Statistical Deep parsing for spanish
Luis Chiruzzo · LA Referencia (Red Federada de Repositorios Institucionales de Publicaciones Científicas) · 2020
This document presents the development of a statistical HPSG parser for Spanish. HPSG is a deep linguistic formalism that combines syntactic and semanticinformation in the same representation, and is capable of elegantly modelingmany linguistic phenomena. Our research consists in the following steps: design of the HPSG grammar, construction of the corpus, implementation of theparsing algorithms, and evaluation of the parsers performance. We created a simple yet powerful HPSG grammar for Spanish that modelsmorphosyntactic information of words, syntactic combinatorial valence, and semantic argument structures in its lexical entries. The grammar uses thirteenvery broad rules for attaching specifiers, complements, modifiers, clitics, relative clauses and punctuation symbols, and for modeling coordinations. In asimplification from standard HPSG, the only type of long range dependency wemodel is the relative clause that modifies a noun phrase, and we use semanticrole labeling as our semantic representation. We transformed the Spanish AnCora corpus using a semi-automatic processand analyzed it using our grammar implementation, creating a Spanish HPSGcorpus of 517,237 words in 17,328 sentences (all of AnCora). We implemented several statistical parsing algorithms and trained them overthis corpus. The implemented strategies are: a bottom-up baseline using bi-lexical comparisons or a multilayer perceptron; a CKY approach that uses theresults of a supertagger; and a top-down approach that encodes word sequencesusing a LSTM network. We evaluated the performance of the implemented parsers and compared them with each other and against other existing Spanish parsers. Our LSTM top-down approach seems to be the best performing parser over our test data, obtaining the highest scores (compared to our strategies and also to externalparsers) according to constituency metrics (87.57 unlabeled F1, 82.06 labeled F1), dependency metrics (91.32 UAS, 88.96 LAS), and SRL (87.68 unlabeled,80.66 labeled), but we must take in consideration that the comparison against the external parsers might be noisy due to the post-processing we needed to do in order to adapt them to our format. We also defined a set of metrics to evaluate the identification of some particular language phenomena, and the LSTM top-down parser out performed the baselines in almost all of these metrics as well.