Modeling Dual Read/Write Paths for Simultaneous Machine Translation

We present Sequential Policy Optimization for Simultaneous Machine Translation (SeqPO-SiMT), a new policy optimization framework that defines the simultaneous machine translation (SiMT) task as a sequential decision making problem, incorporating a tailored reward to enhance translation quality while reducing latency.In contrast to popular Reinforcement Learning from Human Feedback (RLHF) methods, such as PPO and DPO, which are typically applied in single-step tasks, SeqPO-SiMT effectively tackles the multi-step SiMT task.This intuitive framework allows the SiMT LLMs to simulate and refine the SiMT process using a tailored reward.We conduct experiments on six datasets from diverse domains for En Zh and Zh En SiMT tasks, demonstrating that SeqPO-SiMT consistently achieves significantly higher translation quality with lower latency.In particular, SeqPO-SiMT outperforms the supervised fine-tuning (SFT) model by 1.13 points in COMET, while reducing the Average Lagging by 6.17 in the NEWSTEST2021 En Zh dataset.While SiMT operates with far less context than offline translation, the SiMT results of SeqPO-SiMT on 7B LLM surprisingly rival the offline translation of high-performing LLMs, including Qwen-2.5-7B-Instruct and LLaMA-3-8B-Instruct.

Multi-modal simultaneous machine translation fusion of image information

Simultaneous machine translation has recently gained traction thanks to significant quality improvements and the advent of streaming applications. Simultaneous translation systems need to find a trade-off between translation quality and response time, and with this purpose multiple latency measures have been proposed. However, latency evaluations for simultaneous translation are estimated at the sentence level, not taking into account the sequential nature of a streaming scenario. Indeed, these sentence-level latency measures are not well suited for continuous stream translation resulting in figures that are not coherent with the simultaneous translation policy of the system being assessed. This work proposes a stream-level adaptation of the current latency measures based on a re-segmentation approach applied to the output translation, that is successfully evaluated on streaming conditions for a reference IWSLT task.

Simultaneous machine translation, which aims at a realtime translation, is useful in many live scenarios but very challenging due to the trade-off between accuracy and latency. To achieve the balance for both, the model needs to wait for appropriate streaming text (READ policy) and then generates its translation (WRITE policy). However, WRITE policies of previous work either are specific to the method itself due to the end-to-end training or suffer from the input mismatch between training and decoding for the non-end-to-end training. Therefore, it is essential to learn a generic and better WRITE policy for simultaneous machine translation. Inspired by strategies utilized by human interpreters and "wait" policies, we propose a novel adaptive prefix-to-prefix training policy called LEAPT, which allows our machine translation model to learn how to translate source sentence prefixes and make use of the future context. Experiments show that our proposed methods greatly outperform competitive baselines and achieve promising results.

Compared to fully manual translation, post-editing (PE) machine translation (MT) output can save time and reduce errors. Automatic word-level quality estimation (QE) aims to predict the correctness of words in MT output and holds great promise to aid PE by flagging problematic output. Quality of QE is crucial, as incorrect QE might lead to translators missing errors or wasting time on already correct MT output. Achieving accurate automatic word-level QE is very hard, and it is currently not known (i) at what quality threshold QE is actually beginning to be useful for human PE, and (ii), how to best present word-level QE information to translators. In particular, should word-level QE visualization indicate uncertainty of the QE model or not? In this paper, we address both research questions with real and simulated word-level QE, visualizations, and user studies, where time, subjective ratings, and quality of the final translations are assessed. Results show that current word-level QE models are not yet good enough to support PE. Instead, quality levels of > 80% F1 are required. For helpful quality levels, a visualization reflecting the uncertainty of the QE model is preferred. Our analysis further shows that speed gains achieved through QE are not merely a result of blindly trusting the QE system, but that the quality of the final translations also improves. The threshold results from the paper establish a quality goal for future word-level QE research.

Editing Taiwan divination Verses with controlled Language Strategies: Machine-Translation-Mediated Effective Communication Chung-ling Shih Abstract Aimed at fostering machine-translation-mediated communication across languages and cultures, this paper proposes editing Taiwan divination verses from the natural into controlled language to improve the comprehensibility of machine translation (MT) outputs. After editing 160 divination verses, and evaluating the semantic and grammatical accuracy of their English MTs that are produced by online Google Translate (a neural MT system for free), the author has identified several controlled language strategies. The lexical strategies include replacement of archaic Chinese words with vernacular Chinese ones, paraphrasing of culture references and insertion of explanations for metaphors. Grammatical strategies are the use of articles, determiners and possessive cases, and syntactical ones include the addition of conjunctions and the restoration of missing subjects or/and objects. The MT outputs of edited and unedited divination verses are compared. The findings show that the English MT outputs of edited texts have greatly improved their semantic, grammatical and syntactic accuracy, so the effectiveness of controlled language strategies is justified. Due to the effectiveness of editing verses with controlled language strategies, the goal of MT-enabled web-based communication across cultures is achieved. The practical significance is also discussed including culture acquisition and cost reduction. Full Text: PDF DOI: 10.15640/jflcc.v8n2a2

Digitalization is one of the key distinctive features of modern environment and social life. Nowadays more and more functions are transferred to the artificial mind. How effective is the replacement of human activity with computer activity? In the given article, this problem is solved by an example of integration of digital technologies into translation activities. It this paper, emphasis is placed on the quality of machine translation (MT) output of legal texts in the language pair English - Slovak. It studies a Criminal Code formulated in the Slovak language which was translated by a human translator into English and consequently via machine translation system Google Translate (GT) back into Slovak. The back-translation - translation of a translated text back into its original language - as a quality assessment tool to detect discrepancies, mistranslations and inevitable differences between the source text and the target text was used. The quality of MT output was evaluated according to Multidimensional Quality Metrics (MQM) standards with the focus on the dimension of Fluency. The multiple comparisons were applied to determine which issues (errors) in Fluency dimension differ from the others. A statistically significant difference is noticed between Agreement and other issues, as well as between Ambiguity and other issues. The errors in Agreement are related to the differences between the languages: English is considered mostly an analytic language, Slovak represents a synthetic language. The issues in the Ambiguity dimension correlate with the type of the text being examined, since legal texts are characterized by relatively complicated wording and numerous terms; moreover, accuracy and unambiguity need to be preserved. Generally, the MT output is able to provide users with basic information about the text. On the other hand, most of the segments need revision and/or correction; in such cases, human intervention and post-editing is necessary.

Simultaneous translation is a task in which translation begins before the speaker has finished speaking, so it is important to decide when to start the translation process. However, deciding whether to read more input words or start to translate is difficult for language pairs with different word orders such as English and Japanese. Motivated by the concept of pre-reordering, we propose a couple of simple decision rules using the label of the next constituent predicted by incremental constituent label prediction. In experiments on English-to-Japanese simultaneous translation, the proposed method outperformed baselines in the quality-latency trade-off.

The purpose of this research is to test the efficacy of applying automated evaluation techniques, originally devised for the evaluation of human language learners, to the output of machine translation (MT) systems. We believe that these evaluation techniques will provide information about both the human language learning process, the translation process and the development of machine translation systems. This, the first experiment in a series of experiments, looks at the intelligibility of MT output. A language learning experiment showed that assessors can differentiate native from non-native language essays in less than 100 words. Even more illuminating was the factors on which the assessors made their decisions. We tested this to see if similar criteria could be elicited from duplicating the experiment using machine translation output. Subjects were given a set of up to six extracts of translated newswire text. Some of the extracts were expert human translations, others were machine translation outputs. The subjects were given three minutes per extract to determine whether they believed the sample output to be an expert human translation or a machine translation. Additionally, they were asked to mark the word at which they made this decision. The results of this experiment, along with a preliminary analysis of the factors involved in the decision making process will be presented here.

The present paper describes a machine translation (MT) course taught to undergraduate students in the Department of English Language and Literature at Dhofar University in Oman. The course is one of the major requirements for BA in Translation. Fifteen EFL translation students who were in their third year of study were enrolled in the course. The author presents both the theoretical and practical parts of the course. In the theoretical part, the topics covered in the course are outlined. As for the practical part, it focuses on the translation students’ post-editing of online MT output. This is beneficial to the students as free online MT systems can potentially be used as a means for improving student translators’ training and EFL learning. This is achieved through subjecting MT output to analysis or post-editing by the students so that they can focus on the differences between the source and target languages. With this goal in mind, assignments were given to the students to post-edit the Arabic and English MT output of three free online MT systems (Systran, Babylon and Google Translate), discuss the linguistic problems that they spot for each system and choose the one that has the fewest number of errors. The results show that the students, with varying degrees of success, have managed to identify some linguistic errors with the MT output for each MT system and thus produced a better human translation. The paper concludes that there is a need to incorporate MT courses in translation departments in the Arab world, as integrating technology into translation curricula will have great effect on student translators’ training for their future career as professional translators.

We propose a generative framework for simultaneous machine translation. Conventional approaches use a fixed number of source words to translate or learn dynamic policies for the number of source words by reinforcement learning. Here we formulate simultaneous translation as a structural sequence-to-sequence learning problem. A latent variable is introduced to model read or translate actions at every time step, which is then integrated out to consider all the possible translation policies. A re-parameterised Poisson prior is used to regularise the policies which allows the model to explicitly balance translation quality and latency. The experiments demonstrate the effectiveness and robustness of the generative framework, which achieves the best BLEU scores given different average translation latencies on benchmark datasets.

This chapter presents the findings from a process-oriented study on institutional translators’ translation processes focusing on their allocation of cognitive resources and the types of activities performed and their use of digital resources during translation. Screen recorded data, captured online and remotely from seven institutional translators working in-house at the European Union (EU), are analyzed. The findings show that institutional translators spend only half of the time drafting translation, and that a considerable amount of time is spent consulting digital resources, the majority of which are developed and maintained by the EU institutions. In addition, the recorded screen data show that legal domain text more frequently prompts resource consultation than press release text. Automatically generated suggestions by the institutions’ translation tool demand the translators’ evaluation of translation memory and machine translation output. As much as 50% of translation drafting involves evaluation of automated translation output, whereas from scratch translation is considerably less prominent. The study’s findings provide a unique process-oriented perspective on what competences are necessary in an institutional translation context, highlighting the need to develop specialized technological and post-editing competences.

Translation Environment Tools make translators’ work easier by providing them with term lists, translation memories and machine translation output. Ideally, such tools automatically predict whether it is more effortful to post-edit than to translate from scratch, and determine whether or not to provide translators with machine translation output. Current machine translation quality estimation systems heavily rely on automatic metrics, even though they do not accurately capture actual post-editing effort. In addition, these systems do not take translator experience into account, even though novices’ translation processes are different from those of professional translators. In this paper, we report on the impact of machine translation errors on various types of post-editing effort indicators, for professional translators as well as student translators. We compare the impact of MT quality on a product effort indicator (HTER) with that on various process effort indicators. The translation and post-editing process of student translators and professional translators was logged with a combination of keystroke logging and eye-tracking, and the MT output was analyzed with a fine-grained translation quality assessment approach. We find that most post-editing effort indicators (product as well as process) are influenced by machine translation quality, but that different error types affect different post-editing effort indicators, confirming that a more fine-grained MT quality analysis is needed to correctly estimate actual post-editing effort. Coherence, meaning shifts, and structural issues are shown to be good indicators of post-editing effort. The additional impact of experience on these interactions between MT quality and post-editing effort is smaller than expected.

Simultaneous machine translation consists in starting output generation before the entire input sequence is available. Wait-k decoders offer a simple but efficient approach for this problem. They first read k source tokens, after which they alternate between producing a target token and reading another source token. We investigate the behavior of wait-k decoding in low resource settings for spoken corpora using IWSLT datasets. We improve training of these models using unidirectional encoders, and training across multiple values of k. Experiments with Transformer and 2D-convolutional architectures show that our wait-k models generalize well across a wide range of latency levels. We also show that the 2D-convolution architecture is competitive with Transformers for simultaneous translation of spoken language.

While the productivity gains brought about by machine translation (MT) can help translators meet ever-tighter deadlines and respond to pressing demands for publishing content simultaneously in different languages, these tools also impose a workflow that tends to reduce the human translator's role to simply correcting mistakes made by the machine in a one-way process with no real interaction. Thus, although more cost-effective, post-editing of MT output also appears a less creative and enjoyable a task than translation. Adaptive MT, on the other hand, has been advertised as a way to recenter the translation process on the human and foster more genuine interaction with the machine. Said to have been developed for professional translation workflows, the technology enables a dynamic work process that is supposedly very different from the repetitive task that post-editing static MT output can be. This paper presents an experiment with adaptive MT conducted during the 2020-2021 academic year. As part of a course on MT and post-editing, second-year master's students conducted group projects on the Lilt platform. In this paper, students' views on the MT engine are analyzed, with a focus on their interaction with the technology. While students recognize the potential of adaptive MT for empowering the human in the loop, MT quality and CAT ergonomics in general appear to have a greater influence on usability than interaction with the machine.