Privacy-Preserving News Recommendation Model Learning

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 126–136 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 126 An Attentive Recurrent Model for Incremental Prediction of Sentence-final Verbs Wenyan Li Comcast AI Research Lab wenyanl9562@gmail.com Alvin Grissom II Haverford College agrissom@haverford.edu Jordan Boyd-Graber University of Maryland jbg@umiacs.umd.edu Abstract Verb prediction is important for understand- ing human processing of verb-final languages, with practical applications to real-time simul- taneous interpretation from verb-final to verb- medial languages. While previous approaches use classical statistical models, we introduce an attention-based neural model to incremen- tally predict final verbs on incomplete sen- tences in Japanese and German SOV sentences. To offer flexibility to the model, we further in- corporate synonym awareness. Our approach both better predicts the final verbs in Japanese and German and provides more interpretable explanations of why those verbs are selected. 1 Introduction Final verb prediction is fundamental to human language processing in languages with subject- object-verb (SOV) word order, such as German1 and Japanese, (Kamide et al., 2003; Momma et al., 2014; Chow et al., 2018) particularly for simultane- ous interpretation, where an interpreter generates a translation in real time. Instead of waiting until the entire sentence is completed, simultaneous in- terpretation requires translation of the source text units while the interlocutor is speaking. When human simultaneous interpreters trans- late from an SOV language to an SVO one incrementally—without waiting for the final verb at the end of a sentence—they must use strategies to reduce the lag, or delay, between the time they hear the source words and the time they translate them (Wilss, 1978; He et al., 2016). One strategy is final verb prediction: since the verb comes late in the source sentence but early in the target trans- lation, if the verb is predicted in advance, it can be translated before it is heard, allowing for a more 1German is rich in both SOV and SVO sentences. It has been argued that its underlying structure is SOV (Bach, 1962; Koster, 1975), but this is not immediately relevant to our task. German Cazeneuve dankte dort den Män- nern und sagte, ohne deren kühlen Kopf hätte es vielleicht ein “furchtbares Drama” gegeben. English Cazeneuve thanked the men there and said that without their cool heads there might have been a “terrible drama”. Japanese また大和国奈良県の葛城山に 篭り密教の宿曜秘法を習得したとも 言 言 言わ わ わ. English It also said that he was acquainted with a secret lodging accommodation in Kat- suragiyama in Nara Prefecture of Yamato. Figure 1: An example of the verb position difference between SOV and SVO languages, where the final verb in German and Japanese is expected much earlier in their English translation. “simultaneous” (or monotonic) translation (Jörg, 1997; Bevilacqua, 2009; He et al., 2015). Further- more, Chernov et al. (2004) argue that simultane- ous interpreters’ probabilty estimates and predic- tions of the verbal and semantic structure of pre- ceeding messages facilitates simultaneity in human simultaneous interpretation. Like for human translation, simultaneous ma- chine translation (SMT), becomes more monotonic for …

Content truncated to 500 words