Search

Training an automatic speech recognition (ASR) post-processor based on sequence-to-sequence (S2S) requires a parallel pair (e.g., speech recognition result and human post-edited sentence) to construct the dataset, which demands a great amount of human labor. BackTransScription (BTS) proposes a data-building method to mitigate the limitations of the existing S2S based ASR post-processors, which can automatically generate vast amounts of training datasets, reducing time and cost in data construction. Despite the emergence of this novel approach, the BTS-based ASR post-processor still has research challenges and is mostly untested in diverse approaches. In this study, we highlight these challenges through detailed experiments by analyzing the data-centric approach (i.e., controlling the amount of data without model alteration) and the model-centric approach (i.e., model modification). In other words, we attempt to point out problems with the current trend of research pursuing a model-centric approach and alert against ignoring the importance of the data. Our experiment results show that the data-centric approach outperformed the model-centric approach by +11.69, +17.64, and +19.02 in the F1-score, BLEU, and GLEU tests. [ABSTRACT FROM AUTHOR]

The machine translation system aims to translate source language into target language. Recent studies on MT systems mainly focus on neural machine translation. One factor that significantly affects the performance of NMT is the availability of high-quality parallel corpora. However, high-quality parallel corpora concerning Korean are relatively scarce compared to those associated with other high-resource languages, such as German or Italian. To address this problem, AI Hub recently released seven types of parallel corpora for Korean. In this study, we conduct an in-depth verification of the quality of corresponding parallel corpora through Linguistic Inquiry and Word Count (LIWC) and several relevant experiments. LIWC is a word-counting software program that can analyze corpora in multiple ways and extract linguistic features as a dictionary base. To the best of our knowledge, this study is the first to use LIWC to analyze parallel corpora in the field of NMT. Our findings suggest the direction of further research toward obtaining the improved quality parallel corpora through our correlation analysis in LIWC and NMT performance. [ABSTRACT FROM AUTHOR]

With the increasing pervasiveness of smart phones and smart devices, dialogue systems are gaining ever growing attention from both academic and industry. These systems can be broadly classified into two categories, one that is aimed at helping user to gain new knowledge and one that can chat with users without completing any specific tasks. Although dialogue systems are improving substantially, the user experience of such systems are still unsatisfactory as there are no specific rules covering all possible situations of real human–machine dialogue, resulting in breakdowns. There are two technical issues affecting the detection of dialogue breakdown in an open domain conversation: human resources to prepare and annotate a large chunk of conversation data and dialogue histories containing words that don't appear directly in training data. To tackle these issues, we propose a novel encoding method for temporal utterances with memory attention based on end-to-end dialogue breakdown detection. Specifically, long short-term memory (LSTM) is employed to encode each word of all previous user and system utterances. Encoded vectors from LSTM (user and system utterances), along with system and user utterances from sentence embedding, are then stored in memory wherein an attention mechanism is applied to select the most relevant piece of words from system and user utterances for breakdown detection. An evaluation of the proposed approach on a breakdown detection task (DBDC3) showed that the model for single-labeled breakdown detection outperforms other state-of-the-art methods in a classification task. In conclusion, a more effective knowledge gain and management can be achieved by integration of our proposed breakdown detection into dialogue systems. [ABSTRACT FROM AUTHOR]

A mixed-initiative interface conjoins aspects of both adaptable and adaptive interfaces in cases where adaptive customization assistance is added to an adaptable interface, improving the efficacy of customization, efficiency of interactions and user satisfaction. Although many studies showed the efficiency of adaptive customization support, they were either conducted within a laboratory with short-term settings or failed to consider the long-term results of the approach on the elderly. Thus, this study aims to assess the capabilities of adaptive assistance derived from the cognitive and behavioral information of users within an adaptive mixed-initiative UI/UX system (SmartSenior) designed to assist elderly people by improving their familiarity with smart devices. Drawing on cognitive and behavioral data of users, adaptive support was offered by way of customization suggestions that users could accept or disregard at their own discretion. For 10 weeks, 20 senior citizens used SmartSenior, and their actions within the interface were recorded. Half of the test subjects received support and the other half did not. The customization behavior and activity of the two groups were then compared, along with subjective responses concerning the customization support. Results demonstrated that test subjects who were supported made more effective use of SmartSenior's customization features than those who went unsupported. Among the experimental group, subjects accepted most of the customization suggestions provided, and all of them praised the utility of the support and perceived it as beneficial. Moreover, the results show that customization support is more beneficial to users who would never customize of their own volition; such users will be increasingly likely to do so with support. In conclusion, adaptive customization support helps the elderly to more effectively customize their interface, and hence it would helpfully augment the standard adaptable UI/UX. [ABSTRACT FROM AUTHOR]

The adaption of user interface (UI) promises to greatly enhance user experience (UX). This is more evident when we focus on elderly people. However, to date there has yet to be any intelligent, domain independent UI/UX system that caters to such elderly users. In this paper we seek to address this gap, and put forward an intelligent UI/UX system, called SmartSenior, that makes use of semi-supervised learning to execute automatic adaptations that would help elderly users by taking into account both their behavioral data and cognitive responses. SmartSenior initially assesses a user's cognitive capacity and produces a first profile by way of a clustering and classification algorithm. It subsequently produces a personalized UI/UX by altering the profile according to the data on the user's actions. We assess the efficacy of our system by way of an assessment in which elderly users employed the SmartSenior for 8 weeks. This experiment produced results that were, on the whole, satisfactory. [ABSTRACT FROM AUTHOR]

In dialogue systems, understanding user utterances is crucial for providing appropriate responses. Various classification models have been proposed to deal with natural language understanding tasks related to user intention analysis, such as dialogue acts or emotion recognition. However, models that use original lexical features without any modifications encounter the problem of data sparseness, and constructing sufficient training data to overcome this problem is labor-intensive, time-consuming, and expensive. To address this issue, word embedding models that can learn lexical synonyms using vast raw corpora have recently been proposed. However, the analysis of embedding features is not yet sufficient to validate the efficiency of such models. Specifically, using the cosine similarity score as a feature in the embedding space neglects the skewed nature of the word frequency distribution, which can affect the improvement of model performance. This paper describes a novel density-based clustering method that efficiently integrates word embedding vectors into dialogue intention recognition. Experimental results show that our proposed model helps overcome the data sparseness problem seen in previous classification models and can assist in improving the classification performance. [ABSTRACT FROM AUTHOR]

The digitization of the fashion industry diversified consumer segments, and consumers now have broader choices with shorter production cycles; digital technology in the fashion industry is attracting the attention of consumers. Therefore, a system that efficiently supports the searching and recommendation of a product is becoming increasingly important. However, the text-based search method has limitations because of the nature of the fashion industry, in which design is a very important factor. Therefore, we developed an intelligent fashion technique based on deep learning for efficient fashion product searches and recommendations consisting of a Sketch-Product fashion retrieval model and vector-based user preference fashion recommendation model. It was found that the "Precision at 5" of the image-based similar product retrieval model was 0.774 and that of the sketch-based similar product retrieval model was 0.445. The vector-based preference fashion recommendation model also showed positive performance. This system is expected to enhance consumers' satisfaction by supporting users in more effectively searching for fashion products or by recommending fashion products before they begin a search. [ABSTRACT FROM AUTHOR]