ULSAna: Universal Language Semantic Analyzer

Proceedings of Recent Advances in Natural Language Processing, pages 967–972, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_112 967 ULSAna: Universal Language Semantic Analyzer Ondˇrej Praˇz´ak NTIS Faculty of Applied Sciences University of West Bohemia Technick´a 8, 306 14 Plzeˇn ondfa@ntis.zcu.cz Miloslav Konop´ık Dpt. of Computer Science and Engineering Faculty of Applied Sciences University of West Bohemia Technick´a 8, 306 14 Plzeˇn konopik@kiv.zcu.cz Abstract We present a live cross-lingual system ca- pable of producing shallow semantic an- notations of natural language sentences for 51 languages at this time. The domain of the input sentences is in principle uncon- strained. The system uses single training data (in English) for all the languages. The resulting semantic annotations are there- fore consistent across different languages. We use CoNLL Semantic Role Labeling training data and Universal dependencies as the basis for the system. The system is publicly available and supports processing data in batches; therefore, it can be easily used by the community for research tasks. 1 Introduction In this work, we present a major outcome in our journey to build a system capable of producing se- mantic annotations for domain and language un- constrained natural language sentences. Currently, we rely on the Semantic Role Labeling (SRL) an- notation scheme (Gildea and Jurafsky, 2002). The SRL goal is to determine semantic relationships (semantic roles) of given predicates (see examples in Figure 1). Verbs, such as “believe” or “cook”, are natural predicates but certain nouns can be ac- cepted as predicates as well (see the third line in the example). The semantic roles are specific for each predicate; however, the meaning of the roles is mostly shared across predicates. The core roles are denoted by A0 (usually Agent), A1 (usually Patient) and A2. Additional roles are modifier ar- guments (AM-*), restriction arguments (R-*) and others. We selected SRL because we believe that the annotations are simple enough to be general- ized for different languages and target domains but (1) [He]A0 believes [in what he plays] A1 . (2) Can [you] A0 cook [the dinner] A1 ? (3) [The nation‘s] AM-LOC largest [pension]A1 fund, Figure 1: Three examples of shallow semantic an- notations: 1) and 2) are examples of verb predi- cates and 3) of a noun predicate. at the same time expressive enough to bring a use- ful insight into the sentence semantics. In order to be able to produce semantic anno- tations for more languages, we employ the cross- lingual SRL. Cross-lingual SRL takes the training data from a source language (usually English) and builds a language independent model that can be applied to target languages. The advantage of the cross-lingual SRL is that it ensures coherent anno- tation for all supported languages because it trains on single training data for all the languages. This does not apply to the monolingual SRL where the tagsets and annotation guidelines change with ev- ery training dataset. In our approach, we heavily depend on Univer- sal Dependencies (UD) (Nivre et al., 2016). They are the primary means to transfer …

Content truncated to 500 words