Spoken dialogue systems (SDSs) often encounter significant challenges when interacting with users who exhibit unique conversational behaviors, such as minors, particularly in low-resource environments with limited data availability. To address these challenges, we present a novel data augmentation framework for enhancing SDS performance, particularly when handling such user groups. The framework employs large language models to extract and model speaker styles and leverages pre-trained language models to simulate diverse dialogue act (DA) histories, ultimately creating a rich and personalized set of training data. By focusing on both the unique speaking styles and distinctive dialogue behavior trajectories of users, the framework improves DA prediction accuracy, thereby guiding the SDS to more effectively adapt to low-resource users with specialized conversational characteristics. Extensive experiments conducted in low-resource settings validate the effectiveness of this approach, demonstrating its potential to improve SDS adaptability and foster the development of more inclusive and responsive systems.

Large Language Models (LLMs) possess potentially extensive knowledge; however, because their internal processing operates as a black box, directly editing the knowledge embedded within the LLMs is difficult. To address this issue, a method known as local-modification-based knowledge editing has been developed. This method identifies the “knowledge neurons” that encode the target knowledge and adjusts the parameters associated with these neurons to update the stored information. Knowledge neurons are identified by masking the object (o) from sentences representing relational triplets (s, r, o), with the LLM predicting the masked element, and observing its internal activation patterns during the prediction. When the architecture is decoder-based, the predicted object (o) must be located at the end of the sentence. Previous local-modification-based knowledge-editing methods for decoder-based models have assumed subject-verb-object languages and faced challenges when applied to subject-object-verb languages such as Japanese. In this study, we propose a knowledge-editing method that eliminates the need for word order constraints by converting the input used to identify knowledge neurons into a question, where object (o) is the answer. We conducted validation experiments using a known-facts dataset and confirmed that the proposed method is effective for Japanese language, which is a non- subject-verb-object language.

Automatic extraction of location referring expressions (LREs) can facilitate humanities research by enabling the analysis of large collections of historical texts. In this study, we constructed LRE annotation datasets from early modern and modern travelogues. We then evaluated the performance of Transformer-based contemporary language models in extracting LREs from historical texts by combining these datasets with existing datasets of modern disaster records and contemporary travelogues. Our experiments demonstrated the effectiveness of leveraging contemporary annotated data for LRE extraction from historical texts. However, whereas extraction accuracy on contemporary texts was high (maximum F1 score of 0.890), accuracy on historical texts remained low to moderate (maximum F1 scores of 0.506–0.739), indicating that further model enhancements are needed to better adapt contemporary language models to historical text.

Lexical simplification (LS) is the task of making text easier to understand by replacing complex words with simpler equivalents. LS involves the subtask of lexical complexity prediction (LCP). We present MultiLS-Japanese, the first unified LS and LCP dataset targeting non-native Japanese speakers, and one of the ten language-specific MultiLS datasets. We propose methods for LS and LCP based on large language models (LLMs) that outperform existing LLM-based methods on 7 and 8 of the 10 MultiLS languages, respectively, while using only a fraction of their computational cost. Our methods rely on a single prompt across languages and introduce a novel calibrated token-probability scoring technique, G-Scale, for LCP. Our ablations confirmed the benefits of G-Scale and of concrete wording in the LLM prompt. We made the MultiLS-Japanese dataset available online under a CC-BY-SA license, including detailed metadata.

The meanings and relationships of words shift over time. This phenomenon is referred to as semantic shift. Research focused on understanding how semantic shifts occur over multiple time periods is essential for gaining a detailed understanding of semantic shifts. However, detecting change points only between adjacent time periods is insufficient for analyzing detailed semantic shifts, and using BERT-based methods to examine word sense proportions incurs a high computational cost. To address those issues, we propose a simple yet intuitive framework for how semantic shifts occur over multiple time periods by utilizing similarity matrices based on word embeddings. We calculate diachronic word similarity matrices using fast and lightweight word embeddings across arbitrary time periods, enabling a deeper analysis of continuous semantic shifts. By clustering these similarity matrices, we can further categorize words that exhibit similar patterns of semantic shift in an unsupervised manner. To assess the effectiveness of our framework and explore its limitations, we conducted experiments on both English and Japanese datasets across multiple time intervals.

Personalization has garnered attention in the field of dialogue response generation, with the aim of creating responses tailored to individual user preferences by leveraging background information and dialogue history. Previous studies suggest that, in addition to adapting the content of the response, a system's use of a speaking style similar to that of the user can be a factor in increasing user affinity. For personalization that adapts to subjective preferences, discussions on stylistic similarity should be based on user evaluations. However, numerous evaluations of stylistic similarity rely on objective assessments by third parties who are not participants in the dialogue. The distinction between these and subjective evaluations based on user perception has not been sufficiently examined. In this study, we focused on non-task-oriented dialogue settings and constructed a new dataset in both English and Japanese, annotated with manual evaluations of subjective and objective stylistic similarity, along with user dialogue preferences. Our analysis revealed that stylistic similarity as perceived by the user exhibited a high positive correlation with dialogue preference, whereas no clear correlation was observed with objective stylistic similarity. This study provides empirical evidence for the necessity of distinguishing between evaluation subjects in style assessments for personalization.

The k-nearest-neighbor language model (kNN-LM), one of the retrieval-augmented language models, improves the perplexity for given text by directly accessing a large datastore built from any text data during inference. A widely held hypothesis for the success of kNN-LM is that its explicit memory, i.e., the datastore, enhances predictions for long-tail phenomena. However, prior works have primarily shown its ability to retrieve long-tail contexts, leaving the model’s performance remain underexplored in estimating the probabilities of long-tail target tokens during inference. In this paper, we investigate the behavior of kNN-LM on low-frequency tokens, examining prediction probability, retrieval accuracy, token distribution in the datastore, and approximation error of the product quantization. Our experimental results reveal that kNN-LM does not improve prediction performance for low-frequency tokens.