7-2 等長機器翻譯研究及配音字幕的優化.pdf

《7-2 等長機器翻譯研究及配音字幕的優化.pdf》由會員分享，可在線閱讀，更多相關《7-2 等長機器翻譯研究及配音字幕的優化.pdf（35頁珍藏版）》請在三個皮匠報告上搜索。

1、Isometric Machine Translation for Subtitling等長機器翻譯及配音字幕優化楊浩華為-文本機器翻譯實驗室 主任Hao Yang,Director of Text Machine Translation Laboratories,Huawei|01NMT Basics&Trends02What Is Isometric MT03Isometric MT Architecture04Isometric MT Application目錄Content|01NMT Trends|There is a white dog on the grass.草地上有只白色小狗

2、。f(x)convert sentenceFrom some languagesInto another languageMT Problem-機器翻譯，肖桐，朱靖波等，2021SMT ProblemTranslation modelLanguage modelCompute argmaxNMT ProblemEnd-to-End modelNo feature extraction layerNo fine tune layerEncoder-Decoder ModelEncoder modelDecoder modelEncoded vectorHow to configure s5700

3、 arp?Source sentenceC如何如何怎么配置設置S5700S2700?arpap如何配置配置S5700S5700arparp?Target sentenceEncoder-Decoder ModelEncoder：RNNHidden state:encoded vectorDecoder:RNN1.3.1 Seq2Seq with Neural NetworksDecodingGreedy decodingBeam searchSamplingGreedy DecodingAt each step,keep several best hypothesesBeam SearchDe

4、codingSMT VS NMT=33.3 VS 34.81SMT Re-ranking=33.3 VS 36.6How to configure S5700 arp?如何 配置 s5700 arp?如何 配置 s5700 Seq2Seq Model ProblemContext vector is bottleneckPerformance degrades as sentence becomes longerHow to configure s5700 arp?Source sentenceC如何配置設置如何配置SoftmaxAttentionAttentionSMT:33.3Seq2Se

5、q:34.81RNNSearch(RNNAttn):36.15Transformer:41.8*WMT14 EN/FRTransformerEncoder self attentionDecoder self attentionCross attentionPerformanceAshish Vaswani,Attention is all you needhttps:/ Attention is all you needTransformerNo RNNAttentionPosition embeddingOn the WMT 2014 English-to-French a new sin

6、gle-model SOTA BLEU score of 41.8 after training for 3.5 days on eight GPUshttp:/speech.ee.ntu.edu.tw/tlkagk/courses_DLHLP20.htmlTransformerBLEU 41.83.5 days on eight GPUsPre-trained model familyMore Machine Translation TrendsMulti-device 多設備Multi-screen 多屏幕Real-time 實時Phone/TabletWearableCarLaptop

7、Tablet Phone WatchTV DesktopAR/VRSource:https:/redian.news/news/12146;https:/youtu.be/niM4ttonrrA;https:/ is Isometric MT|What Is Isometric MT Isometric MTGenerates translationsEnsures SRC/MT in similar lengthApplication ScenariosAutomatic dubbingSubtitle fittingSimultaneous speech translationLayout

8、 constrained translationWhat Is Isometric MT Isometric MT MetricsTranslation Quality(TQ)BLEU BERTScoreLength Compliance(LC)LC LRBLEUBERTScoreLCIsometric MT TasksIWSTL 2022 Isometric MT TasksObjectives:Translation directions:En De/Fr/Es Subtitle translation scenarios Ensuring quality of sentence tran

9、slation Ensuring consistent length between MT and SRCDifficulties TQ vs.LCSRC:Its the one wheel XR and if you dont know what a one wheel is.MT1:Es ist das eineRad XR und wennSienicht wissen,was ein Rad ist.MT2:Es ist das One Wheel XR und wenn Sienichtwissen,was ein One Wheel ist.SRC:Its basically a

10、motorized one wheel skateboard that can go onMT1:Es ist ein motorisiertesRad-Skateboard,das weiter gehen kannMT2:Es ist im Grundeein motorisiertesSkateboard miteinemRad,das weitergehen kannSRC:This is my new toy.MT1:Das istmein neues.MT2:Das istmein neues Spielzeug.ExampleMore diversityLess missing

11、translation03Isometric MT Architecture|Isometric MT ArchitectureTranslation ModelAT/NATLC StrategyLCD Length Token Method Length Encoding MethodLAB Length-Aware BeamRe-rankEnsemble MT ModelsMT Model ScoringIsometric MT Architecture 1:Model AugmentationAT Model AugmentationLow-resource model augmenta

12、tion Shared decoder/embedding Multilingual model en2de,de2en Data diversification Sampling BT FT+BT R-Drop&ensemblePerformance BLEU 4-10+Isometric MT Architecture 2:LCD StrategyLCD:Length Controlled DecodingLength token Three-category tagging Short 0.9 norm 1.1 AT NAT+LCD+LAB can satisfy TQ+LC04Isom

13、etric MT Applications|Real-World ExperienceWithout Isometric MTWith Isometric MTHuawei Translate PlatformReferenceLi,Zongyao,et al.HW-TSCs Participation in the IWSLT 2022 Isometric Spoken Language Translation.Proceedings of the 19th InternationalConference on Spoken Language Translation(IWSLT 2022).

14、2022.Wang,Minghan,et al.HI-CMLM:Improve CMLM with Hybrid Decoder Input.Proceedings of the 14th International Conference on Natural Language Generation.2021.Lakew,Surafel Melaku,MattiaDi Gangi,and Marcello Federico.Controlling the output length of neural machine translation.arXivpreprint arXiv:1910.10408(2019).Takase,Sho,and Naoaki Okazaki.Positional encoding to control output sequence length.arXivpreprint arXiv:1904.07418(2019).Ghazvininejad,Marjan,et al.Mask-predict:Parallel decoding of conditional masked language models.arXiv preprint arXiv:1904.09324(2019).|Thank You 感謝您的觀看