Counting What Deserves to be Counted for Graph Parsing

Anonymous

Abstract: Graph parsers rely on scoring every subgraphs for building a complete graph. In real syntactic parsing or semantic parsing, every types of subgraphs in terms of syntactic or semantic roles may generate quite unbalanced distribution, which seems not well captured by the current graph paring models. Thus we propose an enhanced model design to let the parser explicitly capture such kind of unbalanced distribution. In detail, we introduce Accumulative Operation-based Induction (AOI) attention mechanism to assign accumulative scores for words. AOI scorer successfully approximates word-level unbalanced distribution. With conceptually simple but general-purpose design, our proposed AOI attention enhancement indeed leads to better parsing performance on a wide range of datasets of different parsing tasks, which verifies the scalability and robustness of capturing diverse subgraph distribution.