Towards Semi-Automatic Generation of Proposition Banks for Low-Resource Languages

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 993–998, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Towards Semi-Automatic Generation of Proposition Banks for Low-Resource Languages Alan Akbik IBM Research Almaden Research Center San Jose, CA 95120 Vishwajeet Kumar IIT Bombay CS and Engineeering Mumbai, India {akbika,yunyaoli}@us.ibm.com vishwajeetkumar86@gmail.com Yunyao Li IBM Research Almaden Research Center San Jose, CA 95120 Abstract Annotation projection based on parallel cor- pora has shown great promise in inexpensively creating Proposition Banks for languages for which high-quality parallel corpora and syn- tactic parsers are available. In this paper, we present an experimental study where we ap- ply this approach to three languages that lack such resources: Tamil, Bengali and Malay- alam. We find an average quality difference of 6 to 20 absolute F-measure points vis-a- vis high-resource languages, which indicates that annotation projection alone is insufficient in low-resource scenarios. Based on these re- sults, we explore the possibility of using an- notation projection as a starting point for in- expensive data curation involving both experts and non-experts. We give an outline of what such a process may look like and present an initial study to discuss its potential and chal- lenges. 1 Introduction Creating syntactically and semantically annotated NLP resources for low-resource languages is known to be immensely costly. For instance, the Proposi- tion Bank (Palmer et al., 2005) was created by an- notating predicate-argument structures in the Penn Treebank (Marcus et al., 1993) with shallow seman- tic labels: frame labels for verbal predicates and role labels for arguments. Similarly, the SALSA (Bur- chardt et al., 2006) resource added FrameNet-style annotations to the TIGER Treebank (Brants et al., 2002), the Chinese Propbank (Xue, 2008) is built on the Chinese Treebank (Xue et al., 2005), and My father bought a house my father a house bought எ� த�ைத ஒ� ��ைட வா��னா� A0 A0 buy.01 A1 A1 buy.01 Figure 1: Annotation projection on a pair of very simple sen- tences. English Propbank frame (buy.01) and role (A0, A1) labels are projected onto aligned Tamil words. Furthermore, the typed dependencies between the words “my father” and “a house” (dotted lines) are projected onto their Tamil equivalents. so forth. Since each such layer of annotation typ- ically requires years of manual work, the accumu- lated costs can be prohibitive for low-resource lan- guages. Recent work on annotation projection offers a way to inexpensively label a target language corpus with linguistic annotation (Pad´o and Lapata, 2009). This only requires a word-aligned parallel corpus of labeled English sentences and their translations in the target language. English labels are then auto- matically projected onto the aligned target language words. Refer to Figure 1 for an example. Low-resource languages. However, previous work that investigated Propbank annotation projection has focused only on languages for which treebanks - and therefore syntactic parsers - already exist. Since syn- tactic information is typically used to increase pro- jection accuracy (Pad´o and Lapata, 2009; Akbik et al., 2015), we must expect this …

Content truncated to 500 words