CraftAssist Instruction Parsing: Semantic Parsing for a Voxel-World Assistant

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 437–442 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 437 Opportunistic Decoding with Timely Correction for Simultaneous Translation Renjie Zheng 1,2,∗ Mingbo Ma 2,∗ Baigong Zheng 2 Kaibo Liu2 Liang Huang 1,2 1Oregon State University, Corvallis, OR, USA 2Baidu Research, Sunnyvale, CA, USA {renjiezheng,mingboma,baigongzheng}@baidu.com {kaiboliu,lianghuang}@baidu.com Abstract Simultaneous translation has many important application scenarios and attracts much atten- tion from both academia and industry recently. Most existing frameworks, however, have dif- ficulties in balancing between the translation quality and latency, i.e., the decoding policy is usually either too aggressive or too conser- vative. We propose an opportunistic decoding technique with timely correction ability, which always (over-)generates a certain mount of ex- tra words at each step to keep the audience on track with the latest information. At the same time, it also corrects, in a timely fashion, the mistakes in the former overgenerated words when observing more source context to ensure high translation quality. Experiments show our technique achieves substantial reduction in la- tency and up to +3.1 increase in BLEU, with revision rate under 8% in Chinese-to-English and English-to-Chinese translation. 1 Introduction Simultaneous translation, which starts translation before the speaker finishes, is extremely useful in many scenarios, such as international conferences, travels, and so on. In order to achieve low latency, it is often inevitable to generate target words with insufficient source information, which makes this task extremely challenging. Recently, there are many efforts towards balanc- ing the translation latency and quality with mainly two types of approaches. On one hand, Ma et al. (2019a) propose very simple frameworks that de- code following a fixed-latency policy such as wait- k. On the other hand, there are many attempts to learn an adaptive policy which enables the model to decide READ or WRITE action on the fly using various techniques such as reinforcement learning (Gu et al., 2017; Alinejad et al., 2018; Grissom II ∗These authors contributed equally. yt <latexit sha1_base64="SeGBdmF0K269BzW9NTI6ylilwtw=">AB6nicbVDLSgNBEOyNrxhfUY9eB oPgKez6QI9BLx4jmgckS5idzCZDZmeXmV4hLPkELx4U8eoXefNvnCR70MSChqKqm+6uIJHCoOt+O4WV1bX1jeJmaWt7Z3evH/QNHGqGW+wWMa6HVDpVC8gQIlbyea0yiQvBWMbqd+64lrI2L1iOE+xEdKBEKR tFKD+Me9soVt+rOQJaJl5MK5Kj3yl/dfszSiCtkhrT8dwE/YxqFEzySambGp5QNqID3rFU0YgbP5udOiEnVumTMNa2FJKZ+nsio5Ex4yiwnRHFoVn0puJ/XifF8NrPhEpS5IrNF4WpJBiT6d+kLzRnKMeWUKaFvZ WwIdWUoU2nZEPwFl9eJs2zqndevby/qNRu8jiKcATHcAoeXEN7qAODWAwgGd4hTdHOi/Ou/Mxby04+cwh/IHz+QN1Bo3r</latexit> ˆy6w t <latexit sha1_base64="KB9f+IQoMdoV0w7Aq9F4Qd4Ea5M=">ACXicbVC5TsNAEF2HK4QrQEmzIkKimwOQ RlBQxkckixsdabdbLK+mB3DLIstzT8Cg0FCNHyB3T8DZvEBSQ8aSn92Y0M8+LBVdgmt9GaWFxaXmlvFpZW9/Y3Kpu7RVlEjKWjQSkex6RDHBQ9YCDoJ1Y8lI4AnW8UaXY79z6TiUXgDacycgAxC7nNKQEtuFdtDApkdEBh6fpbmuQu 3mS3YnRIkBPyQu9WaWTcnwPEKkgNFWi61S+7H9EkYCFQZTqWYMTkYkcCpYXrETxWJCR2TAepqGJGDKySaf5PhAK3sR1KXj9Rf09kJFAqDTzdOT5ZzXpj8T+vl4B/7mQ8jBNgIZ0u8hOBIcLjWHCfS0ZBpJoQKrm+FdMhkYSCDq+iQ 7BmX54n7aO6dVw/vT6pNS6KOMpoD+2jQ2ShM9RAV6iJWoiR/SMXtGb8WS8GO/Gx7S1ZBQzu+gPjM8fJaObQg=</latexit> revision window decoding time t <latexit sha1_base64="t6XaytdIsHwdU4AeCNDSjPNP5sM=">AB6HicbVDLSgNBEOz1GeMr6tHLYBA8hV0f6DHoxWMC5gHJEmYns8mY2dlplcIS7AiwdFvPpJ 3vwbJ8keNLGgoajqprsrSKQw6Lrfzsrq2vrGZmGruL2zu7dfOjhsmjVjDdYLGPdDqjhUijeQIGStxPNaRI3gpGd1O/9cS1EbF6wHC/YgOlAgFo2ilOvZKZbfizkCWiZeTMuSo9Upf3X7M0ogrZJIa0/HcBP2MahRM8kmxmxqeUDaiA96xVNGIGz+bHTohp1bpkzDWthSmfp7IqORMeMosJ0RxaFZ9Kbif14nxfDGz4RKUuSKzReFqSQYk+ nXpC80ZyjHlCmhb2VsCHVlKHNpmhD8BZfXibN84p3UbmqX5art3kcBTiGEzgD6hCvdQgwYw4PAMr/DmPDovzrvzMW9dcfKZI/gD5/MH4m2M/w=</latexit> t + 1 <latexit sha1_base64="tGQ1GTwhm57M/cuPNJRIL/VDcNU=">AB6nicbVDLSgNBEOz1GeMr6tHLYBAEIez6QI9BLx4jmgckS5idzCZDZmeXmV4hLPkELx4U8eoXefNvnCR70MSChqKqm+6uI JHCoOt+O0vLK6tr64WN4ubW9s5uaW+/YeJUM15nsYx1K6CGS6F4HQVK3ko0p1EgeTMY3k785hPXRsTqEUcJ9yPaVyIUjKVHvDU65bKbsWdgiwSLydlyFHrlr46vZilEVfIJDWm7bkJ+hnVKJjk42InNTyhbEj7vG2pohE3fjY9dUyOrdIjYaxtKSRT9fdERiNjRlFgOyOKAzPvTcT/vHaK4bWfCZWkyBWbLQpTSTAmk79JT2jOUI4soUwLeythA6opQ5tO0Ybgzb+8SBpnFe+8cnl/Ua7e5HEU4BCO4 AQ8uIq3EN6sCgD8/wCm+OdF6cd+dj1rk5DMH8AfO5w+4yY1v</latexit> correction 8 > > < > > : <latexit sha1_base64="vOXxVXrmuiv79AW8seb2TK J6MCA=">ACN3icfVDLSgMxFM34dnxVXboJFsVmfGBLotuXImCVaEpJZPeaYOZzJDcEcvQv3Ljb7jTjQtF3PoH prWCL7wkcDjnHu69J8qUtBgE97I6Nj4xOTUtD8zOze/UFpcOrNpbgTURKpScxFxC0pqKFEBReZAZ5ECs6jy4O+ fn4FxspUn2I3g0bC21rGUnB0VLN0xBTEyAoWQVvqIuFo5HXPp+vUPcboP8hnoFufDmZku4OVZqkcVIJB0d8gHI yGdZxs3THWqnIE9AoFLe2HgYZNgpuUAoFPZ/lFjIuLnkb6g5qnoBtFIO7e3TNMS0ap8Z9jXTAfnUPLG2m0Su063 ZsT+1PvmXVs8x3msUmc5ghYfg+JcUxpP0TakgYEq4DXBjpdqWiw0X6KL2XQjhz5N/g7PNSrhV2TnZLlf3h3F MkRWySjZISHZJlRySY1IjgtyQB/JEnr1b79F78V4/Wke8oWeZfCv7R1tL6Y8</latexit> … … … 8 > > < > > : <latexit sha1_base64="vOXxVXrmuiv79AW8seb2TKJ6MCA=">AC N3icfVDLSgMxFM34dnxVXboJFsVmfGBLotuXImCVaEpJZPeaYOZzJDcEcvQv3Ljb7jTjQtF3PoHprWCL7wkcDjnHu69J8qUtBgE97I6Nj4xOTU tD8zOze/UFpcOrNpbgTURKpScxFxC0pqKFEBReZAZ5ECs6jy4O+fn4FxspUn2I3g0bC21rGUnB0VLN0xBTEyAoWQVvqIuFo5HXPp+vUPcboP8h noFufDmZku4OVZqkcVIJB0d8gHIyGdZxs3THWqnIE9AoFLe2HgYZNgpuUAoFPZ/lFjIuLnkb6g5qnoBtFIO7e3TNMS0ap8Z9jXTAfnUPLG2m0S u063ZsT+1PvmXVs8x3msUmc5ghYfg+JcUxpP0TakgYEq4DXBjpdqWiw0X6KL2XQjhz5N/g7PNSrhV2TnZLlf3h3FMkRWySjZISHZJlRySY1 IjgtyQB/JEnr1b79F78V4/Wke8oWeZfCv7R1tL6Y8</latexit> t + 2 <latexit sha1_base64="9+8co5X+9cuaDI/4vU0v8CjHtc=">AB6nicbVDLSgNBEOyNrxhfUY9eBoMgCGE3KnoMevEY0TwgWcLsZJIMmZ1dZnqFsOQTvHhQxKtf5M2/cZLsQRMLGoqbrq7glg Kg67eRWVtfWN/Kbha3tnd294v5Bw0SJZrzOIhnpVkANl0LxOgqUvBVrTsNA8mYwup36zSeujYjUI45j7od0oERfMIpWesCzSrdYcsvuDGSZeBkpQYZat/jV6UsCblCJqkxbc+N0U+pRsEknxQ6ieExZSM64G1LFQ258dPZqRNyYpUe6UfalkIyU39PpDQ0ZhwGtjOkODSL3lT8z2sn2L/2U6HiBLli80X9RBKMyPRv0hOaM5RjSyjTwt5K2JBqytCmU7AheIsvL5NGpeydly/vL0rVmyOPBzBMZyCB1d QhTuoQR0YDOAZXuHNkc6L8+58zFtzTjZzCH/gfP4Auk2NcA=</latexit> correction y<t <latexit sha1_base64="5wmZKBhkP+VlNy8pxWZS+/TmoiA=">AB+nicbVDLSsNAFL2pr1pfqS7dDBbBVUl8oAs XRTcuK9gHtCFMpN26OTBzEQJMZ/ixoUibv0Sd/6NkzYLrR4YOJxzL/fM8WLOpLKsL6OytLyulZdr21sbm3vmPXdrowSQWiHRDwSfQ9LylIO4opTvuxoDjwO150+vC791TIVkU3qk0pk6AxyHzGcFKS65Zz4YBVhPz9I8d7NLlbtmw2paM 6C/xC5JA0q0XfNzOIpIEtBQEY6lHNhWrJwMC8UIp3ltmEgaYzLFYzrQNMQBlU42i56jQ62MkB8J/UKFZurPjQwHUqaBpyeLnHLRK8T/vEGi/AsnY2GcKBqS+SE/4UhFqOgBjZigRPFUE0wE01kRmWCBidJt1XQJ9uKX/5LucdM+aZ7dnjZaV2U dVdiHAzgCG86hBTfQhg4QeIAneIFX49F4Nt6M9/loxSh39uAXjI9vLjKUmg=</latexit> irreversible Figure 1: Besides yt, opportunistic decoding continues to generate additional w words which are represented as ˆy⩽w t . The timely correction only revises this part in future steps. Different shapes denote different words. In this example, from step t to t + 1, all previously op- portunistically decoded words are revised, and an ex- tra triangle word is generated in opportunistic window. From step t + 1 to t + 2, two words from previous op- portunistic window are kept and only the triangle word is revised. et al., 2014), …

Content truncated to 500 words