| 节点id | 节点类型 | 节点颜色 | 中文概括 | 知识单元 | |:——————————————–|:————–|:———–|:———————|:———————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————| | C90-3044&J93-1006 | 研究方法/流程 | Purple | 提出统计方法 | The method proposed there requires a database to be maintained of the syntactic structures of sentences together with the structures of the corresponding translations.This paper proposes a method to solve this problem. | | J93-1006&1995.tmi-1.28 | 研究方法/流程 | Purple | 句子对齐算法 | Algorithms for subsentential alignment have been developed as well at granularities of the character a parallel bilingual corpus may be aligned to the sentence level with reasonable accuracyThe algorithm appears to converge to the correct sentence alignment in only a few iterations. partial alignment of the word level to induce a maximum likelihood alignment of the sentence level | | 1995.tmi-1.28&P96-1021 | 研究方法/流程 | Purple | 反转型录语法(ITG) | inversion transduction grammars or ITGsmaking use of an inversion transduction grammar (ITG) formalism that we recently developed for bilingual language modeling | | P96-1021&C98-2225 | 模型 | Red | 多项式时间算法 | introduced a polynomial-time algorithmA Polynomial-Time Algorithm for Statistical Machine Translation | | C98-2225&P01-1050 | 模型 | Red | 随机语法通道 | noisy channel frameworkMachine Translation with a Stochastic Grammatical Channel | | P01-1050&000223079900018 | 模型 | Red | 利用翻译记忆库 | decoder described in starts from a gloss that uses the translations in the translation memory phrases in the translation memory were automatically extracted from the Viterbi alignments produced by Giza tries to improve on the gloss translation by modifying it incrementally used a statistical translation memory of phrases in conjunction with a statistical translation model uses them in conjunction with a word-for-word statistical translation system extracts phrase translations from automatically aligned corporatranslation memory and a statistical-based translation model set of algorithms that enable us to translate natural language sentences by exploiting both a translation memory and a statistical-based translation model automatically derived translation memory can be used within a statistical framework | | 000223079900018&N03-1017 | 模型 | Red | 联合短语模型 | joint probabilities estimated by proposed a translation model that assumes that lexical correspondences can be established not only at the word level, but at the phrase level as well joint phrase model proposed by introduced a joint-probability model for phrase translation原标题内容 | | N03-1017&N04-1033 | 模型 | Red | 评估短语模型 | or the maximum phrase length the phrase extraction method the underlying word alignment model various aspects of phrase-based systems are comparedWe propose a new phrase-based translation model and decoding algorithm that enables us to evaluate and compare several, previously proposed phrase-based translation models heuristic learning of phrase translations from word-based alignments learning phrases longer than three words | | N04-1033&P05-1033 | 模型 | Red | 单调搜索算法 | simple distortion model that reorders phrases independently of their content, or not at allWe describe a highly efficient monotone search algorithm with a complexity linear in the input sentence length. | | P05-1033&000274500200121 | 模型 | Red | 同步树替换语法 | the rules discussed in this paper are equivalent to productions of synchronous tree substitution grammarsA Hierarchical Phrase-Based Model for Statistical Machine Translation | | 000274500200121&P07-1089 | 模型 | Red | 句法树到字符串 | linguistically syntax-based translation models induce tree-to-string translation rules from parallel texts with linguistic annotationsScalable Inference and Training of Context-Rich Syntactic Translation Models | | P07-1089&P08-1064 | 模型 | Red | 捕获非句法短语 | non-syntactic phrase modelingcapturing nonsyntactic phrase pairs by describing the correspondence between multiple parse trees and one string. | | P08-1064&D08-1024 | 研究方法/流程 | Purple | 软化树约束 | soften the constraints of the input treeautomatically learns aligned tree sequence pairs with mapping probabilities from word-aligned biparsed parallel texts supports multi-level structure reordering of tree typology with larger span | | D08-1024&N09-1025 | 模型 | Red | MIRA调优算法 | MIRA is preferable to MERTwe can obtain results using MIRA that match or surpass MERT in terms of both translation quality and computational cost | | N09-1025&D09-1008 | 研究方法/流程 | Purple | 大规模新特征 | small fraction of the bi-lingual features available11,001 New Features for Statistical Machine Translation | | D09-1008&C10-1050 | 研究方法/流程 | Purple | 集成依赖结构 | proposed a way to integrate dependency structure into target and source side string on hierarchical phrase rulesstate-of-the-art hierarchical MT system source dependency LM scores | | C10-1050&D11-1079 | 模型 | Red | 基于词的重排序 | take a word-based reordering approach for HPBMTHierarchical Phrase-based Machine Translation with Word-based Reordering Model | | D11-1079&D13-1053 | 模型 | Red | 软依赖约束重排序 | improving the reordering of hierarchical phrase-based translation systems external reordering modelsWe derive soft constraints from the source dependency parsing to directly address the reordering problem for the hierarchical phrasebased model Soft Dependency Constraints for Reordering in Hierarchical Phrase-Based Translation | | D13-1053&P14-1129 | 研究方法/流程 | Purple | 分解源句法 | Factored source syntaxFactored Soft Source Syntactic Constraints for Hierarchical Machine Translation | | P14-1129&Q15-1013 | 模型 | Red | 集成与源上下文 | ensembles of networks outperform a single network source-context as inaugments the NNLM with a source context window | | Q15-1013&D15-1248 | 模型 | Red | 关系依赖语言模型 | dependency language model (RDLM) relational dependency language model We propose a language model for dependency structures that is relational rather than configurational as a feature function in string-to-tree SMT from English to German and Russian | | D15-1248&000493806800162 | 模型 | Red | 联合形态句法模型 | syntaxbased SMT resultsimprovements in translation quality of 1.4-1.8 BLEU We propose to model syntactic and morphological structure jointly in a dependency translation model A Joint Dependency Model of Morphological and Syntactic Structure for Statistical Machine Translation | | 000493806800162&000493806800100 | 研究方法/流程 | Purple | 子词单元处理稀疏词 | this problemNeural Machine Translation of Rare Words with Subword Units | | 000493806800100&000493805000058 | 模型 | Red | 混合词-字符模型 | combination of word and charactersAchieving Open Vocabulary Neural Machine Translation with Hybrid Word-Character Models | | 000493805000058&D16-1026 | 模型 | Red | 字符级NMT | sequences of charactersCharacter-based Neural Machine Translation | | D16-1026&Q17-1026 | 研究方法/流程 | Purple | 零资源翻译 | each minibatch is balanced the proportion of each language pair in a single minibatch corresponds to that of the full corpusZero-Resource Translation with Multi-Lingual Neural Machine Translation | | Q17-1026&000493992300092 | 模型 | Red | 全字符级NMT | character level character or sub-word levelsFully Character-Level Neural Machine Translation without Explicit Segmentation | | 000493992300092&000485488905004 | 模型 | Red | 分层级句法翻译 | syntax informationthe target sentence is translated hierarchically from chunks to words, with information in different granularities being leveraged | | 000485488905004&000386658300285 | 数据集 | Green | 多领域平行语料 | discourse parallel corpus 500 thousand Portuguese-English sentence pairs in various domains such as news, laws, microblog etc. 2 million Chinese-EnglishExperimental results on both Chinese English | | 000386658300285&W11-2132 | 研究方法/流程 | Purple | SMT领域自适应 | approachDomain Adaptation for Statistical Machine Translation | | W11-2132&E12-1014 | 研究方法/流程 | Purple | 利用单语数据 | use an existing SMT system to discover parallel sentences within independent monolingual texts, and use them to re-train and enhance the systemInvestigations on Translation Model Adaptation Using Monolingual Data | | E12-1014&P13-1036 | 研究方法/流程 | Purple | 单语诱导短语表 | uses monolingual corpora to induce phrase tablesToward Statistical Machine Translation without Parallel Corpora We extend existing research on bilingual lexicon induction to estimate both lexical and phrasal translation probabilities for MT-scale phrasetables. We estimate the parameters of a phrasebased statistical machine translation system from monolingual corpora instead of a bilingual parallel corpus. | | P13-1036&D14-1061 | 模型 | Red | 贝叶斯哈希采样 | Bayesian inference make decipherment scalablewe propose a new Bayesian inference method Scalable Decipherment for Machine Translation via Hash Sampling | | D14-1061&000493806800185 | 研究方法/流程 | Purple | 联合对齐与破译 | decipherment problemBeyond Parallel Data: Joint Word Alignment and Decipherment Improves Machine Translation | | 000493806800185&D16-1160 | 研究方法/流程 | Purple | 半监督学习 | propose a similar semi-supervised framework similar semi-supervised reconstruction methodSemi-Supervised Learning for Neural Machine Translation | | D16-1160&D16-1249 | 研究方法/流程 | Purple | 利用源端单语数据 | achieve better results monolingual corporaExploiting Source-side Monolingual Data in Neural Machine Translation | | D16-1249&D16-1096 | 研究方法/流程 | Purple | 监督注意力机制 | supervised alignmentsSupervised Attentions for Neural Machine Translation | | D16-1096&000493984800177 | 模型 | Red | 覆盖率模型 | coverage modelCoverage Embedding Models for Neural Machine Translation | | 000493984800177&D17-1012 | 模型 | Red | NMT引入句法 | use of syntactic structures in NMT modelsImproved Neural Machine Translation with a Syntax-Aware Encoder and Decoder | | D17-1012&D17-1209 | 模型 | Red | 隐式句法图解析 | incorporate syntactic information in NMT models in a relatively indirect way (e.g., multi-task learninglearns a latent graph parser as part of the encoder of an attention-based neural machine translation model jointly learns translation and source-side latent graph representations of sentences | | D17-1209&Q18-1017 | 研究方法/流程 | Purple | 图卷积编码器 | Providing explicit linguistic informationGraph Convolutional Encoders for Syntax-aware Neural Machine Translation | | Q18-1017&C18-1263 | 研究方法/流程 | Purple | 调度式多任务学习 | multi-task learning frameworksScheduled Multi-Task Learning: From Syntax to Translation | | C18-1263&000900116903092 | 模型 | Red | 任务特定注意力 | sharing all parameters but the attention mechanismallowing for language-specific specialization of the attention model to a particular language-pair or task Multilingual Neural Machine Translation with Task-Specific Attention compared to a model in which all parameters are shared | | 000900116903092&000736531900047 | 模型 | Red | 大规模多对多模型 | multilingual translation jointly train one translation model that translates multiple language directions at the same time shares representations to improve the translation performance on low-resource languagesmassively multilingual many-to-many models are effective in low resource settings Massively Multilingual Neural Machine Translation training a single model that supports translation from multiple source languages into multiple target languages | | 000736531900047&001181866502027 | 模型 | Red | mBART降噪预训练 | mBART fine-tunes with traditional on-the-fly backtranslation pre-trains on a variety of language configurationsMultilingual Denoising Pre-training for Neural Machine Translation | | 001181866502027&000570978201083 | 研究方法/流程 | Purple | 统一概率框架 | generative modelingour approach results in higher BLEU scores over state-of-the-art unsupervised models We present a probabilistic framework for multilingual neural machine translation that encompasses supervised and unsupervised setups, focusing on unsupervised translation | | 000570978201083&000663162000001 | 模型 | Red | 在线回译 | online backtranslation to improve the performance of non-English pairs multilingual translation models factorize computation when translating to many languages and share information between similar languages English-Centric multilingual model from on the OPUS100 corpus Using backtranslation thus requires the ability to translate in both directions Mining data for each and every language pair is prohibitive - previous work circumvents this issue by focusing only on the 99 pairs that go through English train on 100 directions increasing model capacityExperiments on OPUS-100 (a novel multilingual dataset with 100 languages) Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation stronger modeling capacity OPUS-100 (a novel multilingual dataset with 100 languages) propose random online backtranslation | | 000663162000001&000698663100025 | 模型 | Red | 多对多翻译模型 | multilingual NMT, which supports translation from multiple source languages into multiple target languages with a single modeltraining a single model able to translate between any pair of languages create a true Many-to-Many multilingual translation model that can translate directly between any pair of 100 languages | | 000698663100025&000860727001027 | 模型 | Red | 语言特定子网络(LaSS) | update part of the MT parametersLearning Language Specific Sub-network (LaSS) for each language pair to counter parameter interference | | 000860727001027&000847347700003 | 研究方法/流程 | Purple | 检索增强翻译(KSTER) | semi-parametric models retrieve similar examples from this datastorepropose to learn Kernel-Smoothed Translation with Example Retrieval (KSTER) non-parametric approaches that retrieve similar examples from a database to guide the translation process | | 000847347700003&2023.findings-acl.564 | 研究方法/流程 | Purple | 推理时域外数据 | utilizes out-of-domain bitext during inferenceEfficient Machine Translation Domain Adaptation | | 2023.findings-acl.564&001371932502043 | 研究方法/流程 | Purple | 上下文示例选择 | the quality of the few-shot demonstration impacts the quality of the output.In-context Examples Selection for Machine Translation | | 001371932502043&001156229800010 | 模型 | Red | LLM翻译能力评估 | LLMs can produce fluent and adequate translations, especially for high-resource English-centric language pairs, that are competitive with those of dedicated supervised translation models translation errors, even severely critical ones, obtained via prompting a LLM are different from those produced by traditional machine translation modelsa transformer decoder-only model trained solely with self-supervised learning is able to match specialized supervised state-of-the-art models | | 001156229800010&2023.americasnlp-1.17 | 模型 | Red | LLM幻觉分析 | It remains unclear how well LLMs perform when translating into extremely low-resource languages hallucinations are qualitatively different from conventional translation modelsLLMs produce qualitatively different hallucinations to those of NMT models models struggle with hallucinations primarily in low-resource directions | | 2023.americasnlp-1.17&2023.americasnlp-1.23 | 模型 | Red | kNN-MT | describes systems which make use of kNN and an external data store during decoding submitted four different models for all language pairscombination of fine-tuned M2M100 with $k$NN-MT submitted a total of four systems including GPT-4, a bilingual model, fine-tuned M2M100, and a combination of fine-tuned M2M100 with $k$NN-MT | | 2023.americasnlp-1.23&2023.americasnlp-1.21 | 研究方法/流程 | Purple | AmericasNLP评测 | The 2023 AmericasNLP Shared TaskFindings of the AmericasNLP 2023 Shared Task on Machine Translation into Indigenous Languages | | 2023.americasnlp-1.21&2024.findings-acl.670 | 研究方法/流程 | Purple | 单语数据回译 | back-translations produced from monolingual data constitutions, handbooks, news articlesavailable data from various sources such as constitutions, handbooks and news articles |
Table 2 基于混合权重的主路径
| 节点id | 节点类型 | 节点颜色 | 中文概括 | 知识单元 | |:——————————————————|:——————|:———–|:———————-|:——————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————| | J90-2002&J93-1006 | 研究方法/流程 | Purple | 统计翻译方法 | takes a large corpus of text with aligned translations statistical approach to translationA Statistical Approach to Machine Translation | | C90-3044&J93-1006 | 研究方法/流程 | Purple | 提出统计方法 | The method proposed there requires a database to be maintained of the syntactic structures of sentences together with the structures of the corresponding translations.This paper proposes a method to solve this problem. | | J93-1006&J93-2003 | 研究方法/流程 | Purple | 句子对齐 | automatically obtaining pairs of aligned sentences from parallel corporaText-Translation Alignment | | J93-2003&J99-4005 | 数据集 | Green | 源-信道模型参数估计 | built a source-channel model of translation between English and FrenchThe Mathematics of Statistical Machine Translation: Parameter Estimation | | J99-4005&10.1162_089120103321337458 | 模型 | Red | 解码问题NP-Complete | decoding problem for the IBM-4 translation model is provably NP-completeDecoding complexity in word-replacement translation models this problem is NP-complete | | 10.1162_089120103321337458&10.1162_0891201042544884 | 研究方法/流程 | Purple | 词重排与波束搜索 | similar algorithmWord reordering and a dynamic programming beam search algorithm for statistical machine translation | | 10.1162_0891201042544884&J07-2003 | 模型 | Red | 对齐模板方法 | alignment template translation model phrase-based systemsA phrase-based statistical machine translation approach-the alignment template approach-is described. The alignment template approach to statistical machine translation | | J07-2003&P08-1066 | 评估方法 | Pink | Hiero分层短语模型 | Hiero decoder does not require explicit syntactic representation Hiero system achieved about 1 to 3 point improvement in BLEU similar to similar to the Hiero systemHierarchical Phrase-Based Translation | | P08-1066&E09-1044 | 模型 | Red | 目标端依赖语言模型 | target dependency trees target dependency language model during decodingA New String-to-Dependency Machine Translation Algorithm with a Target Dependency Language Model | | E09-1044&W11-2211 | 研究方法/流程 | Purple | 浅层规则剪枝 | using shallow rulesreductions in the size of the rule sets used in translation modifications to hypothesis expansion in cube pruning | | W11-2211&10.1186_1471-2105-14-146 | 研究方法/流程 | Purple | 利用自动翻译语料 | using a corpus of automatic translations in addition to human translations provides a (small) improvement of translation qualityWe employ bitexts that have been built by automatically translating large amounts of monolingual data as additional parallel training corpora Our results show that integrating a second translation model with only non-hierarchical phrases extracted from the automatically generated bitexts is a reasonable approach | | 10.1186_1471-2105-14-146&10.1016_j.artmed.2014.01.004 | 研究方法/流程 | Purple | 构建MEDLINE平行语料 | propose a method for obtaining in-domain parallel corpora from titles and abstracts of publications in the MEDLINEdatabaseCombining MEDLINE and publisher data to create parallel corpora for the automatic translation of biomedical text | | 10.1016_j.artmed.2014.01.004&W14-3328 | 研究方法/流程 | Purple | 利用Web资源选数据 | use of web resources to complement the training resourcesin-domain training and tuning intelligent training data selection | | W14-3328&W14-3302 | 研究方法/流程 | Purple | WMT14医疗翻译任务 | WMT14 Medical Translation Taskattend the medical summary sentence unconstrained translation task of the Ninth Workshop on Statistical Machine Translation (WMT2014) | | W14-3302&Q15-1013 | 数据集 | Green | WMT 2014评测 | WMT 2014 shared translation taskFindings of the 2014 Workshop on Statistical Machine Translation | | Q15-1013&D15-1248 | 模型 | Red | 关系依赖语言模型 | dependency language model (RDLM) relational dependency language model We propose a language model for dependency structures that is relational rather than configurational as a feature function in string-to-tree SMT from English to German and Russian | | D15-1248&000493806800162 | 模型 | Red | 联合形态句法模型 | syntaxbased SMT resultsimprovements in translation quality of 1.4-1.8 BLEU We propose to model syntactic and morphological structure jointly in a dependency translation model A Joint Dependency Model of Morphological and Syntactic Structure for Statistical Machine Translation | | 000493806800162&000493806800009 | 工具或软件库 | Blue | BPE子词分割 | represent unseen words as sequences of subword units follow The network vocabulary size is 90 000 represent rare words via BPE settings and training procedure described by represent rare words (or morphemes in the case of Turkish) as character bigram sequences Neural Machine Translation of Rare Words with Subword Units segmentation based on the byte pair encoding compression algorithm encoding rare and unknown words as sequences of subword units simple character n-gram models | | 000493806800009&D16-1162 | 评估方法 | Pink | BLEU分数提升 | gains in BLEU gains made by the proposed method state-of-the-art performance on several language pairswe obtain substantial improvements on the WMT 15 task English <-> German (+2.8-3.7 BLEU) obtaining new state-of-the-art results | | D16-1162&000485630703053 | 研究方法/流程 | Purple | 低频词翻译错误 | NMT is prone to generate words that seem to be natural in the target sentence, but do not reflect the original meaning of the source sentence.Neural machine translation (NMT) often makes mistakes in translating low-frequency content words that are essential to understanding the meaning of the sentence. | | 000485630703053&000493984800064 | 背景信息/通用知识 | Brown | 中译英性能提升 | Chinese-to-English translationExperimental results on Chinese-English translation show that the proposed approach achieves significant and consistent improvements over state-of-the-art NMT and SMT systems on multiple NIST test sets. | | 000493984800064&000485488905004 | 研究方法/流程 | Purple | 建模源端句法 | syntax informationModeling Source Syntax for Neural Machine Translation | | 000485488905004&000386658300285 | 数据集 | Green | 多领域平行语料库 | discourse parallel corpus 500 thousand Portuguese-English sentence pairs in various domains such as news, laws, microblog etc. 2 million Chinese-EnglishExperimental results on both Chinese English | | 000386658300285&W11-2132 | 研究方法/流程 | Purple | SMT领域自适应 | approachDomain Adaptation for Statistical Machine Translation | | W11-2132&E12-1014 | 研究方法/流程 | Purple | 利用单语数据 | use an existing SMT system to discover parallel sentences within independent monolingual texts, and use them to re-train and enhance the systemInvestigations on Translation Model Adaptation Using Monolingual Data | | E12-1014&W13-2233 | 数据集 | Green | 利用可比语料 | use comparable corpora to score an existing Spanish-English phrase table extracted from the Europarl corpus compute both phrasal features and lexically smoothed features (using word alignments, like the Moses lexical translation probabilities) we use time-stamped web crawls as well as interlingually linked Wikipedia documents the techniques proposed by and We extend existing research on bilingual lexicon induction to estimate both lexical and phrasal translation probabilities for MT-scale phrasetables we examine an idealization where a phrase-table is given We propose a novel algorithm to estimate reordering probabilities from monolingual data Toward Statistical Machine Translation without Parallel Corpora | | W13-2233&D14-1061 | 研究方法/流程 | Purple | 低资源词典构建 | build a translation lexicon from non parallel or comparable dataCombining Bilingual and Comparable Corpora for Low Resource Machine Translation | | D14-1061&000493806800185 | 研究方法/流程 | Purple | 联合对齐与破译 | decipherment problemBeyond Parallel Data: Joint Word Alignment and Decipherment Improves Machine Translation | | 000493806800185&D16-1160 | 研究方法/流程 | Purple | 半监督学习 | propose a similar semi-supervised framework similar semi-supervised reconstruction methodSemi-Supervised Learning for Neural Machine Translation | | D16-1160&D16-1249 | 研究方法/流程 | Purple | 利用源端单语数据 | achieve better results monolingual corporaExploiting Source-side Monolingual Data in Neural Machine Translation | | D16-1249&D16-1096 | 研究方法/流程 | Purple | 监督注意力机制 | supervised alignmentsSupervised Attentions for Neural Machine Translation | | D16-1096&000493984800177 | 模型 | Red | 覆盖率模型 | coverage modelCoverage Embedding Models for Neural Machine Translation | | 000493984800177&D17-1012 | 模型 | Red | NMT引入句法 | use of syntactic structures in NMT modelsImproved Neural Machine Translation with a Syntax-Aware Encoder and Decoder | | D17-1012&000493992300012 | 模型 | Red | 共享负采样 | shared the negative samples of each target word in a sentence in training time | | 000493992300012&W17-5708 | 模型 | Red | 融合RNNG到NMT | or the decoder modifying the encodercombining the recurrent neural network grammar into the attention-based neural machine translation | | W17-5708&N18-2084 | 研究方法/流程 | Purple | 预训练词嵌入初始化 | pre-training is properly integrated into the NMT system using pre-trained embeddings in NMT pre-trained word embeddings have been used either in standard translation systems pre-training is usefulA Bag of Useful Tricks for Practical Neural Machine Translation: Embedding Layer Initialization and Large Batch Size | | N18-2084&000865723400103 | 数据集 | Green | 低资源翻译挑战 | multilingual training for low-resource translationThe performance of Neural Machine Translation (NMT) systems often suffers in lowresource scenarios where sufficiently largescale parallel corpora cannot be obtained. | | 000865723400103&000900116903092 | 评估方法 | Pink | NMT新语言快速适应 | same four languages as -Azerbeijani (Az), Belarusian (Be), Galician (Gl) and Slovak (Sk) trained many-to-one models from 58 languages into English use tokenized BLEU in order to be comparable with recently published results on this dataset best fine-tuned many-to-one models of Rapid Adaptation of NMT to New Languages by training multilingual models the 59-language TED Talks corpus best many-to-one models in massively multilingual many-to-many models with up to 58 languages to-and-from Englishimproves over other adaptation methods by 1.7 BLEU points average over 4 LRL settings jointly train on both a LRL of interest and a similar high-resourced language to prevent over-fitting Rapid Adaptation of Neural Machine Translation to New Languages massively multilingual models, even without any explicit adaptation, are surprisingly effective massively multilingual seed models achieving BLEU scores of up to 15.5 | | 000900116903092&000736531900047 | 模型 | Red | 大规模多对多模型 | multilingual translation jointly train one translation model that translates multiple language directions at the same time shares representations to improve the translation performance on low-resource languagesmassively multilingual many-to-many models are effective in low resource settings Massively Multilingual Neural Machine Translation training a single model that supports translation from multiple source languages into multiple target languages | | 000736531900047&001181866502027 | 模型 | Red | mBART降噪预训练 | mBART fine-tunes with traditional on-the-fly backtranslation pre-trains on a variety of language configurationsMultilingual Denoising Pre-training for Neural Machine Translation | | 001181866502027&000570978201083 | 研究方法/流程 | Purple | 统一概率框架 | generative modelingour approach results in higher BLEU scores over state-of-the-art unsupervised models We present a probabilistic framework for multilingual neural machine translation that encompasses supervised and unsupervised setups, focusing on unsupervised translation | | 000570978201083&000663162000001 | 数据集 | Green | 在线回译 | online backtranslation to improve the performance of non-English pairs multilingual translation models factorize computation when translating to many languages and share information between similar languages English-Centric multilingual model from on the OPUS100 corpus Using backtranslation thus requires the ability to translate in both directions Mining data for each and every language pair is prohibitive - previous work circumvents this issue by focusing only on the 99 pairs that go through English train on 100 directions increasing model capacityExperiments on OPUS-100 (a novel multilingual dataset with 100 languages) Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation stronger modeling capacity OPUS-100 (a novel multilingual dataset with 100 languages) propose random online backtranslation | | 000663162000001&000923421200003 | 评估方法 | Pink | 非英语中心多语言翻译 | multilingual machine translationBeyond English-Centric Multilingual Machine Translation | | 000923421200003&2023.acl-long.524 | 数据集 | Green | FLORES-101评测基准 | FLORES-101 evaluation set FLORES-101 evaluation dataThe Flores-101 Evaluation Benchmark for Low-Resource and Multilingual Machine Translation | | 2023.acl-long.524&2023.acl-long.859 | 研究方法/流程 | Purple | 偶发双语主义 | present in the training data as a marker for multilingual content attributing LLM MT capabilities to the presence of incidental bilingual examplesSearching for Needles in a Haystack: On the Role of Incidental Bilingualism in PaLM’s Translation Capability PaLM is exposed to over 30 million translation pairs across at least 44 languages incidental bilingualism—the unintentional consumption of bilingual signals, including translation examples | | 2023.acl-long.859&001371932502043 | 评估方法 | Pink | 大模型提示(Prompting) | large-scale models with parameters in the hundreds of billionsPrompting PaLM for Translation: Assessing Strategies and Performance | | 001371932502043&001156229800010 | 模型 | Red | LLM翻译能力与错误分析 | LLMs can produce fluent and adequate translations, especially for high-resource English-centric language pairs, that are competitive with those of dedicated supervised translation models translation errors, even severely critical ones, obtained via prompting a LLM are different from those produced by traditional machine translation modelsa transformer decoder-only model trained solely with self-supervised learning is able to match specialized supervised state-of-the-art models | | 001156229800010&2023.emnlp-main.733 | 研究方法/流程 | Purple | LLM幻觉研究 | apply the minimal perturbations of , including misspelling and title-casing words, and inserting frequent tokens at the beginning of the source sentenceHallucinations in Large Multilingual Translation Models | | 2023.emnlp-main.733&2024.emnlp-main.802 | 研究方法/流程 | Purple | MBR解码与重排序 | reference-based and QE metrics to rerank multiple hypotheses generated by dedicated MT models or large language models (LLMs), aiming to improve translation qualitythe strategy to produce the final translation (instruction-based, quality-based reranking, and minimum Bayes risk (MBR) decoding) Our results show that MBR decoding is a very effective method | | 2024.emnlp-main.802&2024.emnlp-main.1152 | 评估方法 | Pink | 评估指标与人类判断 | examining their correlation with human judgments, as well as their Precision, Recall, and F -score evaluated MT metrics' ability to assess high-quality translationsCan Automatic Metrics Assess High-Quality Translations? | | 2024.emnlp-main.1152&2024.emnlp-main.914 | 背景信息/通用知识 | Brown | 评估指标的可解释性 | not easy to interpretBeyond Correlation: Interpretable Evaluation of Machine Translation Metrics |
