Grammatical error correction (GEC) is commonly referred to as a machine translation task that converts an ungrammatical sentence to a grammatical sentence. This task requires a large amount of parallel data consisting of pairs of ungrammatical and grammatical sentences. However, for the Japanese GEC task, only a limited number of large-scale parallel data are available. Therefore, data augmentation (DA), which generates pseudo-parallel data, is being actively researched. Many previous studies have focused on generating ungrammatical sentences rather than grammatical sentences. To tackle this problem, this study proposes the BERT-DA algorithm, which is a DA algorithm that generates correct sentences using a pre-trained BERT model. In our experiments, we focused on two factors: the source data and the amount of data generated. Considering these elements proved to be more effective for BERT-DA. Based on the evaluation results of multiple domains, the BERT-DA model outperformed the existing system in terms of the Max Match and GLEU+.

Communication through SNS (Social Networking Services) and function of SNS platform can cause two phenomena polarization and echo chambers. In previous study [sasahara 21] about the communication on SNSs regarding a single topic, it has been shown that there is a correlation between these two phenomena and also that SNS functions facilitate the occurrence of these phenomena. In this study, focusing on the fact that real SNS users often deal with multiple topics on a daily basis, we analyzed an agent-based model of SNS communication regarding multiple topics. The analysis revealed that polarization of opinions occurs even when the number of topics increases but that echo chambers are less likely to occur as the number of topics increases. We have also found that there is little correlation between polarization and echo chamber under the condition with more than certain number of topics. Additionally, some functions of SNS such as user recommendations and reposts was found to facilitate echo chambers regardless of the number of topics. This result implies that the findings of previous research are robust even in multiple topics environment. These results suggest that topic diversity and function of SNS platform amendment may be a means of reducing the occurrence of echo chambers. In this model, we cannot take into account the impact of an individual’s interest in a specific topic on such as unfollowing behavior. In future research, we should expand our model to implement these user’s real behavior.

For the current attentive listening agents, the main way to improve the system performance is to design the format of output utterances. However, to achieve better attentive listening performance, it is also important to predict when to start and to end attentive listening. In fact, predicting the timing of attentive listening can be also seen as predicting whether or not the speaker will continue to speak in an occupied multi-unit turn. In this paper, we propose a deep learning model to predict whether or not a speaker will continue speaking in an attentive listening dialog. We focus on the situation in which the respondent continues to respond to a question by the interlocutor or completes the response being produced. Our model has the following three features. First, the input data of our model is designed using domain knowledge about the structure of attentive listening dialog and the characteristics of the vocabulary used therein. Second, we use multimodal information such as text, acoustic features, and characteristic tokens when constructing the model. Third, considering the practicality in actual daily dialogs, we use everyday conversation data collected from a large-scale corpus for constructing the model. The experimental results show that the proposed model achieves the best prediction performance among the models we examine, providing a great potential for prediction of the timing of attentive listening for dialog agents.