Towards AMR-BR: A SemBank for Brazilian Portuguese Language

Towards AMR-BR: A SemBank for Brazilian Portuguese Language Rafael Torres Anchiˆeta, Thiago Alexandre Salgueiro Pardo Interinstitutional Center for Computational Linguistics (NILC) Institute of Mathematical and Computer Sciences, University of S˜ao Paulo rta@usp.br, taspardo@icmc.usp.br Abstract We present in this paper an effort to build an AMR (Abstract Meaning Representation) annotated corpus (a semantic bank) for Brazilian Portuguese. AMR is a recent and prominent meaning representation with good acceptance and several applications in the Natural Language Processing area. Following what has been done for other languages, and using an alignment-based approach for annotation, we annotated the Little Prince book, which went into the public domain and explored some language-specific annotation issues. Keywords: Abstract Meaning Representation (AMR), corpus annotation, Portuguese language 1. Introduction Due to its wide applicability and potentialities, Natural Language Understanding (NLU) has gained interest and fostered research on themes of computational semantics (Oepen et al., 2016). According to Ovchinnikova (2012), NLU is the field of Natural Language Processing (NLP) that deals with machine reading comprehension. The objec- tive of an NLU system is to specify a computational model to interpret one or more input text fragments. The inter- pretation is usually carried out by a semantic parsing tech- nique, which maps natural language into a suitable meaning representation. A meaning representation is one of the most important components in semantic parsing. Its production is moti- vated by the hypothesis that semantics may be used to im- prove many natural language tasks, such as summarization, question answering, textual entailment, and machine trans- lation, among others. In this context, there are several avail- able meaning representations, as the traditional First-Order Logic (FOL), as detailed in Jurafsky and Martin (2009), se- mantic networks (Lehmann, 1992), Universal Networking Language (UNL) (Uchida et al., 1996), and, more recently, the Abstract Meaning Representation (AMR) (Banarescu et al., 2013). In particular, AMR got the attention of the scientific com- munity due to its relatively simpler structure, establishing the connections/relations among nodes/concepts, making them easy to read. Moreover, AMR structures are arguably easier to produce than traditional formal meaning represen- tations (Bos, 2016). According to Banarescu et al. (2013), AMR-annotated corpora are motivated by the need of providing to the NLP community datasets with embedded annotations re- lated to the traditional tasks of NLP, for instance, named en- tity recognition, semantic role labeling, word sense disam- biguation, and coreference. In this sense, the AMR anno- tation especially focuses on the predicate-argument struc- ture as defined in PropBank (Kingsbury and Palmer, 2002; Palmer et al., 2005). Another characteristic of AMR anno- tation is that words that do not significantly contribute to the meaning of a sentence (which are referred as “syntactic sugar” in the original paper) are left out of the annotation, as articles and the infinitive particle “to”. From the currently available datasets, many semantic parsers emerged (Flanigan et al., 2014; Wang et al., 2015; Peng et al., 2015; Goodman et al., 2016; Zhou et al., 2016; Damonte et al., 2017). Furthermore, with the available parsers, some applications …

Content truncated to 500 words