QuickEdit: Editing Text & Translations by Crossing Words Out

Proceedings of NAACL-HLT 2018, pages 272–282 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics QuickEdit: Editing Text & Translations by Crossing Words Out David Grangier Michael Auli Facebook AI Research Menlo Park, CA Abstract We propose a framework for computer- assisted text editing. It applies to translation post-editing and to paraphrasing. Our proposal relies on very simple interactions: a human editor modifies a sentence by marking tokens they would like the system to change. Our model then generates a new sentence which reformulates the initial sentence by avoiding marked words. The approach builds upon neu- ral sequence-to-sequence modeling and intro- duces a neural network which takes as input a sentence along with change markers. Our model is trained on translation bitext by sim- ulating post-edits. We demonstrate the ad- vantage of our approach for translation post- editing through simulated post-edits. We also evaluate our model for paraphrasing through a user study. 1 Introduction Computers can help humans edit text more effi- ciently. In particular, statistical models are used for that purpose, for instance to help correct spelling mistakes (Brill and Moore, 2000) or sug- gest likely completions of a sentence (Bickel et al., 2005). In this work, we rely on statistical learn- ing to enable a computer to rephrase a sentence by only pointing at words that should be avoided. Specifically, we consider the task of reformulat- ing either a sentence, i.e. paraphrasing (Quirk et al., 2004), or a translation, i.e. translation post- editing (Koehn, 2009b). Paraphrasing reformu- lates a sentence with different words preserving its meaning, while translation post-editing takes a candidate translation along with the corresponding source sentence and improves it. Our proposal relies on very simple interactions: a human editor modifies a sentence by selecting tokens they would like the system to replace and no other feedback. Our system then generates a new sentence which reformulates the initial sen- tence by avoiding the word types from the se- lected tokens. Our approach builds upon neural sequence-to-sequence and introduces a neural net- work which takes as input a sentence along with token markers. We introduce a novel attention- based architecture suited to this goal and propose a training procedure based on simulated post-edits on translation bitext (§3). This approach allows to get substantial modifications of the initial sentence – including deletion, reordering and insertion of multiple words – with limited user effort. Our experiments (§4) relies on large scale simu- lated post-edits. They show that our model outper- forms our post-editing baseline by up to 5 BLEU points on WMT’14 English-German and WMT’14 German-English translation. The advantage of our method is also highlighted in monolingual set- tings, where we analyze the quality of the para- phrases generated by our model in a user study. Before introducing our method (§3) and its em- pirical evaluation (§4), we describe related work in the next section. 2 Related Work Our work builds upon previous research on neu- ral machine translation, machine translation post- editing, and computer-assisted …

Content truncated to 500 words