Generating High Quality Proposition Banks for Multilingual Semantic Role Labeling

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 397–407, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Generating High Quality Proposition Banks for Multilingual Semantic Role Labeling Alan Akbik∗ Technische Universit¨at Berlin, Germany alan.akbik@tu-berlin.de Laura Chiticariu Marina Danilevsky Yunyao Li Shivakumar Vaithyanathan IBM Research - AlmadenHuaiyu Zhu 650 Harry Road, San Jose, CA 95120, USA {chiti,mdanile,yunyaoli,vaithyan,huaiyu}@us.ibm.com Abstract Semantic role labeling (SRL) is crucial to natural language understanding as it identi- fies the predicate-argument structure in text with semantic labels. Unfortunately, re- sources required to construct SRL models are expensive to obtain and simply do not exist for most languages. In this paper, we present a two-stage method to enable the construction of SRL models for resource- poor languages by exploiting monolingual SRL and multilingual parallel data. Exper- imental results show that our method out- performs existing methods. We use our method to generate Proposition Banks with high to reasonable quality for 7 languages in three language families and release these resources to the research community. 1 Introduction Semantic role labeling (SRL) is the task of automat- ically labeling predicates and arguments in a sen- tence with shallow semantic labels. This level of analysis provides a more stable semantic representa- tion across syntactically different sentences, thereby enabling a range of NLP tasks such as information extraction and question answering (Shen and Lap- ata, 2007; Maqsud et al., 2014). Projects such as the Proposition Bank (PropBank) (Palmer et al., 2005) spent considerable effort to annotate corpora with semantic labels, in turn enabling supervised learn- ing of statistical SRL parsers for English. Unfor- ∗This work was conducted at IBM. tunately, due to the high costs of manual annota- tion, comparable SRL resources do not exist for most other languages, with few exceptions (Hajiˇc et al., 2009; Erk et al., 2003; Zaghouani et al., 2010; Vaidya et al., 2011). As a cost-effective alternative to manual annota- tion, previous work has investigated the direct pro- jection of semantic labels from a resource rich lan- guage (English) to a resource poor target language (TL) in parallel corpora (Pado, 2007; Van der Plas et al., 2011). The underlying assumption is that orig- inal and translated sentences in parallel corpora are semantically broadly equivalent. Hence, if English sentences of a parallel corpus are automatically la- beled using an SRL system, these labels can be pro- jected onto aligned words in the TL corpus, thereby automatically labeling the TL corpus with seman- tic labels. This way, PropBank-like resources can automatically be created that enable the training of statistical SRL systems for new TLs. However, as noted in previous work (Pado, 2007; Van der Plas et al., 2011), aligned sentences in par- allel corpora often exibit issues such as translation We need to hold people responsible A0 need.01 A1 A0 hold.01 A1 A2 Il faut qu' il y des ait responsables need.01 A1 it needs exist those responsible that there exist.01 A1 need.01 …

Content truncated to 500 words