Delexicalized Cross-lingual Dependency Parsing for Xibe

Proceedings of Recent Advances in Natural Language Processing, pages 1626–1635 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_182 1626 Delexicalized Cross­lingual Dependency Parsing for Xibe He Zhou and Sandra Kübler Indiana University {hzh1, skuebler}@iu.edu Abstract Manually annotating a treebank is time­ consuming and labor­intensive. We conduct delexicalized cross­lingual dependency pars­ ing experiments, where we train the parser on one language and test on our target language. As our test case, we use Xibe, a severely under­resourced Tungusic language. We as­ sume that choosing a closely related language as the source language will provide better re­ sults than more distant relatives. However, it is not clear how to determine those closely re­ lated languages. We investigate three differ­ ent methods: choosing the typologically clos­ est language, using LangRank, and choosing the most similar language based on perplexity. We train parsing models on the selected lan­ guages using UDify and test on different gen­ res of Xibe data. The results show that lan­ guages selected based on typology and per­ plexity scores outperform those predicted by LangRank; Japanese is the optimal source lan­ guage. In determining the source language, proximity to the target language is more impor­ tant than large training sizes. Parsing is also influenced by genre differences, but they have little influence as long as the training data is at least as complex as the target. 1 Introduction For a severely low­resource language, construct­ ing a dependency treebank is labor­intensive and time­consuming, and annotators are difficult to find. Expanding a small treebank via monolin­ gual dependency parsing leads to suboptimal re­ sults since we lack enough training data to train a reliable parser. This situation has led to an increasing interest in techniques for supporting low­resource languages by taking advantage of high­resource languages together with methods for cross­lingual transfer (Meechan­Maddon and Nivre, 2019). This is facilitated by the Univer­ sal Dependencies (UD) project, which has resulted in a treebank collection covering a wide range of language, with the goal of facilitating multilingual parser development (Nivre et al., 2020). The lat­ est release (v2.7) covers 183 treebanks in 104 lan­ guages (Zeman, 2020). In our current work, we carry out preliminary single­source cross­lingual delexicalized dependency parsing experiments for the Xibe language. With this method, we train a parser on the treebank of one source language and parse the target language, with both treebanks delexicalized to abstract away from lexical differ­ ences between the two languages. Choosing the source language is crucial for single­source cross­lingual parsing. The optimal source language needs to be syntactically close to the target language as well as high­resourced. However, it is not obvious how to select this lan­ guage. We investigate three methods for select­ ing the source language: We compare LangRank (Lin et al., 2019) and typology, and we investigate whether using perplexity as a similarity metric can approximate typological knowledge. Then we in­ vestigate whether the size of the source treebank or a genre mismatch affect the quality of the parser. The remainder of this paper is organized as …

Content truncated to 500 words