Dynamic Ensemble of Diversified Encodings for Event Nugget Detection
Kai Ishikawa, Hiroya Takamura, Manabu Okumura · 2019
We propose Dynamic Ensemble method that selects appropriate encoding models for each input to handle a wide variety of event expressions of different types. The main contribution is (1) we focused on compatibility problem of encoding models and a huge variety of linguistic patterns that is difficult to cope with conventional ensemble method, (2) proposed a novel ensemble approach that dynamically selects appropriate encoding models for every input token, and (3) proved the effectiveness of the ensemble method by comparing with official evaluation results of NIST TAC KBP2016 event nugget track. Our proposed method achieved 37.26% in F1 score without syntactic nor semantic parser, and outperformed the score 35.24% of the best system by 2.02% point.