Scaling neural machine translation to 200 languages.

The development of neural techniques has opened up new avenues for research in machine translation. Today, neural machine translation (NMT) systems can leverage highly multilingual capacities and even perform zero-shot translation, delivering promising results in terms of language coverage and quality. However, scaling quality NMT requires large volumes of parallel bilingual data, which are not equally available for the 7,000+ languages in the world1. Focusing on improving the translation qualities of a relatively small group of high-resource languages comes at the expense of directing research attention to low-resource languages, exacerbating digital inequities in the long run. To break this pattern, here we introduce No Language Left Behind-a single massively multilingual model that leverages transfer learning across languages. We developed a conditional computational model based on the Sparsely Gated Mixture of Experts architecture2-7, which we trained on data obtained with new mining techniques tailored for low-resource languages. Furthermore, we devised multiple architectural and training improvements to counteract overfitting while training on thousands of tasks. We evaluated the performance of our model over 40,000 translation directions using tools created specifically for this purpose-an automatic benchmark (FLORES-200), a human evaluation metric (XSTS) and a toxicity detector that covers every language in our model. Compared with the previous state-of-the-art models, our model achieves an average of 44% improvement in translation quality as measured by BLEU. By demonstrating how to scale NMT to 200 languages and making all contributions in this effort freely available for non-commercial use, our work lays important groundwork for the development of a universal translation system.

Languages and future-oriented economic behavior-Experimental evidence for causal effects.

Studies have shown that the use of languages which grammatically associate the future and the present tends to correlate with more future-oriented behavior. We take an experimental approach to go beyond correlation. We asked bilingual research participants, people fluent in two languages (12 language pairs) which differ in the way they encode time, to make a set of future-oriented economic decisions. We find that participants addressed in a language in which the present and the future are marked more distinctly tended to value future events less than participants addressed in a language in which the present and the future are similarly marked. In an additional experiment, bilingual research participants (seven language pairs) were asked to choose whether they wish to complete a more enjoyable task first or later (delayed gratification). When addressed in a language in which the present and the future are marked more distinctly, participants tended to prefer immediate gratification more than when addressed in a language in which the present and the future are marked less distinctly. We shed light on the mechanism in a within-person experiment in which bilingual research participants (nine language pairs) were asked to spatially mark the distance between the present and the future. When participants were addressed in a language in which the present and the future are marked more distinctly, they tended to express more precise temporal beliefs compared with when addressed in a language in which the present and the future are marked less distinctly.

Everyday language input and production in 1,001 children from six continents.

Language is a universal human ability, acquired readily by young children, who otherwise struggle with many basics of survival. And yet, language ability is variable across individuals. Naturalistic and experimental observations suggest that children's linguistic skills vary with factors like socioeconomic status and children's gender. But which factors really influence children's day-to-day language use? Here, we leverage speech technology in a big-data approach to report on a unique cross-cultural and diverse data set: >2,500 d-long, child-centered audio-recordings of 1,001 2- to 48-mo-olds from 12 countries spanning six continents across urban, farmer-forager, and subsistence-farming contexts. As expected, age and language-relevant clinical risks and diagnoses predicted how much speech (and speech-like vocalization) children produced. Critically, so too did adult talk in children's environments: Children who heard more talk from adults produced more speech. In contrast to previous conclusions based on more limited sampling methods and a different set of language proxies, socioeconomic status (operationalized as maternal education) was not significantly associated with children's productions over the first 4 y of life, and neither were gender or multilingualism. These findings from large-scale naturalistic data advance our understanding of which factors are robust predictors of variability in the speech behaviors of young learners in a wide range of everyday contexts.

Cognitive and neural mechanisms of voluntary versus forced language switching in Chinese-English bilinguals: an fMRI study.

The ecological validity of bilingual code-switching has garnered increasing attention in recent years. Contrary to traditional studies that have focused on forced language switching, emerging theories posit that voluntary switching may not incur such a cost. To test these claims and understand differences between forced and voluntary switching, the present study conducted a systematic comparison through both behavioral and neural perspectives. Utilizing fMRI alongside picture-naming tasks, our findings diverge from prior work. Voluntary language switching not only demonstrated switching costs at the behavioral level but also significantly activated brain regions associated with inhibitory control. Direct comparisons of voluntary and forced language switching revealed no significant behavioral differences in switching costs, and both shared several common brain regions that were activated. On the other hand, a nuanced difference between the two types of language switching was revealed by whole-brain analysis: voluntary switching engaged fewer language control regions than forced switching. These findings offer a comprehensive view of the neural and behavioral dynamics involved in bilingual language switching, challenging prior claims that voluntary switching imposes no behavioral or neural costs, and thus providing behavioral and neuroimaging evidence for the involvement of inhibitory control in voluntary language switching.

Functional characterization of the language network of polyglots and hyperpolyglots with precision fMRI.

How do polyglots-individuals who speak five or more languages-process their languages, and what can this population tell us about the language system? Using fMRI, we identified the language network in each of 34 polyglots (including 16 hyperpolyglots with knowledge of 10+ languages) and examined its response to the native language, non-native languages of varying proficiency, and unfamiliar languages. All language conditions engaged all areas of the language network relative to a control condition. Languages that participants rated as higher proficiency elicited stronger responses, except for the native language, which elicited a similar or lower response than a non-native language of similar proficiency. Furthermore, unfamiliar languages that were typologically related to the participants' high-to-moderate-proficiency languages elicited a stronger response than unfamiliar unrelated languages. The results suggest that the language network's response magnitude scales with the degree of engagement of linguistic computations (e.g. related to lexical access and syntactic-structure building). We also replicated a prior finding of weaker responses to native language in polyglots than non-polyglot bilinguals. These results contribute to our understanding of how multiple languages coexist within a single brain and provide new evidence that the language network responds more strongly to stimuli that more fully engage linguistic computations.

Mapping individual aspects of bilingual experience to adaptations in brain structure.

Individual differences in using multiple languages are thought to differentially affect brain structure and function. The present study assessed the neuroanatomical predictions of an emerging theory, the Unifying the Bilingual Experience Trajectories framework, which provides the most comprehensive set of predictions of how individual differences in bilingual experiences lead to specific neural and cognitive adaptations. A total of 140 young adults with variable language experiences were scanned using magnetic resonance imaging and completed demographic questionnaires. Brain structure measures implicated in predictions of the Unifying the Bilingual Experience Trajectories model were extracted and regressed against the model's experiential factors. Consistent with the model's predictions, greater intensity and diversity of bilingual language use resulted in changes in gray matter volume in cortical regions involved in executive control (including inferior frontal gyrus, middle temporal gyrus, angular gyrus, and medial frontal gyrus), indicating adaptations toward handling increased executive control demands. Conversely, duration of bilingual engagement resulted in changes within white matter microstructure (bilateral superior longitudinal fasciculus) and increases in subcortical gray matter (left caudate), indicative of adaptations toward increased efficiency of control. Overall, this research enhances our understanding of how bilingual experiences influence brain structure and provides the first direct empirical evidence for the predictions made by the Unifying the Bilingual Experience Trajectories framework.

Early language dissociation in bilingual minds: magnetoencephalography evidence through a machine learning approach.

Does neural activity reveal how balanced bilinguals choose languages? Despite using diverse neuroimaging techniques, prior studies haven't provided a definitive solution to this problem. Nonetheless, studies involving direct brain stimulation in bilinguals have identified distinct brain regions associated with language production in different languages. In this magnetoencephalography study with 45 proficient Spanish-Basque bilinguals, we investigated language selection during covert picture naming and word reading tasks. Participants were prompted to name line drawings or read words if the color of the stimulus changed to green, in 10% of trials. The task was performed either in Spanish or Basque. Despite similar sensor-level evoked activity for both languages in both tasks, decoding analyses revealed language-specific classification ~100 ms post-stimulus onset. During picture naming, right occipital-temporal sensors predominantly contributed to language decoding, while left occipital-temporal sensors were crucial for decoding during word reading. Cross-task decoding analysis unveiled robust generalization effects from picture naming to word reading. Our methodology involved a fine-grained examination of neural responses using magnetoencephalography, offering insights into the dynamics of language processing in bilinguals. This study refines our understanding of the neural underpinnings of language selection and bridges the gap between non-invasive and invasive experimental evidence in bilingual language production.

Co-learning companionship benefits word learning in a new language: Evidence from a dual-brain EEG examination.

Companionship refers to one's being in the presence of another individual. For adults, acquiring a new language is a highly social activity that often involves learning in the context of companionship. However, the effects of companionship on new language learning have gone relatively underexplored, particularly with respect to word learning. Using a within-subject design, the current study employs electroencephalography to examine how two types of companionship (monitored and co-learning) affect word learning (semantic and lexical) in a new language. Dyads of Chinese speakers of English as a second language participated in a pseudo-word-learning task during which they were placed in monitored and co-learning companionship contexts. The results showed that exposure to co-learning companionship affected the early attention stage of word learning. Moreover, in this early stage, evidence of a higher representation similarity between co-learners showed additional support that co-learning companionship influenced attention. Observed increases in delta and theta interbrain synchronization further revealed that co-learning companionship facilitated semantic access. In all, the similar neural representations and interbrain synchronization between co-learners suggest that co-learning companionship offers important benefits for learning words in a new language.

Phonological properties of logographic words modulate brain activation in bilinguals: a comparative study of Chinese characters and Japanese Kanji.

The brain networks for the first (L1) and second (L2) languages are dynamically formed in the bilingual brain. This study delves into the neural mechanisms associated with logographic-logographic bilingualism, where both languages employ visually complex and conceptually rich logographic scripts. Using functional Magnetic Resonance Imaging, we examined the brain activity of Chinese-Japanese bilinguals and Japanese-Chinese bilinguals as they engaged in rhyming tasks with Chinese characters and Japanese Kanji. Results showed that Japanese-Chinese bilinguals processed both languages using common brain areas, demonstrating an assimilation pattern, whereas Chinese-Japanese bilinguals recruited additional neural regions in the left lateral prefrontal cortex for processing Japanese Kanji, reflecting their accommodation to the higher phonological complexity of L2. In addition, Japanese speakers relied more on the phonological processing route, while Chinese speakers favored visual form analysis for both languages, indicating differing neural strategy preferences between the 2 bilingual groups. Moreover, multivariate pattern analysis demonstrated that, despite the considerable neural overlap, each bilingual group formed distinguishable neural representations for each language. These findings highlight the brain's capacity for neural adaptability and specificity when processing complex logographic languages, enriching our understanding of the neural underpinnings supporting bilingual language processing.

Switching off: disruptive TMS reveals distinct contributions of the posterior middle temporal gyrus and angular gyrus to bilingual speech production.

The role of the left temporoparietal cortex in speech production has been extensively studied during native language processing, proving crucial in controlled lexico-semantic retrieval under varying cognitive demands. Yet, its role in bilinguals, fluent in both native and second languages, remains poorly understood. Here, we employed continuous theta burst stimulation to disrupt neural activity in the left posterior middle-temporal gyrus (pMTG) and angular gyrus (AG) while Italian-Friulian bilinguals performed a cued picture-naming task. The task involved between-language (naming objects in Italian or Friulian) and within-language blocks (naming objects ["knife"] or associated actions ["cut"] in a single language) in which participants could either maintain (non-switch) or change (switch) instructions based on cues. During within-language blocks, cTBS over the pMTG entailed faster naming for high-demanding switch trials, while cTBS to the AG elicited slower latencies in low-demanding non-switch trials. No cTBS effects were observed in the between-language block. Our findings suggest a causal involvement of the left pMTG and AG in lexico-semantic processing across languages, with distinct contributions to controlled vs. "automatic" retrieval, respectively. However, they do not support the existence of shared control mechanisms within and between language(s) production. Altogether, these results inform neurobiological models of semantic control in bilinguals.

Cross-modal and cross-language activation in bilinguals reveals lexical competition even when words or signs are unheard or unseen.

We exploit the phenomenon of cross-modal, cross-language activation to examine the dynamics of language processing. Previous within-language work showed that seeing a sign coactivates phonologically related signs, just as hearing a spoken word coactivates phonologically related words. In this study, we conducted a series of eye-tracking experiments using the visual world paradigm to investigate the time course of cross-language coactivation in hearing bimodal bilinguals (Spanish-Spanish Sign Language) and unimodal bilinguals (Spanish/Basque). The aim was to gauge whether (and how) seeing a sign could coactivate words and, conversely, how hearing a word could coactivate signs and how such cross-language coactivation patterns differ from within-language coactivation. The results revealed cross-language, cross-modal activation in both directions. Furthermore, comparison with previous findings of within-language lexical coactivation for spoken and signed language showed how the impact of temporal structure changes in different modalities. Spoken word activation follows the temporal structure of that word only when the word itself is heard; for signs, the temporal structure of the sign does not govern the time course of lexical access (location coactivation precedes handshape coactivation)-even when the sign is seen. We provide evidence that, instead, this pattern of activation is motivated by how common in the lexicon the sublexical units of the signs are. These results reveal the interaction between the perceptual properties of the explicit signal and structural linguistic properties. Examining languages across modalities illustrates how this interaction impacts language processing.

Event-related Potential Measures of Visual Word Processing in Monolingual and Bilingual Children and Adults: A Focus on Word Frequency Effects.

How does language background influence the neural correlates of visual word recognition in children? To address this question, we used an ERP lexical decision task to examine first-language (L1) and second-language (L2) visual word processing in monolingual and bilingual school-aged children and young adults (n = 123). In particular, we focused on the effects of word frequency (an index of lexical accessibility) on RTs and the N400 ERP component. Behaviorally, we found larger L1 versus L2 word frequency effects among bilingual children, driven by faster and more accurate responses to higher-frequency words (no other language or age group differences were observed). Neurophysiologically, we found larger L1 word frequency effects in bilinguals versus monolinguals (across both age groups), reflected in more negative ERP amplitudes to lower-frequency words. However, the bilingual groups processed L1 and L2 words similarly, despite lower levels of subjective and objective L2 proficiency. Taken together, our findings suggest that divided L1 experience (but not L2 experience) influences the neural correlates of visual word recognition across childhood and adulthood.

The Role of Letter-Speech Sound Integration in Native and Second Language Reading: A Study in Native Japanese Readers Learning English.

The automatic activation of letter-speech sound (L-SS) associations is a vital step in typical reading acquisition. However, the contribution of L-SS integration during nonalphabetic native and alphabetic second language (L2) reading remains unclear. This study explored whether L-SS integration plays a similar role in a nonalphabetic language as in alphabetic languages and its contribution to L2 reading among native Japanese-speaking adults with varying English proficiency. A priming paradigm in Japanese and English was performed by presenting visual letters or symbols, followed by auditory sounds. We compared behavioral and event-related responses elicited by congruent letter-sound pairs, incongruent pairs, and baseline condition (symbol-sound pairs). The behavioral experiment revealed shorter RTs in the congruent condition for Japanese and English tasks, suggesting a facilitation effect of congruency. The ERP experiment results showed an increased early N1 response to Japanese congruent pairs compared to corresponding incongruent stimuli at the left frontotemporal electrodes. Interestingly, advanced English learners exhibited greater activities in bilateral but predominantly right-lateralized frontotemporal regions for the congruent condition within the N1 time window. Moreover, the enhancement of P2 response to congruent pairs was observed in intermediate English learners. These findings indicate that, despite deviations from native language processing, advanced speakers may successfully integrate letters and sounds during English reading, whereas intermediate learners may encounter difficulty in achieving L-SS integration when reading L2. Furthermore, our results suggest that L2 proficiency may affect the level of automaticity in L-SS integration, with the right P2 congruency effect playing a compensatory role for intermediate learners.

Neural substrates of L2-L1 transfer effects on phonological awareness in young Chinese-English bilingual children.

The growing trend of bilingual education between Chinese and English has contributed to a rise in the number of early bilingual children, who were exposed to L2 prior to formal language instruction of L1. The L2-L1 transfer effect in an L1-dominant environment has been well established. However, the threshold of L2 proficiency at which such transfer manifests remains unclear. This study investigated the behavioral and neural processes involved when manipulating phonemes in an auditory phonological task to uncover the transfer effect in young bilingual children. Sixty-two first graders from elementary schools in Taiwan were recruited in this study (29 Chinese monolinguals, 33 Chinese-English bilinguals). The brain activity was measured using fNIRS (functional near-infrared spectroscopy). Bilingual children showed right lateralization to process Chinese and left lateralization to process English, which supports more on the accommodation effect within the framework of the assimilation-accommodation hypothesis. Also, compared to monolinguals, bilingual children showed more bilateral frontal activation in Chinese, potentially reflecting a mixed influence from L2-L1 transfer effects and increased cognitive load of bilingual exposure. These results elucidate the developmental adjustments in the neural substrates associated with early bilingual exposure in phonological processing, offering valuable insights into the bilingual learning process.

Large-scale investigation of white matter structural differences in bilingual and monolingual children: An adolescent brain cognitive development data study.

Emerging research has provided valuable insights into the structural characteristics of the bilingual brain from studies of bilingual adults; however, there is a dearth of evidence examining brain structural alterations in childhood associated with the bilingual experience. This study examined the associations between bilingualism and white matter organization in bilingual children compared to monolingual peers leveraging the large-scale data from the Adolescent Brain Cognitive Development (ABCD) Study. Then, 446 bilingual children (ages 9-10) were identified from the participants in the ABCD data and rigorously matched to a group of 446 monolingual peers. Multiple regression models for selected language and cognitive control white matter pathways were used to compare white matter fractional anisotropy (FA) values between bilinguals and monolinguals, controlling for demographic and environmental factors as covariates in the models. Results revealed significantly lower FA values in bilinguals compared to monolinguals across established dorsal and ventral language network pathways bilaterally (i.e., the superior longitudinal fasciculus and inferior frontal-occipital fasciculus) and right-hemispheric pathways in areas related to cognitive control and short-term memory (i.e., cingulum and parahippocampal cingulum). In contrast to the enhanced FA values observed in adult bilinguals relative to monolinguals, our findings of lower FA in bilingual children relative to monolinguals may suggest a protracted development of white matter pathways associated with language and cognitive control resulting from dual language learning in childhood. Further, these findings underscore the need for large-scale longitudinal investigation of white matter development in bilingual children to understand neuroplasticity associated with the bilingual experience during this period of heightened language learning.

Lost in Translation: Multilingual Analysis of Online Breast Cancer Information.

INTRODUCTION: Patients use the internet to learn more about health conditions. Non-English-speaking patients may face additional challenges. The quality of online breast cancer information, the most common cancer in women, is uncertain. This study aims to examine the quality of online breast cancer information for English and non-English-speaking patients. METHODS: Three search engines were queried using the terms: "how to do a breast examination," "when do I need a mammogram," and "what are the treatment options for breast cancer" in English, Spanish, and Chinese. For each language, 60 unique websites were included and classified by type and information source. Two language-fluent reviewers evaluated website quality using the Journal of American Medical Association benchmark criteria (0-4) and the DISCERN tool (1-5), with higher scores representing higher quality. Scores were averaged for each language. Health On the Net code presence was noted. Inter-rater reliability between reviewers was assessed. RESULTS: English and Spanish websites most commonly originated from US sources (92% and 80%, respectively) compared to Chinese websites (33%, P < 0.001). The most common website type was hospital-affiliated for English (43%) and foundation/advocacy for Spanish and Chinese (43% and 45%, respectively). English websites had the highest and Chinese websites the lowest mean the Journal of American Medical Association (2.2 ± 1.4 versus 1.0 ± 0.8, P = 0.002) and DISCERN scores (3.5 ± 0.9 versus 2.3 ± 0.6, P < 0.001). Health On the Net code was present on 16 (8.9%) websites. Inter-rater reliability ranged from moderate to substantial agreement. CONCLUSIONS: The quality of online information on breast cancer across all three languages is poor. Information quality was poorest for Chinese websites. Improvements to enhance the reliability of breast cancer information across languages are needed.

The role of translation equivalents in bilingual word learning.

Bilingual environments present an important context for word learning. One feature of bilingual environments is the existence of translation equivalents (TEs)-words in different languages that share similar meanings. Documenting TE learning over development may give us insight into the mechanisms underlying word learning in young bilingual children. Prior studies of TE learning have often been confounded by the fact that increases in overall vocabulary size with age lead to greater opportunities for learning TEs. To address this confound, we employed an item-level analysis, which controls for the age trajectory of each item independently. We used Communicative Development Inventory data from four bilingual datasets (two English-Spanish and two English-French; total N = 419) for modeling. Results indicated that knowing a word's TE increased the likelihood of knowing that word for younger children and for TEs that are more similar phonologically. These effects were consistent across datasets, but varied across lexical categories. Thus, TEs may allow bilingual children to bootstrap their early word learning in one language using their knowledge of the other language. RESEARCH HIGHLIGHTS: Bilingual children must learn words that share a common meaning across both languages, that is, translation equivalents, like dog in English and perro in Spanish. Item-level models explored how translation equivalents affect word learning, in addition to child-level (e.g., exposure) and item-level (e.g., phonological similarity) factors. Knowing a word increased the probability of knowing its corresponding translation equivalent, particularly for younger children and for more phonologically-similar translation equivalents. These findings suggest that young bilingual children use their word knowledge in one language to bootstrap their learning of words in the other language.

Development of the Children's English and Spanish Speech Recognition Test: Psychometric Properties, Feasibility, Reliability, and Normative Data.

OBJECTIVES: The Children's English and Spanish Speech Recognition (ChEgSS) test is a computer-based tool for assessing closed-set word recognition in English and in Spanish, with a masker that is either speech-shaped noise or competing speech. The present study was conducted to (1) characterize the psychometric properties of the ChEgSS test, (2) evaluate feasibility and reliability for a large cohort of Spanish/English bilingual children with normal hearing, and (3) establish normative data. DESIGN: Three experiments were conducted to evaluate speech perception in children (4-17 years) and adults (19-40 years) with normal hearing using the ChEgSS test. In Experiment 1, data were collected from Spanish/English bilingual and English monolingual adults at multiple, fixed signal-to-noise ratios. Psychometric functions were fitted to the word-level data to characterize variability across target words in each language and in each masker condition. In Experiment 2, Spanish/English bilingual adults were tested using an adaptive tracking procedure to evaluate the influence of different target-word normalization approaches on the reliability of estimates of masked-speech recognition thresholds corresponding to 70.7% correct word recognition and to determine the optimal number of reversals needed to obtain reliable estimates. In Experiment 3, Spanish/English bilingual and English monolingual children completed speech perception testing using the ChEgSS test to (1) characterize feasibility across age and language group, (2) evaluate test-retest reliability, and (3) establish normative data. RESULTS: Experiments 1 and 2 yielded data that are essential for stimulus normalization, optimizing threshold estimation procedures, and interpreting threshold data across test language and masker type. Findings obtained from Spanish/English bilingual and English monolingual children with normal hearing in Experiment 3 support feasibility and demonstrate reliability for use with children as young as 4 years of age. Equivalent results for testing in English and Spanish were observed for Spanish/English bilingual children, contingent on adequate proficiency in the target language. Regression-based threshold norms were established for Spanish/English bilingual and English monolingual children between 4 and 17 years of age. CONCLUSIONS: The present findings indicate the ChEgSS test is appropriate for testing a wide age range of children with normal hearing in either Spanish, English, or both languages. The ChEgSS test is currently being evaluated in a large cohort of patients with hearing loss at pediatric audiology clinics across the United States. Results will be compared with normative data established in the present study and with established clinical measures used to evaluate English- and Spanish-speaking children. Questionnaire data from parents and clinician feedback will be used to further improve test procedures.