With contributions from the open-source community, a vast amount of instruction tuning (IT) data has emerged. Given the significant resource allocation required by training and evaluating models, it is advantageous to have an efficient method for selecting high-quality IT data. However, existing methods for instruction data selection have limitations such as relying on fragile external APIs, being affected by biases in GPT models, or reducing the diversity of the selected instruction dataset. In this paper, we propose an industrial-friendly, expert-aligned and diversity-preserved instruction data selection method: Clustering and Ranking (CaR). CaR consists of two steps. The first step involves ranking instruction pairs using a scoring model that is well aligned with expert preferences (achieving an accuracy of 84.25%). The second step involves preserving dataset diversity through a clustering process. In our experiment, CaR selected a subset containing only 1.96% of Alpaca’s IT data, yet the underlying AlpaCaR model trained on this subset outperforms Alpaca by an average of 32.1% in GPT-4 evaluations. Furthermore, our method utilizes small models (550M parameters) and requires only 11.2% of the monetary cost compared to existing methods, making it easily deployable in industrial scenarios.
Audio deepfake detection (ADD) is essential for preventing the misuse of synthetic voices that may infringe on personal rights and privacy. Recent zero-shot text-to-speech (TTS) models pose higher risks as they can clone voices with a single utterance. However, the existing ADD datasets are outdated, leading to suboptimal generalization of detection models. In this paper, we construct a new cross-domain ADD dataset comprising over 300 hours of speech data that is generated by five advanced zero-shot TTS models. To simulate real-world scenarios, we employ diverse attack methods and audio prompts from different datasets. Experiments show that, through novel attack-augmented training, the Wav2Vec2-large and Whisper-medium models achieve equal error rates of 4.1% and 6.5% respectively. Additionally, we demonstrate our models’ outstanding few-shot ADD ability by fine-tuning with just one minute of target-domain data. Nonetheless, neural codec compressors greatly affect the detection accuracy, necessitating further research. Our dataset is publicly available (https://github.com/leolya/CD-ADD).
Large Multimodal Models (LMMs) have achieved great success recently, demonstrating a strong capability to understand multimodal information and to interact with human users. Despite the progress made, the challenge of detecting high-risk interactions in multimodal settings, and in particular in speech modality, remains largely unexplored. Conventional research on risk for speech modality primarily emphasises the content (e.g., what is captured as transcription). However, in speech-based interactions, paralinguistic cues in audio can significantly alter the intended meaning behind utterances. In this work, we propose a speech-specific risk taxonomy, covering 8 risk categories under hostility (malicious sarcasm and threats), malicious imitation (age, gender, ethnicity), and stereotypical biases (age, gender, ethnicity). Based on the taxonomy, we create a small-scale dataset for evaluating current LMMs capability in detecting these categories of risk. We observe even the latest models remain ineffective to detect various paralinguistic-specific risks in speech (e.g., Gemini 1.5 Pro is performing only slightly above random baseline). Warning: this paper contains biased and offensive examples.
In this paper, we present an effective method for TextGraphs-17 Shared Task. This task requires selecting an entity from the candidate entities that is relevant to the given question and answer. The selection process is aided by utilizing the shortest path graph in the knowledge graph, connecting entities in the query to the candidate entity. This task aims to explore how to enhance LLMs output with KGs, although current LLMs have certain logical reasoning capabilities, they may not be certain about their own outputs, and the answers they produce may be correct by chance through incorrect paths. In this case, we have introduced a LLM prompt design strategy based on self-ranking and emotion. Specifically, we let the large model score its own answer choices to reflect its confidence in the answer. Additionally, we add emotional incentives to the prompts to encourage the model to carefully examine the questions. Our submissions was conducted under zero-resource setting, and we achieved the second place in the task with an F1-score of 0.8321.
This paper presents a study on strategies to enhance the translation capabilities of large language models (LLMs) in the context of machine translation (MT) tasks. The paper proposes a novel paradigm consisting of three stages: Secondary Pre-training using Extensive Monolingual Data, Continual Pre-training with Interlinear Text Format Documents, and Leveraging Source-Language Consistent Instruction for Supervised Fine-Tuning. Previous research on LLMs focused on various strategies for supervised fine-tuning (SFT), but their effectiveness has been limited. While traditional machine translation approaches rely on vast amounts of parallel bilingual data, our paradigm highlights the importance of using smaller sets of high-quality bilingual data. We argue that the focus should be on augmenting LLMs’ cross-lingual alignment abilities during pre-training rather than solely relying on extensive bilingual data during SFT. Experimental results conducted using the Llama2(CITATION)model, particularly on Chinese-Llama2(CITATION) after monolingual augmentation, demonstrate the improved translation capabilities of LLMs. A significant contribution of our approach lies in Stage2: Continual Pre-training with Interlinear Text Format Documents, which requires less than 1B training data, making our method highly efficient. Additionally, in Stage3, we observed that setting instructions consistent with the source language benefits the supervised fine-tuning process. Experimental results demonstrate that our approach surpasses previous work and achieves superior performance compared to models such as NLLB-54B(CITATION) and GPT3.5-text-davinci-003, despite having a significantly smaller parameter count of only 7B or 13B. This achievement establishes our method as a pioneering strategy in the field of machine translation.
Knowledge Graph Embedding (KGE) is a powerful technique for predicting missing links in Knowledge Graphs (KGs) by learning the entities and relations. Hyperbolic space has emerged as a promising embedding space for KGs due to its ability to represent hierarchical data. Nevertheless, most existing hyperbolic KGE methods rely on tangent approximation and are not fully hyperbolic, resulting in distortions and inaccuracies. To overcome this limitation, we propose LorentzKG, a fully hyperbolic KGE method that represents entities as points in the Lorentz model and represents relations as the intrinsic transformation—the Lorentz transformations between entities. We demonstrate that the Lorentz transformation, which can be decomposed into Lorentz rotation/reflection and Lorentz boost, captures various types of relations including hierarchical structures. Experimental results show that our LorentzKG achieves state-of-the-art performance.
Automatic dubbing aims to translate the speech of a video into another language, ensuring the new speech naturally fits the original video. This paper details Huawei Translation Services Center’s (HW-TSC) submission for IWSLT 2024’s automatic dubbing task, under an unconstrained setting. Our system’s machine translation (MT) component utilizes a Transformer-based MT model and an LLM-based post-editor to produce translations of varying lengths. The text-to-speech (TTS) component employs a VITS-based TTS model and a voice cloning module to emulate the original speaker’s vocal timbre. For enhanced dubbing synchrony, we introduce a parsing-informed pause selector. Finally, we rerank multiple results based on lip-sync error distance (LSE-D) and character error rate (CER). Our system achieves LSE-D of 10.75 and 12.19 on subset1 and subset2 of DE-EN test sets respectively, superior to last year’s best system.
This paper presents HW-TSC’s submission to the IWSLT 2024 Offline Speech Translation Task and Speech-to-Speech Translation Task. The former includes three translation directions: English to German, English to Chinese, and English to Japanese, while the latter only includes the translation direction of English to Chinese. We attend all three tracks (Constraint training, Constrained with Large Language Models training, and Unconstrained training) of offline speech translation task, using the cascade model architecture. Under the constrained training track, we train an ASR model from scratch, and then employ R-Drop and domain data selection to train the NMT model. In the constrained with Large Language Models training track, we use Wav2vec 2.0 and mBART50 for ASR model training initialization, and then train the LLama2-7B-based MT model using continuous training with sentence-aligned parallel data, supervised fine-tuning, and contrastive preference optimization. In the unconstrained training track, we fine-tune the whisper model for speech recognition, and then ensemble the translation results of NMT models and LLMs to produce superior translation output. For the speech-to-speech translation Task, we initially employ the offline speech translation system described above to generate the translated text. Then, we utilize the VITS model to generate the corresponding speech and employ the OpenVoice model for timbre cloning.
This article introduces the process of HW-TSC and the results of IWSLT 2024 Indic Track Speech to Text Translation. We designed a cascade system consisting of an ASR model and a machine translation model to translate speech from one language to another. For the ASR part, we directly use whisper large v3 as our ASR model. Our main task is to optimize the machine translation model (en2ta, en2hi, en2bn). In the process of optimizing the translation model, we first use bilingual corpus to train the baseline model. Then we use monolingual data to construct pseudo-corpus data to further enhance the baseline model. Finally, we filter the parallel corpus data through the labse filtering method and finetune the model again, which can further improve the bleu value. We also selected domain data from bilingual corpus to finetune previous model to achieve the best results.
In this work, we submitted our systems to the low-resource track of the IWSLT 2024 Speech Translation Campaign. Our systems tackled the unconstrained condition of the Dialectal Arabic North Levantine (ISO-3 code: apc) to English language pair. We proposed a cascaded solution consisting of an automatic speech recognition (ASR) model and a machine translation (MT) model. It was noted that the ASR model employed the pre-trained Whisper-large-v3 model to process the speech data, while the MT model adopted the Transformer architecture. To improve the quality of the MT model, it was stated that our system utilized not only the data provided by the competition but also an additional 54 million parallel sentences. Ultimately, we reported that our final system achieved a BLEU score of 24.7 for apc-to-English translation.
This paper outlines our submission for the IWSLT 2024 Simultaneous Speech-to-Text (SimulS2T) and Speech-to-Speech (SimulS2S) Translation competition. We have engaged in all four language directions and both the SimulS2T and SimulS2S tracks: English-German (EN-DE), English-Chinese (EN-ZH), English-Japanese (EN-JA), and Czech-English (CS-EN). For the S2T track, we have built upon our previous year’s system and further honed the cascade system composed of ASR model and MT model. Concurrently, we have introduced an end-to-end system specifically for the CS-EN direction. This end-to-end (E2E) system primarily employs the pre-trained seamlessM4T model. In relation to the SimulS2S track, we have integrated a novel TTS model into our SimulS2T system. The final submission for the S2T directions of EN-DE, EN-ZH, and EN-JA has been refined over our championship system from last year. Building upon this foundation, the incorporation of the new TTS into our SimulS2S system has resulted in the ASR-BLEU surpassing last year’s best score.
This paper introduces HW-TSC’s submission to the IWSLT 2024 Subtitling track. For the automatic subtitling track, we use an unconstrained cascaded strategy, with the main steps being: ASR with word-level timestamps, sentence segmentation based on punctuation restoration, further alignment using CTC or using machine translation with length penalty. For the subtitle compression track, we employ a subtitle compression strategy that integrates machine translation models and extensive rewriting models. We acquire the subtitle text requiring revision through the CPS index, then utilize a translation model to obtain the English version of this text. Following this, we extract the compressed-length subtitle text through controlled decoding. If this method fails to compress the text successfully, we resort to the Llama2 few-shot model for further compression.
This paper presents the submission of Huawei Translate Services Center (HW-TSC) to the WMT24 general machine translation (MT) shared task, where we participate in the English to Chinese (en→zh) language pair. Similar to previous years’ work, we use training strategies such as regularized dropout, bidirectional training, data diversification, forward translation, back translation, alternated training, curriculum learning, and transductive ensemble learning to train the neural machine translation (NMT) model based on the deep Transformer-big architecture. The difference is that we also use continue pre-training, supervised fine-tuning, and contrastive preference optimization to train the large language model (LLM) based MT model. By using Minimum Bayesian risk (MBR) decoding to select the final translation from multiple hypotheses for NMT and LLM-based MT models, our submission receives competitive results in the final evaluation.
Quality estimation (QE) is a crucial technique for evaluating the quality of machine translations without the need for reference translations. This paper focuses on Huawei Translation Services Center’s (HW-TSC’s) submission to the sentence-level QE shared task, named LLMs-enhanced-CrossQE. Our system builds upon the CrossQE architecture from our submission from last year, which consists of a multilingual base model and a task-specific downstream layer. The model input is a concatenation of the source and the translated sentences. To enhance performance, we fine-tuned and ensembled multiple base models, including XLM-R, InfoXLM, RemBERT, and CometKiwi. Specifically, we employed two pseudo-data generation methods: 1) a diverse pseudo-data generation method based on the corruption-based data augmentation technique introduced last year, and 2) a pseudo-data generation method that simulates machine translation errors using large language models (LLMs). Our results demonstrate that the system achieves outstanding performance on sentence-level QE test sets.
The paper presents the submission by HW-TSC in the WMT 2024 Quality-informed Automatic Post Editing (QEAPE) shared task for the English-Hindi (En-Hi) and English-Tamil (En-Ta) language pair. We use LLM for En-Hi and Transformer for EN-ta respectively. For LLM, we first continue pertrain the Llama3, and then use the real APE data to SFT the pre-trained LLM. As for the transformer in En-Ta, we first pre-train a Machine Translation (MT) model by utilizing MT data collected from the web. Then, we fine-tune the model by employing real APE data.We also use the data augmentation method to enhance our model. Specifically, we incorporate candidate translations obtained from an external Machine Translation (MT) system.Given that APE systems tend to exhibit a tendency of ‘over-correction’, we employ a sentence-level Quality Estimation (QE) system to select the final output, deciding between the original translation and the corresponding output generated by the APE model. Our experiments demonstrate that pre-trained MT models are effective when being fine-tuned with the APE corpus of a limited size, and the performance can be further improved with external MT augmentation. our approach improves the HTER by -15.99 points and -0.47 points on En-Hi and En-Ta, respectively.
This paper introduces the submission by Huawei Translation Center (HW-TSC) to the WMT24 Indian Languages Machine Translation (MT) Shared Task. To develop a reliable machine translation system for low-resource Indian languages, we employed two distinct knowledge transfer strategies, taking into account the characteristics of the language scripts and the support available from existing open-source models for Indian languages. For Assamese(as) and Manipuri(mn), we fine-tuned the existing IndicTrans2 open-source model to enable bidirectional translation between English and these languages. For Khasi(kh) and Mizo(mz), we trained a multilingual model as the baseline using bilingual data from this four language pairs as well as additional Bengali data, which share the same language family. This was followed by fine-tuning to achieve bidirectional translation between English and Khasi, as well as English and Mizo. Our transfer learning experiments produced significant results: 23.5 BLEU for en→as, 31.8 BLEU for en→mn, 36.2 BLEU for as→en, and 47.9 BLEU for mn→en on their respective test sets. Similarly, the multilingual model transfer learning experiments yielded impressive outcomes, achieving 19.7 BLEU for en→kh, 32.8 BLEU for en→mz, 16.1 BLEU for kh→en, and 33.9 BLEU for mz→en on their respective test sets. These results not only highlight the effectiveness of transfer learning techniques for low-resource languages but also contribute to advancing machine translation capabilities for low-resource Indian languages.
This article introduces the submission status of the Translation into Low-Resource Languages of Spain task at (WMT 2024) by Huawei Translation Service Center (HW-TSC). We participated in three translation tasks: spanish to aragonese (es2arg), spanish to aranese (es2arn), and spanish to asturian (es2ast). For these three translation tasks, we use training strategies such as multilingual transfer, regularized dropout, forward translation and back translation, labse denoising, transduction ensemble learning and other strategies to neural machine translation (NMT) model based on training deep transformer-big architecture. By using these enhancement strategies, our submission achieved a competitive result in the final evaluation.
This report outlines our approach for the WMT24 Discourse-Level Literary Translation Task, focusing on the Chinese-English language pair in the Constrained Track. Translating literary texts poses significant challenges due to the nuanced meanings, idiomatic expressions, and intricate narrative structures inherent in such works. To address these challenges, we leveraged the Chinese-Llama2 model, specifically enhanced for this task through a combination of Continual Pre-training (CPT) and Supervised Fine-Tuning (SFT). Our methodology includes a novel Incremental Decoding framework, which ensures that each sentence is translated with consideration of its broader context, maintaining coherence and consistency throughout the text. This approach allows the model to capture long-range dependencies and stylistic elements, producing translations that faithfully preserve the original literary quality. Our experiments demonstrate significant improvements in both sentence-level and document-level BLEU scores, underscoring the effectiveness of our proposed framework in addressing the complexities of document-level literary translation.
This paper describes the submissions of Huawei Translation Services Center(HW-TSC) to WMT24 chat translation shared task on English↔Germany (en-de) bidirection. The experiments involved fine-tuning models using chat data and exploring various strategies, including Minimum Bayesian Risk (MBR) decoding and self-training. The results show significant performance improvements in certain directions, with the MBR self-training method achieving the best results. The Large Language Model also discusses the challenges and potential avenues for further research in the field of chat translation.
End-to-end automatic speech recognition (ASR) systems often struggle to recognize rare name entities, such as personal names, organizations and terminologies that are not frequently encountered in the training data. This paper presents Contextual Biasing Whisper (CB-Whisper), a novel ASR system based on OpenAI’s Whisper model that can recognize user-defined name entities by performing open-vocabulary keyword-spotting (KWS) before the decoder. The KWS module leverages text-to-speech (TTS) techniques and a convolutional neural network (CNN) classifier to match the features between the entities and the utterances. To integrate the recognized entities into the Whipser decoder and avoid hallucinations, we carefully crafted multiple prompts with spoken form hints. Experiments show that the KWS module based on Whisper encoder’s features can recognize unseen user-defined keywords effectively. More importantly, the proposed CB-Whisper substantially improves the mixed-error-rate (MER) and entity recall compared to the original Whisper model on three internal datasets and two publicly available datasets including Aishell and ACL datasets that cover English-only, Chinese-only, and code-switching scenarios.
Lexical translation consistency is one of the most common discourse phenomena in Chinese-to-English document-level translation. To better evaluate the performance of lexical translation consistency, previous researches assumes that all repeated source words should be translated consistently. However, constraining translations of repeated source words to be consistent will hurt word diversity and human translators tend to use different words in translation. Therefore, in this paper we construct a test set of 310 bilingual news articles to properly evaluate lexical translation consistency. We manually differentiate those repeated source words whose translations are consistent into two types: true consistency and false consistency. Then based on the constructed test set, we evaluate the performance of lexical translation consistency for several typical NMT systems.
Large Language Models (LLMs) have demonstrated impressive performances across various NLP tasks with just a few prompts via in-context learning. Previous studies have emphasized the pivotal role of well-chosen examples in in-context learning, as opposed to randomly selected instances that exhibits unstable results.A successful example selection scheme depends on multiple factors, while in the context of LLMs-based machine translation, the common selection algorithms only consider the single factor, i.e., the similarity between the example source sentence and the input sentence.In this paper, we introduce a novel approach to use multiple translational factors for in-context example selection by using monotone submodular function maximization.The factors include surface/semantic similarity between examples and inputs on both source and target sides, as well as the diversity within examples.Importantly, our framework mathematically guarantees the coordination between these factors, which are different and challenging to reconcile.Additionally, our research uncovers a previously unexamined dimension: unlike other NLP tasks, the translation part of an example is also crucial, a facet disregarded in prior studies.Experiments conducted on BLOOMZ-7.1B and LLAMA2-13B, demonstrate that our approach significantly outperforms random selection and robust single-factor baselines across various machine translation tasks.
The degree of semantic relatedness of two units of language has long been considered fundamental to understanding meaning. In this paper, we present the system of Huawei Translation Services Center (HW-TSC) for Task 1 of SemEval 2024, which aims to automatically measure the semantic relatedness of sentence pairs in African and Asian languages. The task dataset for this task covers about 14 different languages, These languages originate from five distinct language families and are predominantly spoken in Africa and Asia. For this shared task, we describe our proposed solutions, including ideas and the implementation steps of the task, as well as the outcomes of each experiment on the development dataset. To enhance the performance, we leverage these experimental outcomes and construct an ensemble one. Our results demonstrate that our system achieves impressive performance on test datasets in unsupervised track B and ranked first place for the Punjabi language pair.
Back Translation (BT) is widely used in the field of machine translation, as it has been proved effective for enhancing translation quality. However, BT mainly improves the translation of inputs that share a similar style (to be more specific, translation-liked inputs), since the source side of BT data is machine-translated. For natural inputs, BT brings only slight improvements and sometimes even adverse effects. To address this issue, we propose Text Style Transfer Back Translation (TST BT), which uses a style transfer to modify the source side of BT data. By making the style of source-side text more natural, we aim to improve the translation of natural inputs. Our experiments on various language pairs, including both high-resource and low-resource ones, demonstrate that TST BT significantly improves translation performance against popular BT benchmarks. In addition, TST BT is proved to be effective in domain adaptation so this strategy can be regarded as a generalized data augmentation method. Our training code and text style transfer model are open-sourced.
This paper presents the submission of Huawei Translation Services Center for the IWSLT 2023 dubbing task in the unconstrained setting. The proposed solution consists of a Transformer-based machine translation model and a phoneme duration predictor. The Transformer is deep and multiple target-to-source length-ratio class labels are used to control target lengths. The variation predictor in FastSpeech2 is utilized to predict phoneme durations. To optimize the isochrony in dubbing, re-ranking and scaling are performed. The source audio duration is used as a reference to re-rank the translations of different length-ratio labels, and the one with minimum time deviation is preferred. Additionally, the phoneme duration outputs are scaled within a defined threshold to narrow the duration gap with the source audio.
This paper presents Huawei Translation Service Center (HW-TSC)’s submission on the IWSLT 2023 formality control task, which provides two training scenarios: supervised and zero-shot, each containing two language pairs, and sets constrained and unconstrained conditions. We train the formality control models for these four language pairs under these two conditions respectively, and submit the corresponding translation results. Our efforts are divided into two fronts: enhancing general translation quality and improving formality control capability. According to the different requirements of the formality control task, we use a multi-stage pre-training method to train a bilingual or multilingual neural machine translation (NMT) model as the basic model, which can improve the general translation quality of the base model to a relatively high level. Then, under the premise of affecting the general translation quality of the basic model as little as possible, we adopt domain adaptation and reranking-based transductive learning methods to improve the formality control capability of the model.
This paper presents HW-TSC’s submissions to the IWSLT 2023 Offline Speech Translation task, including speech translation of talks from English to German, Chinese, and Japanese, respectively. We participate in all three conditions (constrained training, constrained with large language models training, and unconstrained training) with models of cascaded architectures. We use data enhancement, pre-training models and other means to improve the ASR quality, and R-Drop, deep model, domain data selection, etc. to improve the translation quality. Compared with last year’s best results, we achieve 2.1 BLEU improvement on the MuST-C English-German test set.
This paper describes our work on the IWSLT2023 Speech-to-Speech task. Our proposed cascaded system consists of an ensemble of Conformer and S2T-Transformer-based ASR models, a Transformer-based MT model, and a Diffusion-based TTS model. Our primary focus in this competition was to investigate the modeling ability of the Diffusion model for TTS tasks in high-resource scenarios and the role of TTS in the overall S2S task. To this end, we proposed DTS, an end-to-end diffusion-based TTS model that takes raw text as input and generates waveform by iteratively denoising on pure Gaussian noise. Compared to previous TTS models, the speech generated by DTS is more natural and performs better in code-switching scenarios. As the training process is end-to-end, it is relatively straightforward. Our experiments demonstrate that DTS outperforms other TTS models on the GigaS2S benchmark, and also brings positive gains for the entire S2S system.
In this paper, we present our submission to the IWSLT 2023 Simultaneous Speech-to-Text Translation competition. Our participation involves three language directions: English-German, English-Chinese, and English-Japanese. Our proposed solution is a cascaded incremental decoding system that comprises an ASR model and an MT model. The ASR model is based on the U2++ architecture and can handle both streaming and offline speech scenarios with ease. Meanwhile, the MT model adopts the Deep-Transformer architecture. To improve performance, we explore methods to generate a confident partial target text output that guides the next MT incremental decoding process. In our experiments, we demonstrate that our simultaneous strategies achieve low latency while maintaining a loss of no more than 2 BLEU points when compared to offline systems.
In this paper, we present our submission to the IWSLT 2023 Simultaneous Speech-to-Speech Translation competition. Our participation involves three language directions: English-German, English-Chinese, and English-Japanese. Our solution is a cascaded incremental decoding system, consisting of an ASR model, an MT model, and a TTS model. By adopting the strategies used in the Speech-to-Text track, we have managed to generate a more confident target text for each audio segment input, which can guide the next MT incremental decoding process. Additionally, we have integrated the TTS model to seamlessly reproduce audio files from the translation hypothesis. To enhance the effectiveness of our experiment, we have utilized a range of methods to reduce error conditions in the TTS input text and improve the smoothness of the TTS output audio.
Prompt tuning has become a new paradigm for model tuning and it has demonstrated success in natural language pretraining and even vision pretraining. The parameter-efficient prompt tuning methods that optimize soft embeddings while keeping the pretrained model frozen demonstrate advantages in low computation costs and almost lossless performance. In this work, we explore the transfer of prompt tuning to multimodal pretrained models. Specifically, we implement prompt tuning to a unified sequence-to-sequence pretrained model by adding a sequence of learnable embeddings to each layer and finetuning the pretrained model on downstream task with only the learnable embeddings being optimized. Experimental results on a series of multimodal understanding and generation tasks demonstrate that our method OFA-PT can achieve comparable performance with finetuning across a series of multimodal generation and understanding tasks. Additionally, it significantly outperforms the unified multimodal pretrained model with other parameter-efficient tuning methods, e.g., Adapter, BitFit. etc. Besides, in comparison with finetuned models, the prompt-tuned models demonstrate improved robustness against adversarial attacks. We further figure out that experimental factors, including prompt length, prompt depth, and reparameteratization, have great impacts on the model performance, and thus we empirically provide a recommendation for the setups of prompt tuning.
Following the idea of “one translation per discourse”, in this paper we aim to improve translation consistency via document-level translation repair (DocRepair), i.e., automatic post-editing on translations of documents. To this end, we propose a lexical translation inconsistency-aware DocRepair to explicitly model translation inconsistency. First we locate the inconsistency in automatic translation. Then we provide translation candidates for those inconsistency. Finally, we propose lattice-like input to properly model inconsistent tokens and phrases and their candidates. Experimental results on three document-level translation datasets show that based on G-Transformer, a state-of-the-art document-to-document (Doc2Doc) translation model, our Doc2Doc DocRepair achieves significant improvement on translation quality in BLEU scores, but also greatly improves lexical translation consistency.
Named Entity Recognition (NER) is one of the most fundamental tasks in natural language processing. Span-level prediction (SpanNER) is more naturally suitable for nested NER than sequence labeling (SeqLab). However, according to our experiments, the SpanNER method is more sensitive to the amount of training data, i.e., the F1 score of SpanNER drops much more than that of SeqLab when the amount of training data drops. In order to improve the robustness of SpanNER in low resource scenarios, we propose a simple and effective method SmartSpanNER, which introduces a Named Entity Head (NEH) prediction task to SpanNER and performs multi-task learning together with the task of span classification. Experimental results demonstrate that the robustness of SpanNER could be greatly improved by SmartSpanNER in low resource scenarios constructed on the CoNLL03, Few-NERD, GENIA and ACE05 standard benchmark datasets.
Tabular mathematical reasoning task requires models to perform multi-step operations including information look-up and numerical calculation, based on heterogeneous data from tables and questions. Existing solutions tend to extend chain-of-thought (CoT) reasoning into powerful large language models (LLMs) to promote multi-hop mathematical reasoning. However, such LLM-based approaches are not a viable solution in the scenario of privatization deployment or limited resources. To address this problem, we revisit small-scale tabular language models (TaLMs) and extend chain-of-thought reasoning into TaLMs for the first time. Specifically, we propose a novel framework, TaCo, which coordinates two TaLMs responsible for CoT generation and answer inference, respectively. Besides, our framework can be combined with an external calculator to enhance accurate numerical calculation. On the TABMWP dataset, TaCo outperforms the state-of-the-art ChatGPT by 9.55% (82.60%→92.15% in accuracy) with much less parameters (0.8B). The code will be released along with the paper.
Computer-aided translation (CAT) aims to enhance human translation efficiency and is still important in scenarios where machine translation cannot meet quality requirements. One fundamental task within this field is Word-Level Auto Completion (WLAC). WLAC predicts a target word given a source sentence, translation context, and a human typed character sequence. Previous works either employ word classification models to exploit contextual information from both sides of the target word or directly disregarded the dependencies from the right-side context. Furthermore, the key information, i.e. human typed sequences, is only used as prefix constraints in the decoding module. In this paper, we propose the INarIG (Iterative Non-autoregressive Instruct Generation) model, which constructs the human typed sequence into Instruction Unit and employs iterative decoding with subwords to fully utilize input information given in the task. Our model is more competent in dealing with low-frequency words (core scenario of this task), and achieves state-of-the-art results on the WMT22 and benchmark datasets, with a maximum increase of over 10% prediction accuracy.
Machine translation (MT) quality estimation (QE) is a crucial task to estimate the quality of MT outputs when reference translations are unavailable. Many studies focus on generating pseudo data using large parallel corpus and achieve remarkable success in the supervised setting. However, pseudo data solutions are less satisfying in unsupervised scenarios because the pseudo labels are inaccurate or the pseudo translations differ from the real ones. To address these problems, we propose to generate pseudo data using the MT model with constrained beam search (CBSQE). CBSQE preserves the reference parts with high MT probabilities as correct translations, while the rest parts as the wrong ones for MT generation. Therefore, CBSQE can reduce the false negative labels caused by synonyms. Overall, beam search will prefer a more real hypothesis with a higher MT generation likelihood. Extensive experiments demonstrate that CBSQE outperforms strong baselines in both supervised and unsupervised settings. Analyses further show the superiority of CBSQE. The code is available at https://github.com/NJUNLP/njuqe.
Despite the subjective nature of semantic textual similarity (STS) and pervasive disagreements in STS annotation, existing benchmarks have used averaged human ratings as gold standard. Averaging masks the true distribution of human opinions on examples of low agreement, and prevents models from capturing the semantic vagueness that the individual ratings represent. In this work, we introduce USTS, the first Uncertainty-aware STS dataset with ∼15,000 Chinese sentence pairs and 150,000 labels, to study collective human opinions in STS. Analysis reveals that neither a scalar nor a single Gaussian fits a set of observed judgments adequately. We further show that current STS models cannot capture the variance caused by human disagreement on individual instances, but rather reflect the predictive confidence over the aggregate dataset.
This paper presents the submission of Huawei Translate Services Center (HW-TSC) to the WMT23 general machine translation (MT) shared task, in which we participate in Chinese↔English (zh↔en) language pair. We use Transformer architecture and obtain the best performance via a variant with larger parameter size. We perform fine-grained pre-processing and filtering on the provided large-scale bilingual and monolingual datasets. We mainly use model enhancement strategies, including Regularized Dropout, Bidirectional Training, Data Diversification, Forward Translation, Back Translation, Alternated Training, Curriculum Learning and Transductive Ensemble Learning. Our submissions obtain competitive results in the final evaluation.
Machine Translation Evaluation is critical to Machine Translation research, as the evaluation results reflect the effectiveness of training strategies. As a result, a fair and efficient evaluation method is necessary. Many researchers have raised questions about currently available evaluation metrics from various perspectives, and propose suggestions accordingly. However, to our knowledge, few researchers has analyzed the difficulty level of source sentence and its influence on evaluation results. This paper presents HW-TSC’s submission to the WMT23 MT Test Suites shared task. We propose a systematic approach for construing challenge sets from four aspects: word difficulty, length difficulty, grammar difficulty and model learning difficulty. We open-source two Multifaceted Challenge Sets for Zh→En and En→Zh. We also present results of participants in this year’s General MT shared task on our test sets.
This paper presents the domain adaptation methods adopted by Huawei Translation Service Center (HW-TSC) to train the neural machine translation (NMT) system on the English↔German (en↔de) language pair of the WMT23 biomedical translation task. Our NMT system is built on deep Transformer with larger parameter sizes. Based on the biomedical NMT system trained last year, we leverage Curriculum Learning, Data Diversification, Forward translation, Back translation, and Transductive Ensemble Learning to further improve system performance. Overall, we believe our submission can achieve highly competitive result in the official final evaluation.
This paper introduces HW-TSC’s submission to the WMT23 Discourse-Level Literary Translation shared task. We use standard sentence-level transformer as a baseline, and perform domain adaptation and discourse modeling to enhance discourse-level capabilities. Regarding domain adaptation, we employ Back-Translation, Forward-Translation and Data Diversification. For discourse modeling, we apply strategies such as Multi-resolutional Document-to-Document Translation and TrAining Data Augmentation.
This paper presents the submission of Huawei Translation Service Center (HW-TSC) to the WMT23 metrics shared task, in which we submit two metrics: KG-BERTScore and HWTSC-EE-Metric. Among them, KG-BERTScore is our primary submission for the reference-free metric, which can provide both segment-level and system-level scoring. While HWTSC-EE-Metric is our primary submission for the reference-based metric, which can only provide system-level scoring. Overall, our metrics show relatively high correlations with MQM scores on the metrics tasks of previous years. Especially on system-level scoring tasks, our metrics achieve new state-of-the-art in many language pairs.
We introduce the submissions of the NJUNLP team to the WMT 2023 Quality Estimation (QE) shared task. Our team submitted predictions for the English-German language pair on all two sub-tasks: (i) sentence- and word-level quality prediction; and (ii) fine-grained error span detection. This year, we further explore pseudo data methods for QE based on NJUQE framework (https://github.com/NJUNLP/njuqe). We generate pseudo MQM data using parallel data from the WMT translation task. We pre-train the XLMR large model on pseudo QE data, then fine-tune it on real QE data. At both stages, we jointly learn sentence-level scores and word-level tags. Empirically, we conduct experiments to find the key hyper-parameters that improve the performance. Technically, we propose a simple method that covert the word-level outputs to fine-grained error span results. Overall, our models achieved the best results in English-German for both word-level and fine-grained error span detection sub-tasks by a considerable margin.
Quality estimation (QE) is an essential technique to assess machine translation quality without reference translations. In this paper, we focus on Huawei Translation Services Center’s (HW-TSC’s) submission to the sentence-level QE shared task, named Ensemble-CrossQE. Our system uses CrossQE, the same model architecture as our last year’s submission, which consists of a multilingual base model and a task-specific downstream layer. The input is the concatenation of the source and the translated sentences. To enhance the performance, we finetuned and ensembled multiple base models such as XLM-R, InfoXLM, RemBERT and CometKiwi. Moreover, we introduce a new corruption-based data augmentation method, which generates deletion, substitution and insertion errors in the original translation and uses a reference-based QE model to obtain pseudo scores. Results show that our system achieves impressive performance on sentence-level QE test sets and ranked the first place for three language pairs: English-Hindi, English-Tamil and English-Telegu. In addition, we participated in the error span detection task. The submitted model outperforms the baseline on Chinese-English and Hebrew-English language pairs.
The paper presents the submission by HW-TSC in the WMT 2023 Automatic Post Editing (APE) shared task for the English-Marathi (En-Mr) language pair. Our method encompasses several key steps. First, we pre-train an APE model by utilizing synthetic APE data provided by the official task organizers. Then, we fine-tune the model by employing real APE data. For data augmentation, we incorporate candidate translations obtained from an external Machine Translation (MT) system. Furthermore, we integrate the En-Mr parallel corpus from the Flores-200 dataset into our training data. To address the overfitting issue, we employ R-Drop during the training phase. Given that APE systems tend to exhibit a tendency of ‘over-correction’, we employ a sentence-level Quality Estimation (QE) system to select the final output, deciding between the original translation and the corresponding output generated by the APE model. Our experiments demonstrate that pre-trained APE models are effective when being fine-tuned with the APE corpus of a limited size, and the performance can be further improved with external MT augmentation. Our approach improves the TER and BLEU scores on the development set by -2.42 and +3.76 points, respectively.
Recently, ChatGPT has shown promising results for Machine Translation (MT) in general domains and is becoming a new paradigm for translation. In this paper, we focus on how to apply ChatGPT to domain-specific translation and propose to leverage Multilingual Knowledge Graph (MKG) to help ChatGPT improve the domain entity translation quality. To achieve this, we extract the bilingual entity pairs from MKG for the domain entities that are recognized from source sentences. We then introduce these pairs into translation prompts, instructing ChatGPT to use the correct translations of the domain entities. To evaluate the novel MKG method for ChatGPT, we conduct comparative experiments on three Chinese-English (zh-en) test datasets constructed from three specific domains, of which one domain is from biomedical science, and the other two are from the Information and Communications Technology (ICT) industry — Visible Light Communication (VLC) and wireless domains. Experimental results demonstrate that both the overall translation quality of ChatGPT (+6.21, +3.13 and +11.25 in BLEU scores) and the translation accuracy of domain entities (+43.2%, +30.2% and +37.9% absolute points) are significantly improved with MKG on the three test datasets.
The widespread use of machine translation (MT) has driven the need for effective automatic quality estimation (AQE) methods. How to enhance the interpretability of MT output quality estimation is well worth exploring in the industry. From the perspective of the alignment of named entities (NEs) in the source and translated sentences, we construct a multilingual knowledge graph (KG) consisting of domain-specific NEs, and design a KG-based interpretable quality estimation (QE) system for machine translations (KG-IQES). KG-IQES effectively estimates the translation quality without relying on reference translations. Its effectiveness has been verified in our business scenarios.
Lexically constrained neural machine translation (NMT) draws much industrial attention for its practical usage in specific domains. However, current autoregressive approaches suffer from high latency. In this paper, we focus on non-autoregressive translation (NAT) for this problem for its efficiency advantage. We identify that current constrained NAT models, which are based on iterative editing, do not handle low-frequency constraints well. To this end, we propose a plug-in algorithm for this line of work, i.e., Aligned Constrained Training (ACT), which alleviates this problem by familiarizing the model with the source-side context of the constraints. Experiments on the general and domain datasets show that our model improves over the backbone constrained NAT model in constraint preservation and translation quality, especially for rare constraints.
In this paper we describe our submission to the shared tasks of the 9th Workshop on Asian Translation (WAT 2022) on NICT–SAP under the team name ”HwTscSU”. The tasks involve translation from 5 languages into English and vice-versa in two domains: IT domain and Wikinews domain. The purpose is to determine the feasibility of multilingualism, domain adaptation or document-level knowledge given very little to none clean parallel corpora for training. Our approach for all translation tasks mainly focused on pre-training NMT models on general datasets and fine-tuning them on domain-specific datasets. Due to the small amount of parallel corpora, we collected and cleaned the OPUS dataset including three IT domain corpora, i.e., GNOME, KDE4, and Ubuntu. We then trained Transformer models on the collected dataset and fine-tuned on corresponding dev set. The BLEU scores greatly improved in comparison with other systems. Our submission ranked 1st in all IT-domain tasks and in one out of eight ALT domain tasks.
Aspect-level multimodal sentiment analysis, which aims to identify the sentiment of the target aspect from multimodal data, recently has attracted extensive attention in the community of multimedia and natural language processing. Despite the recent success in textual aspect-based sentiment analysis, existing models mainly focused on utilizing the object-level semantic information in the image but ignore explicitly using the visual emotional cues, especially the facial emotions. How to distill visual emotional cues and align them with the textual content remains a key challenge to solve the problem. In this work, we introduce a face-sensitive image-to-emotional-text translation (FITE) method, which focuses on capturing visual sentiment cues through facial expressions and selectively matching and fusing with the target aspect in textual modality. To the best of our knowledge, we are the first that explicitly utilize the emotional information from images in the multimodal aspect-based sentiment analysis task. Experiment results show that our method achieves state-of-the-art results on the Twitter-2015 and Twitter-2017 datasets. The improvement demonstrates the superiority of our model in capturing aspect-level sentiment in multimodal data with facial expressions.
In this paper we aim to relieve the issue of lexical translation inconsistency for document-level neural machine translation (NMT) by modeling consistency preference for lexical chains, which consist of repeated words in a source-side document and provide a representation of the lexical consistency structure of the document. Specifically, we first propose lexical-consistency attention to capture consistency context among words in the same lexical chains. Then for each lexical chain we define and learn a consistency-tailored latent variable, which will guide the translation of corresponding sentences to enhance lexical translation consistency. Experimental results on Chinese→English and French→English document-level translation tasks show that our approach not only significantly improves translation performance in BLEU, but also substantially alleviates the problem of the lexical translation inconsistency.
Multimodal sentiment analysis has attracted increasing attention and lots of models have been proposed. However, the performance of the state-of-the-art models decreases sharply when they are deployed in the real world. We find that the main reason is that real-world applications can only access the text outputs by the automatic speech recognition (ASR) models, which may be with errors because of the limitation of model capacity. Through further analysis of the ASR outputs, we find that in some cases the sentiment words, the key sentiment elements in the textual modality, are recognized as other words, which makes the sentiment of the text change and hurts the performance of multimodal sentiment analysis models directly. To address this problem, we propose the sentiment word aware multimodal refinement model (SWRM), which can dynamically refine the erroneous sentiment words by leveraging multimodal sentiment clues. Specifically, we first use the sentiment word position detection module to obtain the most possible position of the sentiment word in the text and then utilize the multimodal sentiment word refinement module to dynamically refine the sentiment word embeddings. The refined embeddings are taken as the textual inputs of the multimodal feature fusion module to predict the sentiment labels. We conduct extensive experiments on the real-world datasets including MOSI-Speechbrain, MOSI-IBM, and MOSI-iFlytek and the results demonstrate the effectiveness of our model, which surpasses the current state-of-the-art models on three datasets. Furthermore, our approach can be adapted for other multimodal feature fusion models easily.
Natural Language Inference (NLI) datasets contain examples with highly ambiguous labels due to its subjectivity. Several recent efforts have been made to acknowledge and embrace the existence of ambiguity, and explore how to capture the human disagreement distribution. In contrast with directly learning from gold ambiguity labels, relying on special resource, we argue that the model has naturally captured the human ambiguity distribution as long as it’s calibrated, i.e. the predictive probability can reflect the true correctness likelihood. Our experiments show that when model is well-calibrated, either by label smoothing or temperature scaling, it can obtain competitive performance as prior work, on both divergence scores between predictive probability and the true human opinion distribution, and the accuracy. This reveals the overhead of collecting gold ambiguity labels can be cut, by broadly solving how to calibrate the NLI network.
Machine translation (MT) metrics often experience poor correlations with human assessments. In terms of MT system evaluation, most metrics pay equal attentions to every sample in an evaluation set, while in human evaluation, difficult sentences often make candidate systems distinguishable via notable fluctuations in human scores, especially when systems are competitive. We find that samples with high entropy values, which though usually count less than 5%, tend to play a key role in MT evaluation: when the evaluation set is shrunk to only the high-entropy portion, correlations with human assessments are actually improved. Thus, in this paper, we propose a fast and unsupervised approach to enhance MT metrics using entropy, expanding the dimension of evaluation by introducing sentence-level difficulty. A translation hypothesis with a significantly high entropy value is considered difficult and receives a large weight in aggregation of system-level scores. Experimental results on five sub-tracks in the WMT19 Metrics shared tasks show that our proposed method significantly enhanced the performance of commonly-used MT metrics in terms of system-level correlations with human assessments, even outperforming existing SOTA metrics. In particular, all enhanced metrics exhibit overall stability in correlations with human assessments in circumstances where only competitive MT systems are included, while the corresponding vanilla metrics fail to correlate with human assessments.
Pre-trained speech Transformers have facilitated great success across various speech processing tasks. However, fine-tuning these encoders for downstream tasks require sufficiently large training data to converge or to achieve state-of-the-art. In text domain this has been partly attributed to sub-optimality of the representation space in pre-trained Transformers. In this work, we take a sober look into pre-trained speech encoders and rewire their representation space without requiring any task-specific labels. Our method utilises neutrally synthesised version of audio inputs along with frame masking to construct positive pairs for contrastive self-supervised learning. When used for augmenting the wav2vec 2 encoder, we observe consistent improvement of isotropy in the representation space. Our experiments on 6 speech processing tasks, exhibit a significant convergence speedup during task fine-tuning as well as consistent task improvement, specially in low-resource settings.
Pre-trained speech Transformers in speech translation (ST) have facilitated state-of-the-art (SotA) results; yet, using such encoders is computationally expensive. To improve this, we present a novel Reducer Adaptor block, RedApt, that could be seamlessly integrated within any Transformer-based speech encoding architecture. Integrating the pretrained wav2vec 2 speech encoder with RedAptbrings 41% speedup, 33% memory reduction with 24% fewer FLOPs at inference. To our positive surprise, our ST model with RedApt outperforms the SotA architecture by an average of 0.68 BLEU score on 8 language pairs from Must-C.
Autoregressive (AR) and Non-autoregressive (NAR) models have their own superiority on the performance and latency, combining them into one model may take advantage of both. Current combination frameworks focus more on the integration of multiple decoding paradigms with a unified generative model, e.g. Masked Language Model. However, the generalization can be harmful on the performance due to the gap between training objective and inference. In this paper, we aim to close the gap by preserving the original objective of AR and NAR under a unified framework. Specifically, we propose the Directional Transformer (Diformer) by jointly modelling AR and NAR into three generation directions (left-to-right, right-to-left and straight) with a newly introduced direction variable, which works by controlling the prediction of each token to have specific dependencies under that direction. The unification achieved by direction successfully preserves the original dependency assumption used in AR and NAR, retaining both generalization and performance. Experiments on 4 WMT benchmarks demonstrate that Diformer outperforms current united-modelling works with more than 1.5 BLEU points for both AR and NAR decoding, and is also competitive to the state-of-the-art independent AR and NAR models.
This paper describes the system for the identifying Plausible Clarifications of Implicit and Underspecified Phrases. This task was set up as an English cloze task, in which clarifications are presented as possible fillers and systems have to score how well each filler plausibly fits in a given context. For this shared task, we propose our own solutions, including supervised proaches, unsupervised approaches with pretrained models, and then we use these models to build an ensemble model. Finally we get the 2nd best result in the subtask1 which is a classification task, and the 3rd best result in the subtask2 which is a regression task.
This paper presents the submissions of Huawei Translate Services Center (HW-TSC) to the WMT 2022 General Machine Translation Shared Task. We participate in 6 language pairs, including Zh↔En, Ru↔En, Uk↔En, Hr↔En, Uk↔Cs and Liv↔En. We use Transformer architecture and obtain the best performance via multiple variants with larger parameter sizes. We perform fine-grained pre-processing and filtering on the provided large-scale bilingual and monolingual datasets. For medium and highresource languages, we mainly use data augmentation strategies, including Back Translation, Self Training, Ensemble Knowledge Distillation, Multilingual, etc. For low-resource languages such as Liv, we use pre-trained machine translation models, and then continue training with Regularization Dropout (R-Drop). The previous mentioned data augmentation methods are also used. Our submissions obtain competitive results in the final evaluation.
Contextual word embeddings extracted from pre-trained models have become the basis for many downstream NLP tasks, including machine translation automatic evaluations. Metrics that leverage embeddings claim better capture of synonyms and changes in word orders, and thus better correlation with human ratings than surface-form matching metrics (e.g. BLEU). However, few studies have been done to examine robustness of these metrics. This report uses a challenge set to uncover the brittleness of reference-based and reference-free metrics. Our challenge set1 aims at examining metrics’ capability to correlate synonyms in different areas and to discern catastrophic errors at both word- and sentence-levels. The results show that although embedding-based metrics perform relatively well on discerning sentence-level negation/affirmation errors, their performances on relating synonyms are poor. In addition, we find that some metrics are susceptible to text styles so their generalizability compromised.
In this paper, we present the contribution of HW-TSC to WMT 2022 Metrics Shared Task. We propose one reference-based metric, HWTSC-EE-BERTScore*, and four referencefree metrics including HWTSC-Teacher-Sim, HWTSC-TLM, KG-BERTScore and CROSSQE. Among these metrics, HWTSC-Teacher-Sim and CROSS-QE are supervised, whereas HWTSC-EE-BERTScore*, HWTSC-TLM and KG-BERTScore are unsupervised. We use these metrics in the segment-level and systemlevel tracks. Overall, our systems achieve strong results for all language pairs on previous test sets and a new state-of-the-art in many sys-level case sets.
This paper presents submissions of the NJUNLP team in WMT 2022Quality Estimation shared task 1, where the goal is to predict the sentence-level and word-level quality for target machine translations. Our system explores pseudo data and multi-task learning. We propose several novel methods to generate pseudo data for different annotations using the conditional masked language model and the neural machine translation model. The proposed methods control the decoding process to generate more real pseudo translations. We pre-train the XLMR-large model with pseudo data and then fine-tune this model with real data both in the way of multi-task learning. We jointly learn sentence-level scores (with regression and rank tasks) and word-level tags (with a sequence tagging task). Our system obtains competitive results on different language pairs and ranks first place on both sentence- and word-level sub-tasks of the English-German language pair.
Quality estimation (QE) is a crucial method to investigate automatic methods for estimating the quality of machine translation results without reference translations. This paper presents Huawei Translation Services Center’s (HW-TSC’s) work called CrossQE in WMT 2022 QE shared tasks 1 and 2, namely sentence- and word- level quality prediction and explainable QE.CrossQE employes the framework of predictor-estimator for task 1, concretely with a pre-trained cross-lingual XLM-RoBERTa large as predictor and task-specific classifier or regressor as estimator. An extensive set of experimental results show that after adding bottleneck adapter layer, mean teacher loss, masked language modeling task loss and MC dropout methods in CrossQE, the performance has improved to a certain extent. For task 2, CrossQE calculated the cosine similarity between each word feature in the target and each word feature in the source by task 1 sentence-level QE system’s predictor, and used the inverse value of maximum similarity between each word in the target and the source as the word translation error risk value. Moreover, CrossQE has outstanding performance on QE test sets of WMT 2022.
This paper presents the submission of Huawei Translation Services Center (HW-TSC) to WMT 2022 Efficiency Shared Task. For this year’s task, we still apply sentence-level distillation strategy to train small models with different configurations. Then, we integrate the average attention mechanism into the lightweight RNN model to pursue more efficient decoding. We tried adding a retrain step to our 8-bit and 4-bit models to achieve a balance between model size and quality. We still use Huawei Noah’s Bolt for INT8 inference and 4-bit storage. Coupled with Bolt’s support for batch inference and multi-core parallel computing, we finally submit models with different configurations to the CPU latency and throughput tracks to explore the Pareto frontiers.
This paper describes the translation systems trained by Huawei translation services center (HW-TSC) for the WMT22 biomedical translation task in five language pairs: English↔German (en↔de), English↔French (en↔fr), English↔Chinese (en↔zh), English↔Russian (en↔ru) and Spanish→English (es→en). Our primary systems are built on deep Transformer with a large filter size. We also utilize R-Drop, data diversification, forward translation, back translation, data selection, finetuning and ensemble to improve the system performance. According to the official evaluation results in OCELoT or CodaLab, our unconstrained systems in en→de, de→en, en→fr, fr→en, en→zh and es→en (clinical terminology sub-track) get the highest BLEU scores among all submissions for the WMT22 biomedical translation task.
This paper describes the submissions of Huawei Translation Services Center (HW-TSC) to WMT22 chat translation shared task on English-Germany (en-de) bidirection with results of zore-shot and few-shot tracks. We use the deep transformer architecture with a lager parameter size. Our submissions to the WMT21 News Translation task are used as the baselines. We adopt strategies such as back translation, forward translation, domain transfer, data selection, and noisy forward translation in task, and achieve competitive results on the development set. We also test the effectiveness of document translation on chat tasks. Due to the lack of chat data, the results on the development set show that it is not as effective as sentence-level translation models.
This paper describes the submissions of Huawei translation services center (HW-TSC) to the WMT22 Very Low Resource Supervised MT task. We participate in all 6 supervised tracks including all combinations between Upper/Lower Sorbian (Hsb/Dsb) and German (De). Our systems are build on deep Transformer with a large filter size. We use multilingual transfer with German-Czech (De-Cs) and German-Polish (De-Pl) parallel data. We also utilize regularized dropout (R-Drop), back translation, fine-tuning and ensemble to improve the system performance. According to the official evaluation results on OCELoT, our supervised systems of all 6 language directions get the highest BLEU scores among all submissions. Our pre-trained multilingual model for unsupervised De2Dsb and Dsb2De translation also gain highest BLEU.
This paper presents the submissions of Huawei Translation Services Center (HW-TSC) to WMT 2022 Word-Level AutoCompletion Task. We propose an end-to-end autoregressive model with bi-context based on Transformer to solve current task. The model uses a mixture of subword and character encoding units to realize the joint encoding of human input, the context of the target side and the decoded sequence, which ensures full utilization of information. We uses one model to solve four types of data structures in the task. During training, we try using a machine translation model as the pre-trained model and fine-tune it for the task. We also add BERT-style MLM data at the fine-tuning stage to improve model performance. We participate in zh→en, en→de, and de→en directions and win the first place in all the three tracks. Particularly, we outperform the second place by more than 5% in terms of accuracy on the zh→en and en→de tracks. The result is buttressed by human evaluations as well, demonstrating the effectiveness of our model.
This paper describes the HW-TSC’s designation of the Offline Speech Translation System submitted for IWSLT 2022 Evaluation. We explored both cascade and end-to-end system on three language tracks (en-de, en-zh and en-ja), and we chose the cascade one as our primary submission. For the automatic speech recognition (ASR) model of cascade system, there are three ASR models including Conformer, S2T-Transformer and U2 trained on the mixture of five datasets. During inference, transcripts are generated with the help of domain controlled generation strategy. Context-aware reranking and ensemble based anti-interference strategy are proposed to produce better ASR outputs. For machine translation part, we pretrained three translation models on WMT21 dataset and fine-tuned them on in-domain corpora. Our cascade system shows competitive performance than the known offline systems in the industry and academia.
This paper presents our work in the participation of IWSLT 2022 simultaneous speech translation evaluation. For the track of text-to-text (T2T), we participate in three language pairs and build wait-k based simultaneous MT (SimulMT) model for the task. The model was pretrained on WMT21 news corpora, and was further improved with in-domain fine-tuning and self-training. For the speech-to-text (S2T) track, we designed both cascade and end-to-end form in three language pairs. The cascade system is composed of a chunking-based streaming ASR model and the SimulMT model used in the T2T track. The end-to-end system is a simultaneous speech translation (SimulST) model based on wait-k strategy, which is directly trained on a synthetic corpus produced by translating all texts of ASR corpora into specific target language with an offline MT model. It also contains a heuristic sentence breaking strategy, preventing it from finishing the translation before the the end of the speech. We evaluate our systems on the MUST-C tst-COMMON dataset and show that the end-to-end system is competitive to the cascade one. Meanwhile, we also demonstrate that the SimulMT model can be efficiently optimized by these approaches, resulting in the improvements of 1-2 BLEU points.
The paper presents the HW-TSC’s pipeline and results of Offline Speech to Speech Translation for IWSLT 2022. We design a cascade system consisted of an ASR model, machine translation model and TTS model to convert the speech from one language into another language(en-de). For the ASR part, we find that better performance can be obtained by ensembling multiple heterogeneous ASR models and performing reranking on beam candidates. And we find that the combination of context-aware reranking strategy and MT model fine-tuned on the in-domain dataset is helpful to improve the performance. Because it can mitigate the problem that the inconsistency in transcripts caused by the lack of context. Finally, we use VITS model provided officially to reproduce audio files from the translation hypothesis.
This paper presents our submissions to the IWSLT 2022 Isometric Spoken Language Translation task. We participate in all three language pairs (English-German, English-French, English-Spanish) under the constrained setting, and submit an English-German result under the unconstrained setting. We use the standard Transformer model as the baseline and obtain the best performance via one of its variants that shares the decoder input and output embedding. We perform detailed pre-processing and filtering on the provided bilingual data. Several strategies are used to train our models, such as Multilingual Translation, Back Translation, Forward Translation, R-Drop, Average Checkpoint, and Ensemble. We investigate three methods for biasing the output length: i) conditioning the output to a given target-source length-ratio class; ii) enriching the transformer positional embedding with length information and iii) length control decoding for non-autoregressive translation etc. Our submissions achieve 30.7, 41.6 and 36.7 BLEU respectively on the tst-COMMON test sets for English-German, English-French, English-Spanish tasks and 100% comply with the length requirements.
Based on large-scale pretrained networks and the liability to be easily overfitting with limited labelled training data of multimodal translation (MMT) is a critical issue in MMT. To this end and we propose a transfer learning solution. Specifically and 1) A vanilla Transformer is pre-trained on massive bilingual text-only corpus to obtain prior knowledge; 2) A multimodal Transformer named VLTransformer is proposed with several components incorporated visual contexts; and 3) The parameters of VLTransformer are initialized with the pre-trained vanilla Transformer and then being fine-tuned on MMT tasks with a newly proposed method named cross-modal masking which forces the model to learn from both modalities. We evaluated on the Multi30k en-de and en-fr dataset and improving up to 8% BLEU score compared with the SOTA performance. The experimental result demonstrates that performing transfer learning with monomodal pre-trained NMT model on multimodal NMT tasks can obtain considerable boosts.
Length prediction is a special task in a series of NAT models where target length has to be determined before generation. However, the performance of length prediction and its influence on translation quality has seldom been discussed. In this paper, we present comprehensive analyses on length prediction task of NAT, aiming to find the factors that influence performance, as well as how it associates with translation quality. We mainly perform experiments based on Conditional Masked Language Model (CMLM) (Ghazvininejad et al., 2019), a representative NAT model, and evaluate it on two language pairs, En-De and En-Ro. We draw two conclusions: 1) The performance of length prediction is mainly influenced by properties of language pairs such as alignment pattern, word order or intrinsic length ratio, and is also affected by the usage of knowledge distilled data. 2) There is a positive correlation between the performance of the length prediction and the BLEU score.
Mask-predict CMLM (Ghazvininejad et al.,2019) has achieved stunning performance among non-autoregressive NMT models, but we find that the mechanism of predicting all of the target words only depending on the hidden state of [MASK] is not effective and efficient in initial iterations of refinement, resulting in ungrammatical repetitions and slow convergence. In this work, we mitigate this problem by combining copied source with embeddings of [MASK] in decoder. Notably. it’s not a straightforward copying that is shown to be useless, but a novel heuristic hybrid strategy — fence-mask. Experimental results show that it gains consistent boosts on both WMT14 En<->De and WMT16 En<->Ro corpus by 0.5 BLEU on average, and 1 BLEU for less-informative short sentences. This reveals that incorporating additional information by proper strategies is beneficial to improve CMLM, particularly translation quality of short texts and speeding up early-stage convergence.
This paper presents the submission of Huawei Translate Services Center (HW-TSC) to the WMT 2021 News Translation Shared Task. We participate in 7 language pairs, including Zh/En, De/En, Ja/En, Ha/En, Is/En, Hi/Bn, and Xh/Zu in both directions under the constrained condition. We use Transformer architecture and obtain the best performance via multiple variants with larger parameter sizes. We perform detailed pre-processing and filtering on the provided large-scale bilingual and monolingual datasets. Several commonly used strategies are used to train our models, such as Back Translation, Forward Translation, Multilingual Translation, Ensemble Knowledge Distillation, etc. Our submission obtains competitive results in the final evaluation.
This paper presents the submission of Huawei Translation Service Center (HW-TSC) to WMT 2021 Triangular MT Shared Task. We participate in the Russian-to-Chinese task under the constrained condition. We use Transformer architecture and obtain the best performance via a variant with larger parameter sizes. We perform detailed data pre-processing and filtering on the provided large-scale bilingual data. Several strategies are used to train our models, such as Multilingual Translation, Back Translation, Forward Translation, Data Denoising, Average Checkpoint, Ensemble, Fine-tuning, etc. Our system obtains 32.5 BLEU on the dev set and 27.7 BLEU on the test set, the highest score among all submissions.
This paper presents the submission of Huawei Translation Services Center (HW-TSC) to the WMT 2021 Large-Scale Multilingual Translation Task. We participate in Samll Track #2, including 6 languages: Javanese (Jv), Indonesian (Id), Malay (Ms), Tagalog (Tl), Tamil (Ta) and English (En) with 30 directions under the constrained condition. We use Transformer architecture and obtain the best performance via multiple variants with larger parameter sizes. We train a single multilingual model to translate all the 30 directions. We perform detailed pre-processing and filtering on the provided large-scale bilingual and monolingual datasets. Several commonly used strategies are used to train our models, such as Back Translation, Forward Translation, Ensemble Knowledge Distillation, Adapter Fine-tuning. Our model obtains competitive results in the end.
This paper presents the submission of Huawei Translation Services Center (HW-TSC) to WMT 2021 Efficiency Shared Task. We explore the sentence-level teacher-student distillation technique and train several small-size models that find a balance between efficiency and quality. Our models feature deep encoder, shallow decoder and light-weight RNN with SSRU layer. We use Huawei Noah’s Bolt, an efficient and light-weight library for on-device inference. Leveraging INT8 quantization, self-defined General Matrix Multiplication (GEMM) operator, shortlist, greedy search and caching, we submit four small-size and efficient translation models with high translation quality for the one CPU core latency track.
This paper describes the submission of Huawei Translation Service Center (HW-TSC) to WMT21 biomedical translation task in two language pairs: Chinese↔English and German↔English (Our registered team name is HuaweiTSC). Technical details are introduced in this paper, including model framework, data pre-processing method and model enhancement strategies. In addition, using the wmt20 OK-aligned biomedical test set, we compare and analyze system performances under different strategies. On WMT21 biomedical translation task, Our systems in English→Chinese and English→German directions get the highest BLEU scores among all submissions according to the official evaluation results.
This paper presents our work in WMT 2021 Quality Estimation (QE) Shared Task. We participated in all of the three sub-tasks, including Sentence-Level Direct Assessment (DA) task, Word and Sentence-Level Post-editing Effort task and Critical Error Detection task, in all language pairs. Our systems employ the framework of Predictor-Estimator, concretely with a pre-trained XLM-Roberta as Predictor and task-specific classifier or regressor as Estimator. For all tasks, we improve our systems by incorporating post-edit sentence or additional high-quality translation sentence in the way of multitask learning or encoding it with predictors directly. Moreover, in zero-shot setting, our data augmentation strategy based on Monte-Carlo Dropout brings up significant improvement on DA sub-task. Notably, our submissions achieve remarkable results over all tasks.
The structural information of Knowledge Bases (KBs) has proven effective to Question Answering (QA). Previous studies rely on deep graph neural networks (GNNs) to capture rich structural information, which may not model node relations in particularly long distance due to oversmoothing issue. To address this challenge, we propose a novel framework GlobalGraph, which models long-distance node relations from two views: 1) Node type similarity: GlobalGraph assigns each node a global type label and models long-distance node relations through the global type label similarity; 2) Correlation between nodes and questions: we learn similarity scores between nodes and the question, and model long-distance node relations through the sum score of two nodes. We conduct extensive experiments on two widely used multi-hop KBQA datasets to prove the effectiveness of our method.
This paper presents our work in the WMT 2020 News Translation Shared Task. We participate in 3 language pairs including Zh/En, Km/En, and Ps/En and in both directions under the constrained condition. We use the standard Transformer-Big model as the baseline and obtain the best performance via two variants with larger parameter sizes. We perform detailed pre-processing and filtering on the provided large-scale bilingual and monolingual dataset. Several commonly used strategies are used to train our models such as Back Translation, Ensemble Knowledge Distillation, etc. We also conduct experiment with similar language augmentation, which lead to positive results, although not used in our submission. Our submission obtains remarkable results in the final evaluation.
The paper presents the submission by HW-TSC in the WMT 2020 Automatic Post Editing Shared Task. We participate in the English-German and English-Chinese language pairs. Our system is built based on the Transformer pre-trained on WMT 2019 and WMT 2020 News Translation corpora, and fine-tuned on the APE corpus. Bottleneck Adapter Layers are integrated into the model to prevent over-fitting. We further collect external translations as the augmented MT candidates to improve the performance. The experiment demonstrates that pre-trained NMT models are effective when fine-tuning with the APE corpus of a limited size, and the performance can be further improved with external MT augmentation. Our system achieves competitive results on both directions in the final evaluation.
This paper describes Huawei’s submissions to the WMT20 biomedical translation shared task. Apart from experimenting with finetuning on domain-specific bitexts, we explore effects of in-domain dictionaries on enhancing cross-domain neural machine translation performance. We utilize a transfer learning strategy through pre-trained machine translation models and extensive scope of engineering endeavors. Four of our ten submissions achieve state-of-the-art performance according to the official automatic evaluation results, namely translation directions on English<->French, English->German and English->Italian.
This paper presents our work in the WMT 2020 Word and Sentence-Level Post-Editing Quality Estimation (QE) Shared Task. Our system follows standard Predictor-Estimator architecture, with a pre-trained Transformer as the Predictor, and specific classifiers and regressors as Estimators. We integrate Bottleneck Adapter Layers in the Predictor to improve the transfer learning efficiency and prevent from over-fitting. At the same time, we jointly train the word- and sentence-level tasks with a unified model with multitask learning. Pseudo-PE assisted QE (PEAQE) is proposed, resulting in significant improvements on the performance. Our submissions achieve competitive result in word/sentence-level sub-tasks for both of En-De/Zh language pairs.
The Predictor-Estimator framework for quality estimation (QE) is commonly used for its strong performance. Where the predictor and estimator works on feature extraction and quality evaluation, respectively. However, training the predictor from scratch is computationally expensive. In this paper, we propose an efficient transfer learning framework to transfer knowledge from NMT dataset into QE models. A Predictor-Estimator alike model named BAL-QE is also proposed, aiming to extract high quality features with pre-trained NMT model, and make classification with a fine-tuned Bottleneck Adapter Layer (BAL). The experiment shows that BAL-QE achieves 97% of the SOTA performance in WMT19 En-De and En-Ru QE tasks by only training 3% of parameters within 4 hours on 4 Titan XP GPUs. Compared with the commonly used NuQE baseline, BAL-QE achieves 47% (En-Ru) and 75% (En-De) of performance promotions.
We propose a unified multilingual model for humor detection which can be trained under a transfer learning framework. 1) The model is built based on pre-trained multilingual BERT, thereby is able to make predictions on Chinese, Russian and Spanish corpora. 2) We step out from single sentence classification and propose sequence-pair prediction which considers the inter-sentence relationship. 3) We propose the Sentence Discrepancy Prediction (SDP) loss, aiming to measure the semantic discrepancy of the sequence-pair, which often appears in the setup and punchline of a joke. Our method achieves two SoTA and a second-place on three humor detection corpora in three languages (Russian, Spanish and Chinese), and also improves F1-score by 4%-6%, which demonstrates the effectiveness of it in humor detection tasks.
This paper describes our work in the WAT 2020 Indic Multilingual Translation Task. We participated in all 7 language pairs (En<->Bn/Hi/Gu/Ml/Mr/Ta/Te) in both directions under the constrained condition—using only the officially provided data. Using transformer as a baseline, our Multi->En and En->Multi translation systems achieve the best performances. Detailed data filtering and data domain selection are the keys to performance enhancement in our experiment, with an average improvement of 2.6 BLEU scores for each language pair in the En->Multi system and an average improvement of 4.6 BLEU scores regarding the Multi->En. In addition, we employed language independent adapter to further improve the system performances. Our submission obtains competitive results in the final evaluation.
The paper presents details of our system in the IWSLT Video Speech Translation evaluation. The system works in a cascade form, which contains three modules: 1) A proprietary ASR system. 2) A disfluency correction system aims to remove interregnums or other disfluent expressions with a fine-tuned BERT and a series of rule-based algorithms. 3) An NMT System based on the Transformer and trained with massive publicly available corpus.
With huge amount of information generated every day on the web, fact checking is an important and challenging task which can help people identify the authenticity of most claims as well as providing evidences selected from knowledge source like Wikipedia. Here we decompose this problem into two parts: an entity linking task (retrieving relative Wikipedia pages) and recognizing textual entailment between the claim and selected pages. In this paper, we present an end-to-end multi-task learning with bi-direction attention (EMBA) model to classify the claim as “supports”, “refutes” or “not enough info” with respect to the pages retrieved and detect sentences as evidence at the same time. We conduct experiments on the FEVER (Fact Extraction and VERification) paper test dataset and shared task test dataset, a new public dataset for verification against textual sources. Experimental results show that our method achieves comparable performance compared with the baseline system.