Translation Latency Research Articles

Characterizing language production across modalities

ABSTRACT This study investigates factors influencing lexical access in language production across modalities (signed and oral). Data from deaf and hearing signers were reanalyzed (Baus and Costa, 2015, On the temporal dynamics of sign production: An ERP study in Catalan Sign Language (LSC). Brain Research, 1609(1), 40-53. https://doi.org/10.1016/j.brainres.2015.03.013; Gimeno-Martínez and Baus, 2022, Iconicity in sign language production: Task matters. Neuropsychologia, 167, 108166. https://doi.org/10.1016/j.neuropsychologia.2022.108166) testing the influence of psycholinguistic variables and ERP mean amplitudes on signing and naming latencies. Deaf signers' signing latencies were influenced by sign iconicity in the picture signing task, and by spoken psycholinguistic variables in the word-to-sign translation task. Additionally, ERP amplitudes before response influenced signing but not translation latencies. Hearing signers' latencies, both signing and naming, were influenced by sign iconicity and word frequency, with early ERP amplitudes predicting only naming latencies. These findings highlight general and modality-specific determinants of lexical access in language production.

Age- and gender-specific normative data on computerized dynamic posturography in a cohort of Danish adults.

Balance performance can be measured with a computerized dynamic posturography (CDP). Frequently used CDP protocols include the sensory organization test (SOT), motor control test (MCT) and adaptation test (ADT). The primary aim of this study was to produce normative data using the Bertec® Balance Advantage CDP on a cohort of healthy adults aged 20-79years for the SOT, MCT and ADT. The secondary aim was to determine age- and gender-related differences. The study population of 120 participants were divided into age groups of 20-29, 30-39years, etc., with an equal number of men and women in each group. Using the Bertec® CDP, the SOT, MCT and ADT protocols were performed. SOT found that the 70-79 age group and men scored lower equilibrium scores (ES) on SOT4, and SOT6. Women had higher ES on SOT1. For the MCT, backward translation. latencies were higher for the 60s and 70s age groups. The 60s age group had higher forward translation latencies. Women had lower latencies on all MCTs. For the ADT toes up test, the 70-79 age group scored higher sway energy. On the toes down test, higher sway energy was registered for the 50-79 age groups. Women had lower sway energy for both ADT's. We have produced normative data for healthy adults aged 20-79years for the SOT, MCT and ADT protocols using the Bertec® Balance Advantage CDP. Balance performance differences might be due to decline in visual acuity, presbyvestibulopathy and differences in body composition.

Sequence-Level Training for Non-Autoregressive Neural Machine Translation

Abstract In recent years, Neural Machine Translation (NMT) has achieved notable results in various translation tasks. However, the word-by-word generation manner determined by the autoregressive mechanism leads to high translation latency of the NMT and restricts its low-latency applications. Non-Autoregressive Neural Machine Translation (NAT) removes the autoregressive mechanism and achieves significant decoding speedup by generating target words independently and simultaneously. Nevertheless, NAT still takes the word-level cross-entropy loss as the training objective, which is not optimal because the output of NAT cannot be properly evaluated due to the multimodality problem. In this article, we propose using sequence-level training objectives to train NAT models, which evaluate the NAT outputs as a whole and correlates well with the real translation quality. First, we propose training NAT models to optimize sequence-level evaluation metrics (e.g., BLEU) based on several novel reinforcement algorithms customized for NAT, which outperform the conventional method by reducing the variance of gradient estimation. Second, we introduce a novel training objective for NAT models, which aims to minimize the Bag-of-N-grams (BoN) difference between the model output and the reference sentence. The BoN training objective is differentiable and can be calculated efficiently without doing any approximations. Finally, we apply a three-stage training strategy to combine these two methods to train the NAT model. We validate our approach on four translation tasks (WMT14 En↔De, WMT16 En↔Ro), which shows that our approach largely outperforms NAT baselines and achieves remarkable performance on all translation tasks. The source code is available at https://github.com/ictnlp/Seq-NAT.

Non-identical cognates yield facilitation in translation – does the way foreign vocabulary is learned affect its processing?

Abstract Cognates (words sharing form and meaning across languages, e.g. Polish–English FILM–FILM (identical cognates) or TUNNEL–TUNEL (non-identical cognates)) are processed faster than single-language words (cognate facilitation effect), which is modulated by a number of factors. Here, we extended the study reported by Lijewska and Chmiel (2015) and tested the influence of learning experience on non-identical cognate processing in a translation task in 2 experiments with Polish-German–English trilinguals and with German–English bilinguals. The trilinguals learned English and German via Polish, whereas the bilinguals learnt English via German. They translated Polish–English non-identical cognates (MUSZTARDA–SENF–MUSTARD), German–English non-identical cognates (TRAWA–GRAS–GRASS) and controls from English into Polish and into German (trilinguals) or only into German (bilinguals). We found no evidence for the influence of learning experience on cognate processing but the novel finding was a reliable facilitation effect obtained when only non-identical cognates were tested in a translation task with trilinguals (in the accuracy data) as well as with bilinguals (in translation latencies). Additionally, the reported study pointed to the important role of language proficiency in processing.

Latent-Variable Non-Autoregressive Neural Machine Translation with Deterministic Inference Using a Delta Posterior

Although neural machine translation models reached high translation quality, the autoregressive nature makes inference difficult to parallelize and leads to high translation latency. Inspired by recent refinement-based approaches, we propose LaNMT, a latent-variable non-autoregressive model with continuous latent variables and deterministic inference procedure. In contrast to existing approaches, we use a deterministic inference algorithm to find the target sequence that maximizes the lowerbound to the log-probability. During inference, the length of translation automatically adapts itself. Our experiments show that the lowerbound can be greatly increased by running the inference algorithm, resulting in significantly improved translation quality. Our proposed model closes the performance gap between non-autoregressive and autoregressive approaches on ASPEC Ja-En dataset with 8.6x faster decoding. On WMT'14 En-De dataset, our model narrows the gap with autoregressive baseline to 2.0 BLEU points with 12.5x speedup. By decoding multiple initial latent variables in parallel and rescore using a teacher model, the proposed model further brings the gap down to 1.0 BLEU point on WMT'14 En-De task with 6.8x speedup.

DUCATI

Conventional on-chip TLB hierarchies are unable to fully cover the growing application working-set sizes. To make things worse, Last-Level TLB (LLT) misses require multiple accesses to the page table even with the use of page walk caches. Consequently, LLT misses incur long address translation latency and hurt performance. This article proposes two low-overhead hardware mechanisms for reducing the frequency and penalty of on-die LLT misses. The first, Unified CAche and TLB (UCAT) , enables the conventional on-die Last-Level Cache to store cache lines and TLB entries in a single unified structure and increases on-die TLB capacity significantly. The second, DRAM-TLB , memoizes virtual to physical address translations in DRAM and reduces LLT miss penalty when UCAT is unable to fully cover total application working-set. DRAM-TLB serves as the next larger level in the TLB hierarchy that significantly increases TLB coverage relative to on-chip TLBs. The combination of these two mechanisms, DUCATI , is an address translation architecture that improves GPU performance by 81%; (up to 4.5×) while requiring minimal changes to the existing system design. We show that DUCATI is within 20%, 5%, and 2% the performance of a perfect LLT system when using 4KB, 64KB, and 2MB pages, respectively.

Analysis of Memory Access Latency Considering Page Faults and TLB Misses in NVM Storage

As high performance NVM storage emerges, memory system configurations optimized for HDD should be revisited. This paper explores the performance of memory systems that use NVM as a storage device and discusses how such systems can be managed efficiently. Specifically, we analyze the memory access time separately for address translation latency and data access latency as the page size, read-ahead, and storage performances are varied. As the page fault cost becomes small under NVM storage, we observe that the bottleneck of memory systems can be shifted to address translation. We show that determining an appropriate page size can improve the address translation latency without increasing data access latency. We also show that turning off the read-ahead option is helpful in reducing data access latency. We expect that our new architecture with appropriate configurations will be helpful in the design of emerging memory systems.

It Is Not Necessary to Retrieve the Phonological Nodes of Context Objects for Chinese Speakers.

The issue of how activation is transmitted from semantic to phonological level in spoken production remains controversial. Recent evidences from alphabetic languages support a cascaded view. However, given the different architecture of phonological encoding in non-alphabetic languages, it is not clear whether this view applies in Chinese, as a non-alphabetic script. We therefore investigated whether the not-to-be named pictures activate their phonological properties in Chinese speech production. In Experiment 1, participants were presented a target English word and a context picture (semantically related or unrelated, phonologically related or unrelated to target word in Chinese) and were asked to translate the English word into a Chinese word. The translation latencies were faster in the semantically related condition than in the unrelated condition. By contrast, no difference between phonologically related and unrelated was observed. In Experiment 2, in order to promote participants phonological sensitivity in a word-translation task, we increased the proportion of phonologically related trials from 25 to 50%. In Experiment 3, we employed a word association task that was more sensitive to phonological activation of context objects than a word translation task. The phonological activation of context objects were absent again in Experiments 2 and 3. Bayes Factor analysis suggested that the absence of phonological activation of context pictures was reliable. Results consistently revealed that only target lemma could activate the corresponding phonological node to guide articulation whereas no phonological activation of non-target lemma’s in Chinese. The present findings thus support a discrete model in Chinese spoken word production, which was contrastive with the cascaded view in alphabetic languages production.

Directionality and context effects in word translation tasks performed by conference interpreters

AbstractProfessional interpreters employed by international institutions usually work into their L1 from their L2, while freelance interpreters tend to work both into and from their L1. A study was devised to see if the long-term interpreting unidirectional practice (in the L2–L1 direction only), in contrast to bidirectional practice (in the L2–L1 and L1–L2 direction), influences the speed of lexical retrieval manifested through shorter translation latencies. Forty-eight professional conference interpreters produced oral translations of nouns presented in isolation, in high context constraint sentences and in low context constraint sentences. Contrary to predictions, unidirectional interpreters did not manifest directionality asymmetry and their L2–L1 translation latencies were not shorter than L1–L2 translation latencies. Surprisingly, the L2–L1 direction advantage was found in the group of bidirectional interpreters. The data suggest that the dominant directionality in interpreting practice has little impact on the strength of interlingual lexical links in the interpreter’s mental lexicon or that other factors (such as language use, exposure and immersion) might offset any such impact. The study also revealed an expected context effect, which shows that interpreters use semantic constraint to anticipate sentence-final words.

Z-MAP

Existing space management and address mapping schemes for flash-based Solid-State-Drive (SSD) operate either at page or block granularity, with inevitable limitations in terms of memory requirement, performance, garbage collection, and scalability. To overcome these limitations, we proposed a novel space management and address mapping scheme for flash referred to as Z-MAP, which manages flash space at granularity of Zone. Each Zone consists of multiple numbers of flash blocks. Leveraging workload classification, Z-MAP explores Page-mapping Zone (Page Zone) to store random data and handle a large number of partial updates, and Block-mapping Zone (Block Zone) to store sequential data and lower the overall mapping table. Zones are dynamically allocated and a mapping scheme for a Zone is determined only when it is allocated. Z-MAP uses a small part of Flash memory or phase change memory as a streaming Buffer Zone to log data sequentially and migrate data into Page Zone or Block Zone based on workload classification. A two-level address mapping is designed to reduce the overall mapping table and address translation latency. Z-MAP classifies data before it is permanently stored into Flash memory so that different workloads can be isolated and garbage collection overhead can be minimized. Z-MAP has been extensively evaluated by trace-driven simulation and a prototype implementation on OpenSSD. Our benchmark results conclusively demonstrate that Z-MAP can achieve up to 76% performance improvement, 81% mapping table reduction, and 88% garbage collection overhead reduction compared to existing Flash Translation Layer (FTL) schemes.

PS-TLB

Traversing the page table during virtual to physical address translation causes pipeline stalls when misses occur in the translation-lookaside buffer (TLB). State-of-the-art translation proposals typically optimize a single aspect of translation performance (e.g., translation sharing, context switch performance, etc.) with potential trade-offs of additional hardware complexity, increased translation latency, or reduced scalability. In this article, we propose the partial sharing TLB (PS-TLB), a fast and scalable solution that reduces off-chip translation misses without sacrificing the timing-critical requirement of on-chip translation. We introduce the partial sharing buffer (PSB) which leverages application page sharing characteristics using minimal additional hardware resources. Compared to the leading TLB proposal that leverages sharing, PS-TLB provides a more than 45% improvement in translation latency with a 9% application speedup while using fewer storage resources. In addition, the page classification and PS-TLB architecture provide further optimizations including an over 30% reduction of interprocessor interrupts for coherence, and reduced context switch misses with fewer resources compared with existing methods.

The Distractor Picture Paradox in Speech Production: Evidence from the Word Translation Task

Several naming studies show that distractor pictures, even when intentionally ignored by the speaker, are still capable of activating their respective phonological representations. However results from word translation studies suggest that distractor pictures are only conceptually activated. Here we tested the reliability of the word translation results. In four experiments, bilingual participants translated words from one language to the other one while ignoring the presentation of pictures. In Experiment 1a phonological related pictures sped up translation latencies. However, the effect disappeared when the percentage of related trials was reduced (Experiment 1b). In Experiment 2a translation latencies were faster when the words were accompanied by semantically related pictures than by unrelated pictures. Importantly, the effect was still reliable when the proportion of related trials was reduced and the total number of semantic categories was increased (Experiment 2b). Theoretical implications of the influence of distractor pictures during speech production are discussed.

Number of meanings and concreteness: Consequences of ambiguity within and across languages

We examined the effects of concreteness and ambiguity on language processing. In Experiment 1, English–Spanish bilinguals translated words with a single translation. Contrary to past findings, we observed no concrete-word advantage in translation latency. In Experiment 2, English–Spanish bilinguals translated words with one and more than one translation. Words with multiple translations were translated more slowly and showed the typical concrete-word advantage. Words with one translation showed a reversal of the typical concrete-word advantage in latency. Further, concrete words were uninfluenced by ambiguity. In Experiment 3, we explored whether the interaction between concreteness and ambiguity was a general property of the language processing system. Supporting this idea, in a monolingual lexical decision task, we found an interaction between concreteness and number of meanings analogous to the interaction in translation. We discuss the common mechanism that may lead to this interaction in both within and cross-language processing.

A chatterbox is a box: Morphology in German word production

Two experiments are reported in which university students translated visually presented English words into German, while German distractor words were simultaneously presented. Distractors were morphologically related, merely form-related or unrelated to the German translations (target words). The transparency of the semantic relation between target words and morphological distractors was also varied. Morphological distractors facilitated word-translation latencies irrespective of their semantic transparency, replicating results obtained with other tasks. Thus, in German word production, effects of morphological complexity seem to be largely independent of semantics. Morphological facilitation is also not due to mere form relatedness, since phonological distractors had no impact on translation latencies, relative to unrelated distractors. Our data corroborate the usefulness of word-translation for investigating spoken word production, in particular, for morphological processing.