Search Results (25)

As large language models (LLMs) gain traction among researchers and practitioners, particularly in digital marketing for tasks such as customer feedback analysis and automated communication, concerns remain about the reliability and consistency of their outputs. This study investigates annotation bias in LLMs by comparing human and AI-generated annotation labels across sentiment, topic, and aspect dimensions in hotel booking reviews. Using the HRAST dataset, which includes 23,114 real user-generated review sentences and a synthetically generated corpus of 2000 LLM-authored sentences, we evaluate inter-annotator agreement between a human expert and three LLMs (ChatGPT-3.5, ChatGPT-4, and ChatGPT-4-mini) as a proxy for assessing annotation bias. Our findings show high agreement among LLMs, especially on synthetic data, but only moderate to fair alignment with human annotations, particularly in sentiment and aspect-based sentiment analysis. LLMs display a pronounced neutrality bias, often defaulting to neutral sentiment in ambiguous cases. Moreover, annotation behavior varies notably with task design, as manual, one-to-one prompting produces higher agreement with human labels than automated batch processing. The study identifies three distinct AI biases—repetition bias, behavioral bias, and neutrality bias—that shape annotation outcomes. These findings highlight how dataset complexity and annotation mode influence LLM behavior, offering important theoretical, methodological, and practical implications for AI-assisted annotation and synthetic content generation. Full article

Despite the growing number of bilinguals worldwide, research on how bilingualism influences grammatical development in children with learning disabilities remains limited. This may be due to challenges in assessing language in these children, given the heterogeneity of their disabilities, lack of appropriate tools, and variability in language background and exposure common in bilingual populations. This pilot study investigates grammatical abilities in bilingual versus monolingual children with Down syndrome using the LITMUS Sentence Repetition Task, specifically designed for bilingual populations. Sentence repetition tasks are widely used for assessing grammar in neurotypical children and children with language impairments and are part of many omnibus language assessments. Ten children with Down syndrome aged 5–8 were recruited: five bilingual, speakers of British English and various home languages, and five monolingual, age- and language-matched. Both groups produced a high proportion of ungrammatical repetitions, with more omissions of verbs than nouns, function words than content words, and significant difficulties producing complex structures such as relative clauses, wh-questions, and passives. However, qualitative analyses showed that bilingual children speaking morphologically rich home languages (e.g., Polish, Greek) appeared to have fewer difficulties with some function words (e.g., prepositions) and were able to produce complex structures like passives and wh-questions, unlike their monolingual peers. Although the small sample limits generalisability, two insights emerge: First, sentence repetition may be of limited use in assessing expressive grammar in children with Down syndrome due to frequent ungrammatical responses. Second, while both groups showed similar challenges, bilingualism—especially with richly inflected home languages—may support specific grammatical skills. These findings support existing evidence that bilingualism does not hinder grammatical development in children with Down syndrome and suggest that parents should not avoid dual-language input. Further research is needed to determine whether bilingualism confers specific benefits in grammatical morpheme use and complex syntactic constructions. Full article

The intelligent development of continuous casting quality engineering is an essential step for the efficient production of high-quality billets. However, there are many quality defects that require strong expertise for handling. In order to reduce reliance on expert experience and improve the intelligent management level of billet quality knowledge, we focus on constructing a Domain-Specific Knowledge Graph (DSKG) for the quality engineering of continuous casting. To achieve joint extraction of billet quality defects entity and relation, we propose a Self-Attention Partition and Recombination Model (SAPRM). SAPRM divides domain-specific sentences into three parts: entity-related, relation-related, and shared features, which are specifically for Named Entity Recognition (NER) and Relation Extraction (RE) tasks. Furthermore, for issues of entity ambiguity and repetition in triples, we propose a semi-supervised incremental learning method for knowledge alignment, where we leverage adversarial training to enhance the performance of knowledge alignment. In the experiment, in the knowledge extraction part, the NER and RE precision of our model achieved 86.7% and 79.48%, respectively. RE precision improved by 20.83% compared to the baseline with sequence labeling method. Additionally, in the knowledge alignment part, the precision of our model reached 99.29%, representing a 1.42% improvement over baseline methods. Consequently, the proposed model with the partition mechanism can effectively extract domain knowledge, cand the semi-supervised method can take advantage of unlabeled triples. Our method can adapt the domain features and construct a high-quality knowledge graph for the quality engineering of continuous casting, providing an efficient solution for billet defect issues. Full article

Focus on form (FonF) teaching interventions have been widely employed to help second language learners notice a target grammatical form while their attention is on meaning, i.e., establishing new form–meaning connections. These interventions can be input-based, focusing on the processing of input (i.e., Processing Instruction), or output-based, focusing on production within communicative activities (i.e., Dictogloss). The current pilot study explored whether such teaching interventions would be beneficial for children with DLD. The study employed Processing Instruction and Dictogloss for the teaching of passives in two groups of Greek school-aged children with DLD. The study applied pre-tests and (delayed) post-tests to explore (a) the potential (long-term) effectiveness of the interventions, and (b) potential differences in their effectiveness within this population. The findings indicate that both Processing Instruction and Dictogloss can be promising interventions for Greek children with DLD since they both led to learning gains and retention two weeks after the interventions across all tasks (comprehension, production, and sentence repetition). However, Dictogloss was more effective in production, while Processing Instruction in sentence repetition (when accuracy scores are measured). Full article

Introduction. Primary progressive aphasia (PPA) is a clinical syndrome characterized by a progressive deterioration in language and speech. It is classified into three variants based on symptom patterns: logopenic, semantic, and non-fluent. Due to the lack of fully reliable and valid screening tests for diagnosing PPA and its variants, a Spanish version of the Mini Linguistic State Examination (MLSE) has recently been introduced. Materials and methods. This study aimed to describe the language impairments in a small sample of six patients with PPA and classify them into the three variants using the decision tree and syndrome guide proposed by the MLSE authors. Results. The findings demonstrate the test’s utility in classifying some PPA variants through a qualitative analysis of patient performance and error types. The study revealed a 50% accuracy rate for the decision tree and an 83.33% accuracy rate when using the syndrome guide. Discussion. This discrepancy arises because the decision tree often classified cases as logopenic variant PPA (lvPPA) when working memory was significantly impaired. Specifically, it tended to misclassify patients with semantic, motor, or speech impairments as having lvPPA due to its reliance on the sentence repetition task for assessing working memory. Full article

This study investigates the relationship between vocabulary knowledge and inference-making during real-time reading comprehension in Spanish-speaking adolescents, addressing an important gap in the literature. A large sample of adolescents (n = 265) aged between 11 and 18 were asked to perform a self-paced reading task, which required integrating words across sentence boundaries. The study compared two conditions: repetition and inference. In the repetition condition, a critical word appeared in both context and target sentences, while in the inference condition, the context allowed the inference of the critical word. Vocabulary knowledge was assessed using a standardized receptive vocabulary test. The results showed that adolescents with larger vocabularies exhibited faster reading times across conditions, particularly in the inference condition, where a stronger vocabulary facilitated more efficient word-to-text integration. The interaction between vocabulary and reading condition indicated that a larger vocabulary mitigates the cognitive cost of inference-making, supporting more effective reading comprehension. These findings highlight the role of vocabulary knowledge in reducing the cognitive load associated with inference generation during reading, offering important implications for educational strategies aimed at improving adolescent literacy. Full article

Introduction: The objective of the current study was two-fold. First, it aimed to estimate receptive vocabulary and expressive syntax skills in L2 Italian among early sequential/simultaneous bilingual children of migrant single-mother families with very low socioeconomic status (SES). This objective was achieved by matching the participants’ performance with normative data. Secondly, this study aimed to identify which individual and language exposure factors contributed to learning L2 vocabulary and syntax. Methods: Twenty-four early sequential/simultaneous bilingual children (age range = 5.10–12.4 years) and their mothers were enrolled. Mothers answered questions about linguistic biography and demographic information. Children completed Lexical Comprehension, Sentence Repetition, and Non-Word Repetition tasks from the Language Assessment Battery for 4–12-year-olds to, respectively, assess receptive vocabulary, expressive syntax, and phonological processing. Moreover, non-verbal intellectual functioning was evaluated by the Raven’s Test. Results/Discussion: Compared to normative data, 20 children showed lower receptive vocabulary abilities (<−1.5 SD), 24 lower expressive syntax skills (−2DS), and 7 children lower phonological processing (<−1.5 DS). Moreover, L2 phonological processing and the length of L2 exposure in an educational context positively predicted L2 receptive vocabulary as well as L2 expressive syntax skills. To date, performance in L2 among early sequential/simultaneous bilingual children from migrant households and very low SES remains underexplored. Future efforts need to be directed towards the understanding of factors that impact oral competence in L2, considering that these children will also be exposed to written L2 in the school context. Full article

Medical imaging description and disease diagnosis are vitally important yet time-consuming. Automated diagnosis report generation (DRG) from medical imaging description can reduce clinicians’ workload and improve their routine efficiency. To address this natural language generation task, fine-tuning a pre-trained large language model (LLM) is cost-effective and indispensable, and its success has been witnessed in many downstream applications. However, semantic inconsistency of sentence embeddings has been massively observed from undesirable repetitions or unnaturalness in text generation. To address the underlying issue of anisotropic distribution of token representation, in this study, a contrastive learning penalized cross-entropy (CLpCE) objective function is implemented to enhance the semantic consistency and accuracy of token representation by guiding the fine-tuning procedure towards a specific task. Furthermore, to improve the diversity of token generation in text summarization and to prevent sampling from unreliable tail of token distributions, a diversity contrastive search (DCS) decoding method is designed for restricting the report generation derived from a probable candidate set with maintained semantic coherence. Furthermore, a novel metric named the maximum of token repetition ratio (maxTRR) is proposed to estimate the token diversity and to help determine the candidate output. Based on the LLM of a generative pre-trained Transformer 2 (GPT-2) of Chinese version, the proposed CLpCE with DCS (CLpCEwDCS) decoding framework is validated on 30,000 desensitized text samples from the “Medical Imaging Diagnosis Report Generation” track of 2023 Global Artificial Intelligence Technology Innovation Competition. Using four kinds of metrics evaluated from n-gram word matching, semantic relevance, and content similarity as well as the maxTRR metric extensive experiments reveal that the proposed framework effectively maintains semantic coherence and accuracy (BLEU-1, 0.4937; BLEU-2, 0.4107; BLEU-3, 0.3461; BLEU-4, 0.2933; METEOR, 0.2612; ROUGE, 0.5182; CIDER, 1.4339) and improves text generation diversity and naturalness (maxTRR, 0.12). The phenomenon of dull or repetitive text generation is common when fine-tuning pre-trained LLMs for natural language processing applications. This study might shed some light on relieving this issue by developing comprehensive strategies to enhance semantic coherence, accuracy and diversity of sentence embeddings. Full article

A large body of cross-linguistic research has shown that complex constructions, such as subordinate constructions, are vulnerable in bilingual DLD children, whereas they are robust in bilingual children with typical language development; therefore, they are argued to constitute a potential clinical marker for identifying DLD in bilingual contexts, especially when the majority language is assessed. However, it is not clear whether this also applies to heritage contexts, particularly in contexts in which the heritage language is affected by L2 contact-induced phenomena, as in the case of Heritage Turkish in Germany. In this study, we compare subordination using data obtained from 13 Turkish heritage children with and without DLD (age range 5; 1–11; 6) to 10 late successive (lL2) BiTDs (age range 7; 2–12; 2) and 10 Turkish adult heritage bilinguals (age range 20; 3–25; 10) by analyzing subordinate constructions using both Standard and Heritage Turkish as reference varieties. We further investigate which background factors predict performance in subordinate constructions. Speech samples were elicited using the sentence repetition task (SRT) from the TODİL standardized test battery and the Multilingual Assessment Instrument for Narratives (MAIN). A systematic analysis of a corpus of subordinate clauses constructed with respect to SRT and MAIN narrative production comprehension tasks shows that heritage children with TD and DLD may not be differentiated through these tasks, especially when their utterances are scored using the Standard Turkish variety as a baseline; however, they may be differentiated if the Heritage Turkish is considered as the baseline. The age of onset in the second language (AoO_L2) was the leading performance predictor in subordinate clause production in SRT and in both tasks of MAIN regardless of using Standard Turkish or Heritage Turkish as reference varieties in scoring. Full article

In this study, we explored the linguistic consolidation processes associated with bilingual processing using an experimental paradigm novel in bilingualism research, i.e., sentence repetition. We tested 46 L1-German L2-English bilinguals immersed in the L2 context. Firstly, we compared participants’ sentence repetition accuracy in single-language sentences and in sentences involving code-switches. Secondly, we investigated the processing cost associated with different types of code-switching, i.e., alternation, insertion, and dense code-switching. Finally, we assessed the following potential predictors of repetition accuracy: regular usage of different code-switching types, executive functions (working memory and inhibitory control), as well as relevant bilingualism variables (proficiency, dominance, and immersion). Our first finding was that bilinguals displayed reduced repetition accuracy in sentences involving code-switches compared to single-language sentences, but only when the single-language sentences were in the participants’ L1. This suggests that any processing costs associated with code-switching are modulated by bilinguals’ language background. Moreover, bilinguals’ poor performance in L2 compared to L1 single-language sentences, despite reporting high levels of L2 exposure frequency, highlights the importance of age of acquisition and dominance profiles for language processing. In terms of code-switching, our results revealed that bilinguals’ repetition accuracy differed across different types of code-switching. The processing effort associated with different types of code-switching in the sentence repetition task was primarily driven by the structural depth and the degree of mixing of the involved code-switch, i.e., dense forms of code-switching involving high levels of linguistic co-activation were harder to repeat than alternations involving unintegrated language switching. This effect partially converged with bilinguals’ sociolinguistic practices because bilinguals also reported lower exposure frequency to dense code-switching, but no direct correlations were observed at the level of individual differences. In terms of general cognitive functions, repetition accuracy was modulated by working memory but not by inhibitory control. By investigating this issue, we hope to contribute to our understanding of language processing in the face of cross-linguistic consolidation processes. Full article

As a typical sequence to sequence task, sign language production (SLP) aims to automatically translate spoken language sentences into the corresponding sign language sequences. The existing SLP methods can be classified into two categories: autoregressive and non-autoregressive SLP. The autoregressive methods suffer from high latency and error accumulation caused by the long-term dependence between current output and the previous poses. And non-autoregressive methods suffer from repetition and omission during the parallel decoding process. To remedy these issues in SLP, we propose a novel method named Pyramid Semi-Autoregressive Transformer with Rich Semantics (PSAT-RS) in this paper. In PSAT-RS, we first introduce a pyramid Semi-Autoregressive mechanism with dividing target sequence into groups in a coarse-to-fine manner, which globally keeps the autoregressive property while locally generating target frames. Meanwhile, the relaxed masked attention mechanism is adopted to make the decoder not only capture the pose sequences in the previous groups, but also pay attention to the current group. Finally, considering the importance of spatial-temporal information, we also design a Rich Semantics embedding (RS) module to encode the sequential information both on time dimension and spatial displacement into the same high-dimensional space. This significantly improves the coordination of joints motion, making the generated sign language videos more natural. Results of our experiments conducted on RWTH-PHOENIX-Weather-2014T and CSL datasets show that the proposed PSAT-RS is competitive to the state-of-the-art autoregressive and non-autoregressive SLP models, achieving a better trade-off between speed and accuracy. Full article

Refugee children tend to show low emotional well-being and weak executive functions that may have consequences on language and therefore complicate a potential diagnosis of Developmental Language Disorder (DLD) in this population. We assessed the performance of 140 children living in Switzerland aged 5 to 8 (20 monolinguals, 86 non-refugee bilinguals, 34 refugee bilinguals) on LITMUS language tasks (nonword repetition, sentence repetition, parental questionnaire), standardized language tasks, memory and executive function tasks. Parents also filled in the Child Behavior Checklist providing a measure of their child’s emotional well-being. Results indicate that refugee children are more emotionally vulnerable and show weaker performance in memory and executive functions tasks compared to non-refugee children, in line with the existing literature. Moreover, when compared to non-refugee bilingual children with similar length of exposure to French, refugee children are disadvantaged on all language tasks. Whereas emotional well-being does not predict language performance, memory and executive functions show up as predictors of both LITMUS and standardized language tasks, although in an unsystematic way. It is concluded that refugee children are at risk across the board and that a better understanding of the complex interplay between well-being, executive functions, memory and language is needed in order to build more appropriate diagnostic tools for these children. Full article

The present study investigated bilingual language control among preschool children in a sentence repetition task containing unilingual stimuli and codeswitched stimuli within prepositional phrases (PPs). Cross-language errors, that is, codeswitches that were not part of the stimulus sentences, were taken as evidence of difficulties in language control. Specifically, we investigated cross-language errors as a function of stimulus sentence type (codeswitched or unilingual), CS site within the PP, directionality (English or Hebrew stimulus sentences), and group status (children with typical language development (TLD), and children with Developmental Language Disorder (DLD)). We also examined cross-language errors in terms of word class and locus in the sentence. The participants were 65 English (home language)–Hebrew (societal language) bilinguals with TLD and 13 with DLD, ages 5;5–6;10 (M = 5;11). Stimulus sentences contained five codeswitch conditions within prepositional phrases, for example, a codeswitched preposition (P) or a codeswitched preposition, determiner and noun (P+DET+N), and a ‘no switch’ condition. The stimuli were 36 English and 36 Hebrew sentences (+24 fillers) matched for semantic content and syntax. English sentences contained switches to Hebrew, and Hebrew sentences contained switches to English. The results showed more cross-language errors for codeswitched than unilingual sentence stimuli. The children with TLD showed a directionality effect, producing more cross-language errors in Hebrew sentence stimuli than in English, but the children with DLD did not. The children with DLD had more cross-language errors than their peers with TLD for English stimuli. Most cross-language errors appeared in the sentence-final, adverbial temporal phrase. Findings are discussed in terms of language co-activation and competition in order to account for the difference in performance on unilingual versus codeswitched stimuli and in light of sociopragmatic and psycholinguistic factors to account for the directionality effect among children with TLD and the lack thereof among children with DLD. Full article

A high-performance versatile computer-assisted pronunciation training (CAPT) system that provides the learner immediate feedback as to whether their pronunciation is correct is very helpful in learning correct pronunciation and allows learners to practice this at any time and with unlimited repetitions, without the presence of an instructor. In this paper, we propose deep learning-based techniques to build a high-performance versatile CAPT system for mispronunciation detection and diagnosis (MDD) and articulatory feedback generation for non-native Arabic learners. The proposed system can locate the error in pronunciation, recognize the mispronounced phonemes, and detect the corresponding articulatory features (AFs), not only in words but even in sentences. We formulate the recognition of phonemes and corresponding AFs as a multi-label object recognition problem, where the objects are the phonemes and their AFs in a spectral image. Moreover, we investigate the use of cutting-edge neural text-to-speech (TTS) technology to generate a new corpus of high-quality speech from predefined text that has the most common substitution errors among Arabic learners. The proposed model and its various enhanced versions achieved excellent results. We compared the performance of the different proposed models with the state-of-the-art end-to-end technique of MDD, and our system had a better performance. In addition, we proposed using fusion between the proposed model and the end-to-end model and obtained a better performance. Our best model achieved a 3.83% phoneme error rate (PER) in the phoneme recognition task, a 70.53% F1-score in the MDD task, and a detection error rate (DER) of 2.6% for the AF detection task. Full article

The logopenic variant of Primary Progressive Aphasia (lvPPA), a syndromic disorder centered on language impairment, often presents variable underlying neurodegenerative pathologies such as Alzheimer Disease (AD). Actual language assessment tests and lumbar puncture, focused on AD diagnosis, cannot precisely distinguish the symptoms, or predict their progression at onset time. We analyzed acoustic markers, aiming to discriminate lvPPA and AD as well as the influence of AD biomarkers on acoustic profiles at the beginning of the disease. We recruited people with AD (n = 8) and with lvPPA (n = 8), with cerebrospinal fluid biomarker profiles determined by lumbar puncture. The participants performed a sentence repetition task that allows assessing potential lvPPA phonological loop deficits. We found that temporal and prosodic markers significantly differentiate the lvPPA and AD group at an early stage of the disease. Biomarker and acoustic profile comparisons discriminated the two lvPPA subgroups according to their biomarkers. For lvPPA with AD biomarkers, acoustic profile equivalent to an atypical AD form with a specific alteration of the phonological loop is shown. However, lvPPA without AD biomarkers has an acoustic profile approximating the one for DLFT. Therefore, these results allow us to classify lvPPA differentially from AD based on acoustic markers from a sentence repetition task. Furthermore, our results suggest that acoustic analysis would constitute a clinically efficient alternative to refused lumbar punctures. It offers the possibility to facilitate early, specific, and accessible neurodegenerative diagnosis and may ease early care with speech therapy, preventing the progression of symptoms. Full article