A Survey on Zero Pronoun Translation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3325–3339 July 9-14, 2023 ©2023 Association for Computational Linguistics A Survey on Zero Pronoun Translation Longyue Wang∗, Siyou Liu∗, Mingzhou Xu, Linfeng Song, Shuming Shi, Zhaopeng Tu Tencent AI Lab {vinnylywang,lifengjin,shumingshi,zptu}@tencent.com guofeng-ai@googlegroups.com Abstract Zero pronouns (ZPs) are frequently omitted in pro-drop languages (e.g. Chinese, Hungar- ian, and Hindi), but should be recalled in non- pro-drop languages (e.g. English). This phe- nomenon has been studied extensively in ma- chine translation (MT), as it poses a significant challenge for MT systems due to the difficulty in determining the correct antecedent for the pronoun. This survey paper highlights the ma- jor works that have been undertaken in zero pronoun translation (ZPT) after the neural revo- lution so that researchers can recognize the cur- rent state and future directions of this field. We provide an organization of the literature based on evolution, dataset, method, and evaluation. In addition, we compare and analyze compet- ing models and evaluation metrics on different benchmarks. We uncover a number of insight- ful findings such as: 1) ZPT is in line with the development trend of large language model; 2) data limitation causes learning bias in lan- guages and domains; 3) performance improve- ments are often reported on single benchmarks, but advanced methods are still far from real- world use; 4) general-purpose metrics are not reliable on nuances and complexities of ZPT, emphasizing the necessity of targeted metrics; 5) apart from commonly-cited errors, ZPs will cause risks of gender bias. 1 Introduction Pronouns play an important role in natural lan- guage, as they enable speakers to refer to people, objects, or events without repeating the nouns that represent them. Zero pronoun (ZP)1 is a complex phenomenon that appears frequently in pronoun- dropping (pro-drop) languages such as Chinese, Hungarian, and Hindi. Specifically, pronouns are often omitted when they can be pragmatically work.∗Longyue Wang and Siyou Liu contributed equally to this 1ZP is also called dropped pronoun. The linguistic concept is detailed in Appendix §A.3. or grammatically inferable from intra- and inter- sentential contexts (Li and Thomson, 1979). Since recovery of such ZPs generally fails, this poses difficulties for several generation tasks, including dialogue modelling (Su et al., 2019), question an- swering (Tan et al., 2021), and machine transla- tion (Wang, 2019). When translating texts from pro-drop to non-pro- drop languages (e.g. Chinese⇒English), this phe- nomenon leads to serious problems for translation models in terms of: 1) completeness, since transla- tion of such invisible pronouns cannot be normally reproduced; 2) correctness, because understanding the semantics of a source sentence needs to identi- fying and resolving the pronominal reference. Figure 1 shows ZP examples in three typological patterns determined by language family (detailed in Appendix §A.1). Taking a full-drop language for instance, the first-person subject and third-person object pronouns are omitted in Hindi input while these pronouns are all compulsory in English trans- lation. This is not a problem for human beings since we can …

Content truncated to 500 words