Neural Modeling of Multi-Predicate Interactions for Japanese Predicate Argument Structure Analysis

Abstract

The performance of Japanese predicate argument structure (PAS) analysis has improved in recent years thanks to the joint modeling of interactions between multiple predicates. However, this approach relies heavily on syntactic information predicted by parsers, and suffers from error propagation. To remedy this problem, we introduce a model that uses grid-type recurrent neural networks. The proposed model automatically induces features sensitive to multi-predicate interactions from the word sequence information of a sentence. Experiments on the NAIST Text Corpus demonstrate that without syntactic information, our model outperforms previous syntax-dependent models.

Multi-Task Learning for

“…Predicate Argument Structure Analysis The first set of rows in Table 2 shows the results of previous models. Ouchi+ 17 is the model from the Multi-Seq model in (Ouchi et al, 2017). M&I 17 is the model in (Matsubayashi and Inui, 2017).…”

“…Iida et al (2006), Imamura et al (2009), and Sasano and Kurohashi (2011) also analyze Inter-zero, which is a difficult task because the whole document must be searched. Following existing research (Ouchi et al, 2015(Ouchi et al, , 2017Inui, 2017, 2018;Taira et al, 2008), we only focus on three categories where arguments and their predicate (event-noun) are in the same sentence. In addition, we exclude the Bunsetsu category from the PASA evaluation following Ouchi et al (2017) and Matsubayashi and Inui (2018).…”

“…Our single model is based on an end-to-end approach (Zhou and Xu, 2015;Ouchi et al, 2017;Matsubayashi and Inui, 2018). Additionally, we add new features.…”

“…The RNN layers are made up of L layers of stacked bi-directional GRU. Additionally, we apply the residual connections (He et al, 2016) following Ouchi et al (2017); Matsubayashi and Inui (2018). At each time step t, the hidden state h…”

Neural Adversarial Training for Semi-supervised Japanese Predicate-argument Structure Analysis

“…The annotated corpora consist of several tens of thousands sentences, and it is difficult to learn predicate-argument relations or selectional preferences from such small-scale corpora. The state-of-the-art models for Japanese PAS analysis achieve an accuracy of around 50% for zero pronouns (Ouchi et al, 2015;Shibata et al, 2016;Iida et al, 2016;Ouchi et al, 2017;Matsubayashi and Inui, 2017).…”