The current study investigated the Lombard effect on acoustic features of discourse and prosodic words produced by native Chinese speakers. The results showed that on the discourse level, the overall intensity was significantly increased, vowel duration was significantly lengthened, while the fundamental frequency (F0) was slightly increased for the speech produced in noise conditions compared to that produced in quiet conditions. Further analysis illustrated the Lombard effect on acoustic features of prosodic words. The results showed that the high-pitched point and low-pitched point were significantly increased, the vowel duration was significantly lengthened, while the pitch range was slightly wider for the speech produced in noise conditions compared to that produced in quiet conditions. These findings shed light on the Lombard effect on the acoustic characteristics of prosodic production among native Chinese speakers in noise conditions.
This study presents a quantitative evaluation of translation quality and computational efficiency between semantic and phonetic translation strategies for Chinese scientific terms, using Large Language Models (LLMs). Focusing on two representative domains—the Chinese periodic table and heterocyclic compound nomenclature—we propose a back-translation framework (LLM-BT) specifically designed for evaluating scientific terminology translation, conducting Chinese→ English→ Chinese cycles with Grok, DeepSeek, and GPT-4 models. BLEU scores are used to assess translation accuracy. Results show that semantic translation of element names achieves near-perfect performance (BLEU ≈1.00), leveraging the compositional properties of Chinese characters. In contrast, phonetic translation of heterocyclic compounds yields lower BLEU scores (0.55-0.65), reflecting the inherent challenges of transliteration. Multilingual comparisons using Google Translate (Chinese, Japanese, Korean, English) further validate the superiority of semantic translation strategies (BLEU 0.94-0.99) over phonetic approaches (0.16-0.36). Also, byte-level analysis reveals that Chinese semantic and pictophonetic names are more compact (3 bytes per element) compared to Japanese (13.73 bytes) and Korean (8.97 bytes) transliterations. These findings highlight the dual advantages of semantic translation for both accuracy and computational efficiency. We advocate prioritizing semantic strategies in scientific terminology standardization and cross-linguistic knowledge transfer.
Currently, speech synthesis has found increasingly widespread applications across various fields. Although the quality of synthesized speech continues to improve, challenges remain in emotional speech synthesis, particularly in achieving naturalness and authentic emotional expression. Additionally, traditional emotional speech synthesis models often suffer from low generation efficiency due to their complex processing pipelines. To address these two issues, a LightGrad-based emotional speech synthesis model is proposed in this paper, featuring a constructed feature fusion module, along with a gated style token layer and a prosody predictor incorporated into the model. The gated style token layer helps generate emotional style embedding vectors to capture more subtle and complex emotional details. The prosody predictor provides prosodic information for audio generation to improve the rhythm, pausing patterns, and natural transitions of intonation in speech. In addition, dilated causal convolutions are used to replace depthwise separable convolutions, enhancing the model's ability to model sequential data, while optimal transport conditional flow matching (OT-CFM) is employed to improve the diffusion decoder, effectively increasing the synthesis efficiency of the model. Experimental results demonstrate that the proposed model can generate emotionally expressive speech while maintaining high synthesis quality. Moreover, training on both English and Chinese corpora verifies its cross-lingual capability.
Sarcasm's nature of inverting the implied sentiment makes it difficult to perceive. While sarcasm detection has seen significant development in recent years, most research has focused on the English language, limiting its applicability to sentiment-related tasks in low-resource languages like Filipino. In this work, we address this gap by leveraging synthetic data and large language models (LLMs) to create a textual Filipino sarcasm detection dataset and assess LLM performance on the task. We utilize LLMs due to their capability to generate data quickly with controlled distribution at a much lower cost and the opportunity to test their understanding of the subtle forms of sarcasm, as well as the Filipino language's linguistic and cultural nuances. Our dataset of 1008 synthetic and 100 real-world tweets enabled Gemini 2.0 Flash to achieve 96% accuracy with few-shot prompting on synthetic data. However, performance dropped to 74% on real-world data with zero-shot prompting. This discrepancy highlights the limitation of synthetic data in capturing the authenticity of real-world language and points to the need for fine-tuning LLMs to better adapt to the nuances of Filipino sarcasm detection.
In Indonesia, researchers are expected to conduct studies aligned with national priority research areas defined by strategic policy documents. Researchers submit proposals to obtain government funding, and these propos-als are evaluated based on their relevance to these national priorities. Traditionally, this evaluation involves resource-intensive manual scoring, prompting a need for automation. Recently, large language models (LLMs) have enabled automated scoring for various document-assessment tasks. However, standard fine-tuning-based approaches are impractical because of frequently changing policies, making it challenging to collect sufficient training data. Furthermore, in-context learning approaches typically assign high relevancy scores to nearly all proposals, given their general alignment with national priorities, thereby hindering effective differentiation. A potential solution is pairwise comparison using LLMs, but proposals are typically submitted in PDF format containing extensive text, figures, and tables, often exceeding current LLM context window limits. To overcome these issues, we propose summarizing each proposal into a concise format suitable for relevancy scoring, then employing LLM - based pairwise comparisons combined with the Bradley-Terry model to estimate relevancy scores. Applying this zero-shot method to an Indonesian proposal dataset demonstrates substantial improvements over direct zero-shot scoring and notably surpasses a standard supervised fine-tuning baseline.
Malaysian English, being a low-resource creole language, presents unique challenges for natural language processing tasks such as Named Entity Recognition (NER) and Relational Extraction (RE). In this paper, we propose a methodology utilizing Human-in-the-Loop (HITL) Annotation to address these challenges and enhance the annotation process for NER and RE tasks in Malaysian English. By implementing this methodology, we effectively expanded the MEN Dataset from 6,061 entities to 12,456 entities and from 4,095 relation instances to 7,794 relation instances. This promising outcome serves as an encouragement to expand resources for any low-resource language by implementing the discussed methodology.
Previous research has demonstrated that short-term second language (L2) training can induce perceptual modifications in native language (L1) sound processing. The current study systematically investigated this phenomenon through experiments with native Chinese learners who had received short-term intensive Spanish training. The study comprised two experiments employing pre-test-training-post-test designs with distinct task configurations. Experiment A implemented a sequential protocol consisting of: Chinese consonant identification, Spanish consonant training and Spanish consonant identification tasks. Experiment B featured a modified design with: /p/-/w/ continuum identification and the Spanish training regimen identical to Experiment A. Results revealed that: four days of Spanish training altered L1 consonant perception, with /w/ showing the most notable variability among affected consonants; and the categorical boundary for /w/ perception shifted significantly in terms of formant rise time, while boundary width remained stable. These findings provide robust evidence that brief L2 exposure can induce rapid drift in L1 phonological categories, particularly affecting boundary positions in categorical perception without altering perceptual sharpness.
Dependency parsing has long been a cornerstone of natural language understanding, allowing machines to dissect and learn syntactic structures of sentences. However, traditional methods face certain limitations when applied to spoken language, mainly due to their reliance on a two-step process: first, they transcribe speech into text using Automatic Speech Recognition (ASR), then, they parse the resulting text. This pipeline introduces potential ASR errors, particularly in noisy or resource-scare environments, leading to degraded parsing accuracy and loss of prosodic information, such as intonation and stress. This information is critical to understanding the syntax of spoken language. Our work addresses these challenges by pioneering an end-to-end approach that directly generates dependency trees from the English language, eliminating the need for an intermediate transcription step. Our proposed method aligns speech with syntactic structures, preserving prosodic and acoustic cues, which improves parsing. This multi-modal strategy allows the model to capture the nuances of speech, such as silence and emphasis, which are not present in text-based parsing. Our proposed approaches ensure a seamless fusion of auditory and linguistic information, enabling a direct speech-to-tree (S2T) mapping, which eliminates the error-prone ASR bottleneck. This not only mitigates the impact of transcription inaccuracies, but also leverages the rich contextual information embedded in speech to produce more accurate syntactic representations. Our work paves the way for more natural and effective spoken language understanding, offering a transformative perspective on how machines can interpret embedded human speech structure in a single, unified step, with potential applications in real-time spoken dialogue systems, accessibility tools, and beyond.
Traditional datasets come with limited diversity in terms of language and linguistic features. This makes them language-biased and less generalized towards other languages. In another way, it reduces the effectiveness in a real-world scenarios. On the other side, manually creating the dataset is time-consuming, expensive, and resource-intensive. In order to minimize the limitation, in this paper, by using the large language models, a largescale, code-mixed (English+Hindi) dataset called CoSSHI is created for various tasks, such as deepfake detection, accent identification (60+ hours), and speech synthesis. The proposed dataset is created using the state-of-the-art LLM models (e.g., GPT-4o, Gemini 2.5 Pro, Grok-3, Mistral Large, Llama 4), text-to-speech models (160+ hours, e.g., XTTSv2, XTTSv1.1, Bark, gTTS, IndicTTS, YoursTTS, VITS), and vocoder (320+ hours, e.g., HiFiGAN, BigVGAN). CoSSHI is evaluated on various tasks using different measures, such as precision for deepfake detection, accuracy and precision for accent identification, and subjective and objective measures for speech synthesis. The results indicate that our dataset offers competitive performance compared to existing datasets.
This study investigates the perceptual grounding of chinese color words by examining the alignment between semantic similarity and perceptual color proximity. A total of 120 high-frequency color expressions derived from six basic color terms were analyzed. For each expression, 100 related images were collected and processed to extract 13 perceptual features in the JzAzBz color space, capturing dimensions such as brightness, hue, and color entropy. Semantic similarity was computed using word vectors trained with the FastText model on the CCI 3.0 corpus. Hierarchical clustering revealed coherent groupings within both semantic and perceptual spaces. A significant but moderate positive correlation (Pearson r = 0.16, p <. 001) between the two similarity matrices supports the hypothesis of partial perceptual grounding. Residual analysis further identified aligned and deviated pairs, highlighting the interplay between perceptual features and color words. These findings provide empirical support for embodied theories of meaning and offer a scalable method for multimodal semantic modeling.
Despite significant advancements in emotional speech synthesis technology, existing synthesis systems still face challenges in achieving fine-grained emotional expressiveness while maintaining computational efficiency. To address these limitations, we propose a novel speech synthesis framework based on the Valence-Arousal-Dominance (VAD) three-dimensional emotional model, which integrates continuous emotion feature extraction with multi-feature codebook. Within this framework, we first construct a multi-dimensional codebook encompassing both acoustic and emotional features by extracting VAD characteristics from training set audio samples. An intelligent retrieval strategy is then employed to train a codebook matcher for optimal feature matching. Additionally, the text encoder is enhanced by incorporating adaptive normalization layers, rotary positional encoding, and local attention mechanisms. Experimental results demonstrate that the proposed framework significantly enhances the emotional expressiveness of synthesized speech while maintaining computational efficiency.
This study examines whether Mandarin Lombard speech (speech produced in noise) enhances intelligibility for Pakistani learners. In the experiment, participants heard two kinds of speech where Mandarin sentences spoken in quiet (plain speech) and in noisy conditions (Lombard speech) and then they need to identified target words of the two speeches both in quiet and noisy environments. Results showed benefits of Lombard speech in both quiet and noisy conditions, especially at the end of a sentence. These findings suggest that Lombard speech offers an intelligibility benefit for non-native listeners, which may provide insights for second language learning.
Summarization is an important task in the field of Natural Language Processing (NLP). With the widespread adoption of pre-trained language models, current summarization research has generally entered a stage dominated by abstractive summarization methods, and most of the research focuses on multi-document and cross-lingual summarization. Despite growing interest in both multi-document and cross-lingual summarization, existing research typically treats them as separate challenges, leaving the field of cross-lingual multi-document summarization blank. This gap is particularly significant in real-world applications. Taking news summarization, the main application field of summarization, as an example, international news reports are often not written in the reader's native language, and the same event may be reported in multiple languages. These real-world scenarios emphasize the growing need to develop models capable of cross-lingual summarization, and more broadly, multilingual multi-document summarization. In this work, we propose an end-to-end approach capable of handling both cross-lingual and multilingual multi-document summarization tasks. We validate the effectiveness of our approach through experiments on both self-built datasets and public datasets.
This study investigates the impact of bidialectalism on L2 perception, focusing on the role of activated dialect modes. Chongqing Dialect is a regional dialect spoken in Chongqing, located in southwestern China. One of its prominent features is the lack of distinction between /n/-/l/, whereas Standard Mandarin, China's official language, distinguishes between /n/ and /l/. In this study, twenty Chongqing Dialect-Standard Mandarin speakers completed identification and discrimination tasks for English word-initial consonants /n/-/I/ under both Chongqing Dialect and Standard Mandarin activation modes. Results show that dialect mode activation affects participants' perception of /n/ and /l/. Females outperformed males in perception tasks, and participants were more accurate in distinguishing /n/ and /l/ before the high vowel /I/. This study contributes to a deeper understanding of the perceptual patterns of bidialectal speakers.
This study investigates the reception of the Chinese feminist film Her Story among male users on Hupu, a male-dominated online community, to uncover underlying gender dynamics in contemporary China. Adopting a hybrid quantitative-qualitative approach, the research combines Python-generated high-frequency wordlist analysis with critical discourse analysis (CDA) to scrutinize 443 anonymized comments. The findings reveal two primary discursive purposes: blaming the film for fueling gender antagonism and guarding patriarchy against the rise of female subjectivity in China. The analysis demonstrates how the posters employed nomination, predication, and argumentation strategies to delegitimize the film's feminist theme and present it as a threat to traditional gender hierarchies. The study contributes to understanding the intersection of gender, media reception, and online discourse in China, offering insights into the tensions between emerging feminism and entrenched patriarchal ideologies.
This study investigates how syntactic structures (active vs. passive) and focus location (match vs. mismatch) influence Chinese EFL learners' perception of English pitch accent. The experiments analyzed responses from 21 Chinese learners in tasks requiring identification of accented words and prosodic correctness in dialogues. Results revealed that: (1) Learners exhibited significantly lower accuracy in perceiving pitch accents at sentence-final positions in passive structures, with prolonged reaction times indicating higher cognitive load. (2)When matching pitch accent to focus, learners demonstrated sensitivity to noun positions in both structures, yet passive sentences showed stronger interference. These findings suggest that the syntactic complexity of passive sentences impedes effective integration of pitch accent and focus, likely due to increased working memory demands. The study supports the L2 Intonation Learning Theory and highlights pedagogical implications for intonation training.
This paper takes implicit negative adverbs related to expectational negation in Chinese, English, Korean, and some Chinese minority languages as examples to illustrate the characteristics of the expectation-negating category, and constructs the conceptual space and semantic map of expectation-negation. First, the synchronic semantics of the expectation negating adverbs in different languages are summarized, and their semantic features and conceptual structures are analyzed one by one in combination with real language data. Then, based on semantic associations, a conceptual space is constructed to explain the core characteristics of the expectation negation category, and a semantic map is developed to reveal the commonalities and differences of expectation-negating categories in Chinese, English, Korean, and other languages. The study finds that expectational negation belongs to the category of pragmatic negation. Chinese, English, and Korean share commonalities in the types of semantic nodes, with their synchronic evolution sequences converging. The core of the conceptual space is at the content level, while the periphery is at the speech act level.
This work addresses the critical challenge of transforming static educational content into interactive learning experiences in secondary education, with a specific focus on multilingual environments in Sri Lanka. We propose LK-Active- Learner - an AI-driven platform that leverages knowledge graphs and Retrieval-Augmented Generation (GraphRAG) to create contextually relevant, personalized assessment materials in both Sinhala and English. By combining document AI for text extraction, knowledge graph construction, and question generation techniques, our system aims to enhance student engagement and comprehension while supporting educators through automated examination material development. This paper presents the theoretical framework, system architecture, and evaluation methodology for the proposed platform. This research contributes to educaional technology advancement in low-resource languages and offers a scalable framework for supporting active learning strategies in diverse linguistic contexts.
Globally, with over 7, 000 languages, only about 20 are well-resourced. Low-resource languages face bottlenecks in traditional supervised learning and fine-tuning of pre-trained models due to data scarcity and informal text usage. Research on sentence representation for low-resource languages is crucial for overcoming data scarcity constraints, enhancing semantic modeling capabilities, and promoting the equitable development of natural language processing technology across the global linguistic landscape, as well as advancing various downstream task applications. This paper surveys the evolution of research in lowresource sentence representation, focusing on analyzing the core concepts, effectiveness, and limitations of traditional methods, cross-lingual knowledge transfer, and contrastive representation learning methods. It also discusses prospects for the development of low-resource sentence representation learning.
