataset, (2) aggregate the inference results, and (3) finally incorporate the aggregated meta-information for the actual task. We show the effectiveness in text clustering tasks, empowering LLMs to perform text-to-text-based clustering and leading to improvements on several datasets. Furthermore, we explore the generated class labels for clustering, showing how the LLM understands the task through data.

Hwiyeol Jo, Taiwoo Park, Hyunwoo Lee et al.

Although there has been a growing interest among industries in integrating generative LLMs into their services, limited experience and scarcity of resources act as a barrier in launching and servicing large-scale LLM-based services. In this paper, we share our experiences in developing and operating generative AI models within a national-scale search engine, with a specific focus on the sensitiveness of user queries. We propose a taxonomy for sensitive search queries, outline our approaches, and present a comprehensive analysis report on sensitive queries from actual users. We believe that our experiences in launching generative AI search systems can contribute to reducing the barrier in building generative LLM-based services.

Jaeseo Lim, Hwiyeol Jo, Byoung-Tak Zhang et al.

Although the use of active learning to increase learners' engagement has recently been introduced in a variety of methods, empirical experiments are lacking. In this study, we attempted to align two experiments in order to (1) make a hypothesis for machine and (2) empirically confirm the effect of active learning on learning. In Experiment 1, we compared the effect of a passive form of learning to active form of learning. The results showed that active learning had a greater learning outcomes than passive learning. In the machine experiment based on the human result, we imitated the human active learning as a form of knowledge distillation. The active learning framework performed better than the passive learning framework. In the end, we showed not only that we can make build better machine training framework through the human experiment result, but also empirically confirm the result of human experiment through imitated machine experiments; human-like active learning have crucial effect on learning performance.

Hwiyeol Jo, Byoung-Tak Zhang

We introduce a training method for both better word representation and performance, which we call GROVER (Gradual Rumination On the Vector with maskERs). The method is to gradually and iteratively add random noises to word embeddings while training a model. GROVER first starts from conventional training process, and then extracts the fine-tuned representations. Next, we gradually add random noises to the word representations and repeat the training process from scratch, but initialize with the noised word representations. Through the re-training process, we can mitigate some noises to be compensated and utilize other noises to learn better representations. As a result, we can get word representations further fine-tuned and specialized on the task. When we experiment with our method on 5 text classification datasets, our method improves model performances on most of the datasets. Moreover, we show that our method can be combined with other regularization techniques, further improving the model performance.

Hwiyeol Jo, Ceyda Cinarel

We propose a novel and simple method for semi-supervised text classification. The method stems from the hypothesis that a classifier with pretrained word embeddings always outperforms the same classifier with randomly initialized word embeddings, as empirically observed in NLP tasks. Our method first builds two sets of classifiers as a form of model ensemble, and then initializes their word embeddings differently: one using random, the other using pretrained word embeddings. We focus on different predictions between the two classifiers on unlabeled data while following the self-training framework. We also use early-stopping in meta-epoch to improve the performance of our method. Our method, Delta-training, outperforms the self-training and the co-training framework in 4 different text classification datasets, showing robustness against error accumulation.

Hwiyeol Jo

Retrofitting techniques, which inject external resources into word representations, have compensated the weakness of distributed representations in semantic and relational knowledge between words. Implicitly retrofitting word vectors by expansional technique outperforms retrofitting in word similarity tasks with word vector generalization. In this paper, we propose unsupervised extrofitting: expansional retrofitting (extrofitting) without external semantic lexicons. We also propose deep extrofitting: in-depth stacking of extrofitting and further combinations of extrofitting with retrofitting. When experimenting with GloVe, we show that our methods outperform the previous methods on most of word similarity tasks while requiring only synonyms as an external resource. Lastly, we show the effect of word vector enrichment on text classification task, as a downstream task.

Hwiyeol Jo, Jeong Ryu

Starting with the idea that sentiment analysis models should be able to predict not only positive or negative but also other psychological states of a person, we implement a sentiment analysis model to investigate the relationship between the model and emotional state. We first examine psychological measurements of 64 participants and ask them to write a book report about a story. After that, we train our sentiment analysis model using crawled movie review data. We finally evaluate participants' writings, using the pretrained model as a concept of transfer learning. The result shows that sentiment analysis model performs good at predicting a score, but the score does not have any correlation with human's self-checked sentiment.

Hwiyeol Jo, Stanley Jungkyu Choi

We propose post-processing method for enriching not only word representation but also its vector space using semantic lexicons, which we call extrofitting. The method consists of 3 steps as follows: (i) Expanding 1 or more dimension(s) on all the word vectors, filling with their representative value. (ii) Transferring semantic knowledge by averaging each representative values of synonyms and filling them in the expanded dimension(s). These two steps make representations of the synonyms close together. (iii) Projecting the vector space using Linear Discriminant Analysis, which eliminates the expanded dimension(s) with semantic knowledge. When experimenting with GloVe, we find that our method outperforms Faruqui's retrofitting on some of word similarity task. We also report further analysis on our method in respect to word vector dimensions, vocabulary size as well as other well-known pretrained word vectors (e.g., Word2Vec, Fasttext).

Hwiyeol Jo, Soo-Min Kim, Jeong Ryu

As the first step to model emotional state of a person, we build sentiment analysis models with existing deep neural network algorithms and compare the models with psychological measurements to enlighten the relationship. In the experiments, we first examined psychological state of 64 participants and asked them to summarize the story of a book, Chronicle of a Death Foretold (Marquez, 1981). Secondly, we trained models using crawled 365,802 movie review data; then we evaluated participants' summaries using the pretrained model as a concept of transfer learning. With the background that emotion affects on memories, we investigated the relationship between the evaluation score of the summaries from computational models and the examined psychological measurements. The result shows that although CNN performed the best among other deep neural network algorithms (LSTM, GRU), its results are not related to the psychological state. Rather, GRU shows more explainable results depending on the psychological state. The contribution of this paper can be summarized as follows: (1) we enlighten the relationship between computational models and psychological measurements. (2) we suggest this framework as objective methods to evaluate the emotion; the real sentiment analysis of a person.

Hwiyeol Jo, Yohan Moon, Jong In Kim et al.

Remembering an event is affected by personal emotional status. We examined the psychological status and personal factors; depression (Center for Epidemiological Studies - Depression, Radloff, 1977), present affective (Positive Affective and Negative Affective Schedule, Watson et al., 1988), life orient (Life Orient Test, Scheier & Carver, 1985), self-awareness (Core Self Evaluation Scale, Judge et al., 2003), and social factor (Social Support, Sarason et al., 1983) of undergraduate students (N=64) and got summaries of a story, Chronicle of a Death Foretold (Gabriel Garcia Marquez, 1981) from them. We implement a sentimental analysis model based on convolutional neural network (LeCun & Bengio, 1995) to evaluate each summary. From the same vein used for transfer learning (Pan & Yang, 2010), we collected 38,265 movie review data to train the model and then use them to score summaries of each student. The results of CES-D and PANAS show the relationship between emotion and memory retrieval as follows: depressed people have shown a tendency of representing a story more negatively, and they seemed less expressive. People with full of emotion - high in PANAS - have retrieved their memory more expressively than others, using more negative words then others. The contributions of this study can be summarized as follows: First, lightening the relationship between emotion and its effect during times of storing or retrieving a memory. Second, suggesting objective methods to evaluate the intensity of emotion in natural language format, using a sentimental analysis model.