• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.18 no.9
• /
• pp.2464-2482
• /
• 2024

The semantic understanding of numbers requires association with context. However, powerful neural networks overfit spurious correlations between context and numbers in training corpus can lead to the occurrence of contextual bias, which may affect the network's accurate estimation of number magnitude when making inferences in real-world data. To investigate the resilience of current methodologies against contextual bias, we introduce a novel out-of-distribution (OOD) numerical question-answering (QA) dataset that features specific correlations between context and numbers in the training data, which are not present in the OOD test data. We evaluate the robustness of different numerical encoding and decoding methods when confronted with contextual bias on this dataset. Our findings indicate that encoding methods incorporating more detailed digit information exhibit greater resilience against contextual bias. Inspired by this finding, we propose a digit-aware position embedding strategy, and the experimental results demonstrate that this strategy is highly effective in improving the robustness of neural networks against contextual bias.

• Journal of Intelligence and Information Systems
• /
• v.28 no.2
• /
• pp.127-146
• /
• 2022

Recently, as word embedding has shown excellent performance in various tasks of deep learning-based natural language processing, researches on the advancement and application of word, sentence, and document embedding are being actively conducted. Among them, cross-language transfer, which enables semantic exchange between different languages, is growing simultaneously with the development of embedding models. Academia's interests in vector alignment are growing with the expectation that it can be applied to various embedding-based analysis. In particular, vector alignment is expected to be applied to mapping between specialized domains and generalized domains. In other words, it is expected that it will be possible to map the vocabulary of specialized fields such as R&D, medicine, and law into the space of the pre-trained language model learned with huge volume of general-purpose documents, or provide a clue for mapping vocabulary between mutually different specialized fields. However, since linear-based vector alignment which has been mainly studied in academia basically assumes statistical linearity, it tends to simplify the vector space. This essentially assumes that different types of vector spaces are geometrically similar, which yields a limitation that it causes inevitable distortion in the alignment process. To overcome this limitation, we propose a deep learning-based vector alignment methodology that effectively learns the nonlinearity of data. The proposed methodology consists of sequential learning of a skip-connected autoencoder and a regression model to align the specialized word embedding expressed in each space to the general embedding space. Finally, through the inference of the two trained models, the specialized vocabulary can be aligned in the general space. To verify the performance of the proposed methodology, an experiment was performed on a total of 77,578 documents in the field of 'health care' among national R&D tasks performed from 2011 to 2020. As a result, it was confirmed that the proposed methodology showed superior performance in terms of cosine similarity compared to the existing linear vector alignment.

• Journal of Broadcast Engineering
• /
• v.24 no.1
• /
• pp.48-57
• /
• 2019

The purpose of this study is to understand the intention of the inquirer from the single text type question in Goal-oriented dialogue. Goal-Oriented Dialogue system means a dialogue system that satisfies the user's specific needs via text or voice. The intention analysis process is a step of analysing the user's intention of inquiry prior to the answer generation, and has a great influence on the performance of the entire Goal-Oriented Dialogue system. The proposed model was used for a daily chemical products domain and Korean text data related to the domain was used. The analysis is divided into a speech-act which means independent on a specific field concept-sequence and which means depend on a specific field. We propose a classification method using the word embedding model and the CNN as a method for analyzing speech-act and concept-sequence. The semantic information of the word is abstracted through the word embedding model, and concept-sequence and speech-act classification are performed through the CNN based on the semantic information of the abstract word.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.26 no.1
• /
• pp.7-14
• /
• 2022

Sensor Registry System (SRS) has been devised for maintaining semantic interoperability of data on heterogeneous sensor networks. SRS measures the connectability of the mobile device to ambient sensors based on positions and only provides metadata of sensors that may be successfully connected. The step of identifying the ambient sensors which can be successfully connected is called sensor filtering. Improving the performance of sensor filtering is one of the core issues of SRS research. In reality, GPS sometimes shows the wrong position and thus leads to failed sensor filtering. Therefore, this paper proposes a new sensor filtering strategy using geographical embedding and neural network-based path prediction. This paper also evaluates the service provision rate with the Monte Carlo approach. The empirical study shows that the proposed method can compensate for position abnormalities and is an effective model for sensor filtering in SRS.

• The Korean Journal of Applied Statistics
• /
• v.34 no.3
• /
• pp.295-308
• /
• 2021

In recent years, as many people post their interests on social media or store documents in digital form due to the development of the internet and computer technologies, the amount of text data generated has exploded. Accordingly, the demand for technology to create valuable information from numerous document data is also increasing. In this study, through statistical techniques, we investigate how the meanings of Korean words change over time by using the presidential speech records and newspaper articles public data. Using this, we present a strategy that can be utilized in the study of the synchronic change of Hangeul. The purpose of this study is to deviate from the study of the theoretical language phenomenon of Hangeul, which was studied by the intuition of existing linguists or native speakers, to derive numerical values through public documents that can be used by anyone, and to explain the phenomenon of changes in the meaning of words.

• Smart Media Journal
• /
• v.11 no.3
• /
• pp.31-38
• /
• 2022

In the document extraction summary study, various methods for selecting important sentences based on the relationship between sentences were proposed. In the Korean document summary using the summation similarity of sentences, the summation similarity of the sentences was regarded as the amount of sentence information, and the summary sentences were extracted by selecting important sentences based on this. However, the problem is that it does not take into account the various importance that each sentence contributes to the entire document. Therefore, in this study, we propose a document extraction summary method that provides a summary by selecting important sentences based on the amount of quantitative and semantic information in the sentence. As a result, the extracted sentence agreement was 58.56% and the ROUGE-L score was 34, which was superior to the method using only the combined similarity. Compared to the deep learning-based method, the extraction method is lighter, but the performance is similar. Through this, it was confirmed that the method of compressing information based on semantic similarity between sentences is an important approach in document extraction summary. In addition, based on the quickly extracted summary, the document generation summary step can be effectively performed.

• The Journal of the Korea Contents Association
• /
• v.22 no.8
• /
• pp.33-43
• /
• 2022

Labeled graphs are used to represent entities, their relationships, and their structures in real data such as knowledge graphs and protein interactions. With the rapid development of IT and the explosive increase in data, there has been a need for a subgraph matching technology to provide information that the user is interested in. In this paper, we propose an approximate Top-k labeled subgraph matching scheme that considers the semantic similarity of labels and the difference in graph structure. The proposed scheme utilizes a learning model using FastText in order to consider the semantic similarity of a label. In addition, the label similarity graph(LSG) is used for approximate subgraph matching by calculating similarity values between labels in advance. Through the LSG, we can resolve the limitations of the existing schemes that subgraph expansion is possible only if the labels match exactly. It supports structural similarity for a query graph by performing searches up to 2-hop. Based on the similarity value, we provide k subgraph matching results. We conduct various performance evaluations in order to show the superiority of the proposed scheme.

• Annual Conference on Human and Language Technology
• /
• 2017.10a
• /
• pp.116-121
• /
• 2017

cQA(Community-based Question Answering) 시스템은 온라인 커뮤니티를 통해 사용자들이 질문을 남기고 답변을 작성할 수 있도록 만들어진 시스템이다. 신규 질문이 인입되면, 기존에 축적된 cQA 저장소에서 해당 질문과 가장 유사한 질문을 검색하고, 그 질문에 대한 답변을 신규 질문에 대한 답변으로 대체할 수 있다. 하지만, 키워드 매칭을 사용하는 전통적인 검색 방식으로는 문장에 내재된 의미들을 이용할 수 없다는 한계가 있다. 이를 극복하기 위해서는 의미적으로 동일한 문장들로 학습이 되어야 하지만, 이러한 데이터를 대량으로 확보하기에는 어려움이 있다. 본 논문에서는 질문이 제목과 내용으로 분리되어 있는 대량의 cQA 셋에서, 질문 제목과 내용을 의미 벡터 공간으로 사상하고 두 벡터의 상대적 거리가 가깝게 되도록 학습함으로써 의사(pseudo) 유사 의미의 성질을 내재화 하였다. 또한, 질문 제목과 내용의 의미 벡터 표현(representation)을 위하여, semi-training word embedding과 CNN(Convolutional Neural Network)을 이용한 딥러닝 기법을 제안하였다. 유사 질문 검색 실험 결과, 제안 모델을 이용한 검색이 키워드 매칭 기반 검색보다 좋은 성능을 보였다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.10 no.9
• /
• pp.4568-4584
• /
• 2016

Steganography based on text generation has become a hot research topic in recent years. However, current text-generation methods which generate texts of normal style have either semantic or syntactic flaws. Note that texts of special genre, such as poem, have much simpler language model, less grammar rules, and lower demand for naturalness. Motivated by this observation, in this paper, we propose a text steganography that utilizes Markov chain model to generate Ci-poetry, a classic Chinese poem style. Since all Ci poems have fixed tone patterns, the generation process is to select proper words based on a chosen tone pattern. Markov chain model can obtain a state transfer matrix which simulates the language model of Ci-poetry by learning from a given corpus. To begin with an initial word, we can hide secret message when we use the state transfer matrix to choose a next word, and iterating until the end of the whole Ci poem. Extensive experiments are conducted and both machine and human evaluation results show that our method can generate Ci-poetry with higher naturalness than former researches and achieve competitive embedding rate.

• Annual Conference of KIPS
• /
• 2017.04a
• /
• pp.941-944
• /
• 2017

심층 학습에 기반을 둔 통계적 언어모형에서 가장 중요한 작업은 단어의 분산 표현(Distributed Representation)이다. 단어의 분산 표현은 단어 자체가 가지는 의미를 다차원 공간에서 벡터로 표현하는 것으로서, 워드 임베딩(word embedding)이라고도 한다. 워드 임베딩을 이용한 심층 학습 기반 통계적 언어모형은 전통적인 통계적 언어모형과 비교하여 성능이 우수한 것으로 알려져 있다. 그러나 워드 임베딩 역시 자료 부족분제에서 벗어날 수 없다. 특히 학습데이터에 나타나지 않은 단어(unknown word)를 처리하는 것이 중요하다. 본 논문에서는 고품질 한국어 워드 임베딩을 위하여 단어의 의미적 계층정보를 이용한 워드 임베딩 방법을 제안한다. 기존연구에서 제안한 워드 임베딩 방법을 그대로 활용하되, 학습 단계에서 목적함수가 입력 단어의 하위어, 동의어를 반영하여 계산될 수 있도록 수정함으로써 단어의 의미적 계층청보를 반영할 수 있다. 본 논문에서 제안한 워드 임베딩 방법을 통해 생성된 단어 벡터의 유추검사(analog reasoning) 결과, 기존 방법보다 5%가 증가한 47.90%를 달성할 수 있었다.