End-to-end Semantic Role Labeling with Neural Transition-based Model

Authors: Hao Fei, Meishan Zhang, Bobo Li, Donghong Ji12803-12811

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on Co NLL09 and Universal Proposition Bank show that our final model can produce state-of-the-art performance, and meanwhile keeps highly efficient in decoding. We also conduct detailed experimental analysis for a deep understanding of our proposed model.
Researcher Affiliation Academia 1 Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University, Wuhan, China 2 Department of School of New Media and Communication, Tianjin University, Tianjin, China {hao.fei, boboli, dhji}@whu.edu.cn, mason.zms@gmail.com
Pseudocode No The paper describes the transition system and actions in prose and with an example table (Figure 2(a)), but does not include a formal pseudocode block or algorithm figure.
Open Source Code Yes Our codes are open at https://github.com/scofield7419/Transition SRL.
Open Datasets Yes We employ two dependency-based SRL benchmarks: Co NLL09 English, and Universal Proposition Bank (UPB) for other total seven languages.
Dataset Splits Yes Each dataset comes with its own train, develop and test sets.
Hardware Specification No No specific hardware details (like GPU models, CPU types, or memory specifications) used for running experiments are mentioned in the paper.
Software Dependencies No The paper mentions software like 'fasttext', 'ELMo', 'BERT', and 'XLNet', but does not provide specific version numbers for these or any other software dependencies. It only specifies 'base-cased-version' for BERT and 'base-version' for XLNet, which are model variants, not software versions.
Experiment Setup Yes The hidden sizes in Bi LSTM, Tree LSTM and Stack-LSTM are 200. We use the two layer version of Bi LSTM, Stack-LSTM. The dimension of transition states is 150 universally. We adopt the Adam optimizer with initial learning rate of 1e-5. ζ is 0.2, and beam size B is 32, according to our developing experiments. We train the model by mini-batch size in [16,32] with early-stop strategy.