Incremental Re-training for Post-editing SMT
Year: 2010
Type: conference-paper
Abstract: A method is presented for incremental retraining\nof an SMT system, in which a local\nphrase table is created and incrementally updated\nas a file is translated and post-edited.\nIt is shown that translation data from within\nthe same file has higher value than other\ndomain-specific data. In two technical domains,\nwithin-file data increases BLEU score\nby several full points. Furthermore, a strong\nrecency effect is documented; nearby data\nwithin the file has greater value than more\ndistant data. It is also shown that the value\nof translation data is strongly correlated with\na metric defined over new occurrences of ngrams.\nFinally, it is argued that the incremental\nre-training prototype could serve as the basis\nfor a practical system which could be interactively\nupdated in real time in a post-editing\nsetting. Based on the results here, such an interactive\nsystem has the potential to dramatically\nimprove translation quality.
Authors: Daniel Hardt, Jakob Elming
Institution: University of Copenhagen
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3170166771