Digital Library[ Search Result ]
Improving Retrieval Models through Reinforcement Learning with Feedback
Min-Taek Seo, Joon-Ho Lim, Tae-Hyeong Kim, Hwi-Jung Ryu, Du-Seong Chang, Seung-Hoon Na
http://doi.org/10.5626/JOK.2024.51.10.900
Open-domain question answering involves the process of retrieving clues through search to solve problems. In such tasks, it is crucial that the search model provides appropriate clues, as this directly impacts the final performance. Moreover, information retrieval is an important function frequently used in everyday life. This paper recognizes the significance of these challenges and aims to improve performances of search models. Just as the recent trend involves adjusting outputs in decoder models using Reinforcement Learning from Human Feedback (RLHF), this study seeks to enhance search models through the use of reinforcement learning. Specifically, we defined two rewards: the loss of the answer model and the similarity between the retrieved documents and the correct document. Based on these, we applied reinforcement learning to adjust the probability score of the top-ranked document in the search model's document probability distribution. Through this approach, we confirmed the generality of the reinforcement learning method and its potential for further performance improvements.
PrefixLM for Korean Text Summarization
Kun-Hui Lee, Seung-Hoon Na, Joon-Ho Lim, Tae-Hyeong Kim, Du-Seong Chang
http://doi.org/10.5626/JOK.2022.49.6.475
In this paper, we examine the effectiveness of PrefixLM that consists of half of the parameters of the T5"s encoder-decoder architecture for Korean text generation tasks. Different from T5 where input and output sequences are separately provided, the transformer block of PrefixLM takes a single sequence that concatenates both input and output sequences. By designing the attention mask, PrefixLM performs uni- and bi-directional attentions on input and output sequences, respectively, thereby enabling to perform two roles of encoder and decoder with a single transformer block. Experiment results on Korean abstractive document summarization task show that PrefixLM leads to performance increases of 2.17 and 2.78 more than 2 in Rouge-F1 score over BART and T5, respectively, implying that the PrefixLM is promising in Korean text generation tasks.
2-Phase Passage Re-ranking Model based on Neural-Symbolic Ranking Models
Yongjin Bae, Hyun Kim, Joon-Ho Lim, Hyun-ki Kim, Kong Joo Lee
http://doi.org/10.5626/JOK.2021.48.5.501
Previous researches related to the QA system have focused on extracting exact answers for the given questions and passages. However, when expanding the problem from machine reading comprehension to open domain question answering, finding the passage containing the correct answer is as important as machine reading comprehension. DrQA reported that Exact Match@Top1 performance decreased from 69.5 to 27.1 when the QA system had the initial search step. In the present work, we have proposed the 2-phase passage reranking model to improve the performance of the question answering system. The proposed model integrates the results of the symbolic and neural ranking models to re-rank them again. The symbolic ranking model was trained based on the CatBoost algorithm and manual features between the question and passage. The neural model was trained based on the KorBERT model by fine-tuning. The second stage model was trained based on the neural regression model. We maximized the performance by combining ranking models with different characters. Finally, the proposed model showed the performance of 85.8% via MRR and 82.2% via BinaryRecall@Top1 measure while evaluating 1,000 questions. Each performance was improved by 17.3%(MRR) and 22.3%(BR@Top1) compared with the baseline model.
Evaluating of Korean Machine Reading Comprehension Generalization Performance via Cross-, Blind and Open-Domain QA Dataset Assessment
http://doi.org/10.5626/JOK.2021.48.3.275
Machine reading comprehension (MRC) entails identification of the correct answer in a paragraph when a natural language question and paragraph are provided. Recently, fine-tuning based on a pre-trained language model yields the best performance. In this study, we evaluated the ability of machine-reading comprehension method to generalize question and paragraph pairs, rather than similar training sets. Towards this end, the cross-evaluation between datasets and blind evaluation was performed. The results showed a correlation between generalization performance and datasets such as answer length and overlap ratio between question and paragraph. As a result of blind evaluation, the evaluation dataset with the long answer and low lexical overlap between the questions and paragraphs resulted in less than 80% performance. Finally, the generalized performance of the MRC model under the open domain QA environment was evaluated, and the performance of the MRC using the searched paragraph was found to be degraded. According to the MRC task characteristics, the difficulty and differences in generalization performance depend on the relationship between the question and the answer, suggesting the need for analysis of different evaluation sets.
Korean Dependency Parsing using Token-Level Contextual Representation in Pre-trained Language Model
http://doi.org/10.5626/JOK.2021.48.1.27
Dependency parsing is a problem of disambiguating sentence structure by recognizing dependencies and labels between words in sentences. In contrast to previous studies that have applied additional RNNs to the pre-trained language model, this paper proposes a dependency parsing method that uses fine-tuning alone to maximize the self-attention mechanism of the pre-trained language model, and also proposes a technique for using relative distance parameters and SEP tokens. In the results of evaluating the Sejong parsing corpus of TTA standard guidelines, the KorBERT_base model showed 95.73% UAS and 93.39% LAS while the KorBERT_large model showed 96.31% UAS and 94.17% LAS. This represents an improvement of about 3% compared to the results of previous studies that did not use the pre-trained language model. Next, the results of the word-morpheme mixed transformation corpus of the previous study showed that the KorBERT_base model was 94.19% UAS and that the KorBERT_large model was 94.76% UAS.
Korean Dependency Parsing using the Self-Attention Head Recognition Model
http://doi.org/10.5626/JOK.2019.46.1.22
Dependency parsing is the problem solving of structural ambiguities of natural language in sentences. Recently, various deep learning techniques have been applied and shown high performance. In this paper, we analyzed deep learning based dependency parsing problem in three stages. The first stage was a representation step for a word (eojeol) that is a unit of dependency parsing. The second stage was a context reflecting step that reflected the surrounding word information for each word. The last stage was the head word and dependency label recognition step. In this paper, we propose the max-pooling method that is widely used in the CNN model for a word representation. Moreover, we apply the Minimal-RNN Unit that has less computational complexity than the LSTM and GRU for contextual representation. Finally, we propose a Self-Attention Head Recognition Model that includes the relative distance embedding between each word for the head word recognition, and applies multi-task learning to the dependency label recognition simultaneously. For the evaluation, the SEJONG phrase-structure parsing corpus was transformed according to the TTA Standard Dependency Guideline. The proposed model showed the accuracy of parsing for UAS 93.38% and LAS 90.42%.
Korean Semantic Role Labeling Using Semantic Frames and Synonym Clusters
Soojong Lim, Joon-Ho Lim, Chung-Hee Lee, Hyun-Ki Kim
Semantic information and features are very important for Semantic Role Labeling(SRL) though many SRL systems based on machine learning mainly adopt lexical and syntactic features. Previous SRL research based on semantic information is very few because using semantic information is very restricted. We proposed the SRL system which adopts semantic information, such as named entity, word sense disambiguation, filtering adjunct role based on sense, synonym cluster, frame extension based on synonym dictionary and joint rule of syntactic-semantic information, and modified verb-specific numbered roles, etc. According to our experimentations, the proposed present method outperforms those of lexical-syntactic based research works by about 3.77 (Korean Propbank) to 8.05 (Exobrain Corpus) F1-scores.
Syllable-based Korean POS Tagging Based on Combining a Pre-analyzed Dictionary with Machine Learning
Chung-Hee Lee, Joon-Ho Lim, Soojong Lim, Hyun-Ki Kim
This study is directed toward the design of a hybrid algorithm for syllable-based Korean POS tagging. Previous syllable-based works on Korean POS tagging have relied on a sequence labeling method and mostly used only a machine learning method. We present a new algorithm integrating a machine learning method and a pre-analyzed dictionary. We used a Sejong tagged corpus for training and evaluation. While the machine learning engine achieved eojeol precision of 0.964, the proposed hybrid engine achieved eojeol precision of 0.990. In a Quiz domain test, the machine learning engine and the proposed hybrid engine obtained 0.961 and 0.972, respectively. This result indicates our method to be effective for Korean POS tagging.
Search
Journal of KIISE
- ISSN : 2383-630X(Print)
- ISSN : 2383-6296(Electronic)
- KCI Accredited Journal
Editorial Office
- Tel. +82-2-588-9240
- Fax. +82-2-521-1352
- E-mail. chwoo@kiise.or.kr