Combining ELECTRA and Adaptive Graph Encoding for Frame Identification

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1671–1679 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1671 Combining ELECTRA and Adaptive Graph Encoding for Frame Identification Fabio Tamburini FICLIT-University of Bologna, Italy fabio.tamburini@unibo.it Abstract This paper presents contributions in two directions: first we propose a new system for Frame Identification (FI), based on pre-trained text encoders trained discriminatively and graphs embedding, producing state of the art performance and, second, we take in consideration all the extremely different procedures used to evaluate systems for this task performing a complete evaluation over two benchmarks and all possible splits and cleaning procedures used in the FI literature. Keywords: frame semantics, frame identification, evaluation 1. Introduction Frame Semantics (Fillmore, 1982; Fillmore and Baker, 2010) is one of the more successful semantic theory and FrameNet1 (Fillmore et al., 2003) is an invaluable resource both for linguistic and computational analy- ses. A semantic frame represents an event or scenario, and possesses frame elements that participate in the event/scenario with different roles. On the computational side, considering the Frame Se- mantic Parsing task, studies in literature concentrated on three major subtasks (Das et al., 2014): • target identification, involving the selection of a word token, or word token sequences (the target predicate), evoking some frame in a given sen- tence. In frame semantics, verbs, nouns, adjec- tives, and even prepositions can evoke frames un- der certain conditions; • frame identification aims at finding the exact frame evoked by a target predicate in a given sen- tence, a sort of word-sense disambiguation task using frames as senses; • argument identification (a.k.a. semantic role labeling) is the task of identifying words and phrases as the appropriate arguments of the target predicate, the so called Frame Elements, and their specific roles. See Figure 1 for some examples and a more detailed explanation of the three tasks and a brief description of FrameNet frame-to-frame relation structure. In this paper, we focus our attention on the Frame Iden- tification (FI) task. There is a long series of studies in literature for solving the FI task (Das and Smith, 2011; Das et al., 2014; Hermann et al., 2014; Swayamdipta et al., 2017; Hartmann et al., 2017; Yang and Mitchell, 2017; Botschen et al., 2018; Peng et al., 2018; Sikos and Padó, 2019; Popov and Sikos, 2019; Tan and Na, 1https://framenet.icsi.berkeley.edu/ fndrupal/ 2019; Chen et al., 2021; Jiang and Riloff, 2021; Su et al., 2021), both using traditional techniques and various neural network approaches. Examining in detail the most recent literature present- ing state of the art systems based on Deep Neural Net- works, we found four main studies. Tan and Na (2019) proposed a very simple method based on transformers (Vaswani et al., 2017; Devlin et al., 2019) and positional attention obtaining good per- formance results. The solution by Chen et al. (2021) consists of four modules, a Bidirectional-LSTM encoder module for encoding the input sentence and three decoder mod- ules, for …

Content truncated to 500 words