Digital Library[ Search Result ]
Solving Factual Inconsistency in Abstractive Summarization using Named Entity Fact Discrimination
Jeongwan Shin, Yunseok Noh, Hyun-Je Song, Seyoung Park
http://doi.org/10.5626/JOK.2022.49.3.231
Factual inconsistency in abstractive summarization is a problem that a generated summary can be factually inconsistent with a source text. Previous studies adopted a span selection that replaced entities in the generated summary with entities in the source text because most inconsistencies are related to incorrect entities. These studies assumed that all entities in the generated summary were inconsistent and tried to replace all entities with other entities. However, this was problematic because some consistent entities could be replaced and masked, so information on consistent entities was lost. This paper proposes a method that sequentially executes a fact discriminator and a fact corrector to solve this problem. The fact discriminator determines the inconsistent entities, and the fact corrector replaces only the inconsistent entities. Since the fact corrector corrects only the inconsistent entities, it utilizes the consistent entities. Experiments show that the proposed method boosts the factual consistency of system-generated summaries and outperforms the baselines in terms of both automatic metrics and human evaluation.
Noise Injection for Natural Language Sentence Generation from Knowledge Base
http://doi.org/10.5626/JOK.2020.47.10.965
Generating a natural language sentence from Knowledge base is an operation of entering a triple in the Knowledge base to generate triple information, which is a natural language sentence containing the relationship between the entities. To solve the task of generating sentences from triples using a deep neural network, learning data consisting of many pairs of triples and natural language sentences are required. However, it is difficult to learn the model because the learning data composed in Korean is not yet released. To solve the deficiency of learning data, this paper proposes an unsupervised learning method that extracts keywords based on Korean Wikipedia sentence data and generates learning data using a noise injection technique. To evaluate the proposed method, we used gold-standard dataset produced by triples and sentence pairs. Consequently, the proposed noise injection method showed superior performances over normal unsupervised learning on various evaluation metrics including automatic and human evaluations.
A Knowledge Graph Embedding-based Ensemble Model for Link Prediction
http://doi.org/10.5626/JOK.2020.47.5.473
Knowledge bases often suffer from their limited applicability due to missing information in their entities and relations. Link prediction has been investigated to complete the missing information and makes a knowledge base more useful. The existing studies on link prediction often rely on knowledge graph embedding and have shown trade-off in their performance. In this paper, we propose an ensemble model for knowledge graph embedding to improve quality of link prediction. The proposed model combines multiple knowledge graph embeddings that have unique characteristics. In this way, the ensemble model is able to consider various aspects of the entries within a knowledge base and reduce the variation of accuracy depending on hyper-parameters. Our experiment shows that the proposed model outperforms other knowledge graph embedding methods by 13.5% on WN18 and FB15K dataset.
Language Style Transfer Based on Surface-Level Neutralization
Wooyong Choi, Yunseok Noh, Seyoung Park
http://doi.org/10.5626/JOK.2020.47.2.162
Two main concerns of language style transfer such as sentiment transfer are neutralization of a stylized sentence and re-stylization of the neutralized sentence with a target style. Generally, neutralization is accomplished by learning a neutralized latent space by adversarial learning. However, this neutralization method suffers from the difficulty of maintaining the original content after style transfer. In this paper, we propose a two-step language style transfer method comprised of a surface-level neutralization that removes style words and a target-style word prediction for the removed words. For this, a self-attentive style classifier and style-specific word predictors are used for the surface neutralization and style word generation, respectively. To evaluate the proposed method, several experiments of language style transfer were conducted with Yelp and Amazon review datasets and Caption dataset. As a result, the proposed method shows superior performance over baseline methods on various evaluation metrics including automatic and human evaluations.
Single Sentence Summarization with an Event Word Attention Mechanism
Ian Jung, Su Jeong Choi, Seyoung Park
http://doi.org/10.5626/JOK.2020.47.2.155
The purpose of summarization is to generate short text that preserves important information in the source sentences. There are two approaches for the summarization task. One is an extractive approach and other is an abstractive approach. The extractive approach is to determine if words in a source sentence are retained or not. The abstractive approach generates the summary of a given source sentence using the neural network such as the sequence-to-sequence model and the pointer-generator. However, these approaches present a problem because such approaches omit important information such as event words. This paper proposes an event word attention mechanism for sentence summarization. Event words serve as the key meaning of a given source sentence, since they express what occurs in the source sentence. The event word attention weights are calculated by event information of each words in the source sentence and then it combines global attention mechanism. For evaluation, we used the English and Korean dataset. Experimental results show that, the model of adopting event attention outperforms the existing models.
News Stream Summarization for an Event based on Timeline
Ian Jung, Su Jeong Choi, Seyoung Park
http://doi.org/10.5626/JOK.2019.46.11.1140
This paper explores the summarization task in news stream, as it is continuously produced and has sequential characteristic. Timeline based summarization is widely adopted in news stream summarization because timeline can represent events sequentially. However, previous work relies on the time of collection of news article, thus they cannot consider for dates other than out of the collected period. In addition, previous work lacked consideration of conciseness, informativeness, and coherence. To address these problems, we propose a news stream summarization model with an expanded timeline. The model takes into consideration the expanded timeline by using time points that are referenced in given news articles and selects sentences that are concise, informative and consistent with neighboring time points. First, we constitute expanded timeline by selecting dates which are from all identified time points in the news articles. Then, we extract sentences as summary with consideration of informativeness based on keyword for each time points, and on coherence between two consecutive time points, and on continuity of named entities except for long sentence in the articles. Experimental results show that the proposed model generated higher quality summaries compared to previous work.
Solving for Redundant Repetition Problem of Generating Summarization using Decoding History
Jaehyun Ryu, Yunseok Noh, Su Jeong Choi, Seyoung Park, Seong-Bae Park
http://doi.org/10.5626/JOK.2019.46.6.535
Neural attentional sequence-to-sequence models have achieved great success in abstractive summarization. However, the model is limited by several challenges including repetitive generation of words, phrase and sentences in the decoding step. Many studies have attempted to address the problem by modifying the model structure. Although the consideration of actual history of word generation is crucial to reduce word repetition, these methods, however, do not consider the decoding history of generated sequence. In this paper, we propose a new loss function, called ‘Repeat Loss’ to avoid repetitions. The Repeat Loss directly prevents the model from repetitive generation of words by giving a loss penalty to the generation probability of words already generated in the decoding history. Since the propose Repeat Loss does not need a special network structure, the loss function is applicable to any existing sequence-to-sequence models. In experiments, we applied the Repeat Loss to a number of sequence-to-sequence model based summarization systems and trained them on both Korean and CNN/Daily Mail summarization datasets. The results demonstrate that the proposed method reduced repetitions and produced high-quality summarization.
Scene Generation from a Sentence by Learning Object Relation
Yongmin Shin, Su Jeong Choi, Seong-Bae Park, Seyoung Park
http://doi.org/10.5626/JOK.2019.46.5.431
In communication between humans and machines, location information is crucial. However, it is sometimes omitted. While humans can infer omitted information, machines cannot. Thus, certain problems can occur when generating scenes from sentences. In order to solve this problem, previous studies have found an explicit relation in the sentence, then inferred an implicit relation by using prior probability. However, such methods are not suitable for Korean, as it has morphologically productivity. In this paper, we suggest a scene-generation method for Korean. Frist, we find an explicit relation by using an RNN-based artificial neural network. Then, to infer implicit information, we use the prior probability of relations. Finally, we prepare a scene tree with the obtained information, then generate a scene using that tree. In order to evaluate the scene generation, we measure the accuracy of the model dealing with the relationship and assign a human score to the generated scene. As a result, the method is proven to be effective with excellent performance and evaluation.
Regularizing Korean Conversational Model by Applying Denoising Mechanism
Tae-Hyeong Kim, Yunseok Noh, Seong-Bae Park, Se-Yeong Park
http://doi.org/10.5626/JOK.2018.45.6.572
A conversation system is a system that responds appropriately to input utterances. Recently, the sequence-to-sequence framework has been widely used as a conversation-learning model. However, the conversation model learned in such a way often generates a safe and dull response that does not provide appropriate information or sophisticated meaning. In addition, this model is also useless for input utterances appearing in various forms, such as with changed ending words or changed word order. To solve these problems, we propose a denoising response generation model applying a denoising mechanism. By injecting noise into original input, the proposed method creates a model that will stochastically experience new input made up of items that were not included in the original data during the training process. This data augmentation effect regularizes the model and allows the realization of a robust model. We evaluate our model using 90k input utterances-responses from Korean conversation pair data. The proposed model achieves better results compared to a baseline model on both ROUGE F1 score and qualitative evaluations by human annotators.
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