Strongly Incremental Constituency Parsing with Graph Neural Networks

Authors: Kaiyu Yang, Jia Deng

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our parser on Penn Treebank (PTB) and Chinese Treebank (CTB). On PTB, it outperforms existing parsers trained with only constituency trees; and it performs on par with state-of-the-art parsers that use dependency trees as additional training data. On CTB, our parser establishes a new state of the art.
Researcher Affiliation Academia Kaiyu Yang Princeton University kaiyuy@cs.princeton.edu Jia Deng Princeton University jiadeng@cs.princeton.edu
Pseudocode No The paper does not contain any explicit pseudocode blocks or algorithms labeled as such.
Open Source Code Yes Code is available at https://github.com/princeton-vl/ attach-juxtapose-parser.
Open Datasets Yes We evaluate our model for constituency parsing on two standard benchmarks: Chinese Treebank 5.1 (CTB) [46] and the Wall Street Journal part of Penn Treebank (PTB) [24].
Dataset Splits Yes PTB consists of 39,832 training examples; 1,700 validation examples; and 2,416 testing examples. Whereas CTB consists of 17,544/352/348 examples for training/validation/testing respectively.
Hardware Specification Yes The model is implemented in Py Torch [29] and takes 2 3 days to train on a single Nvidia Ge Force GTX 2080 Ti GPU. These experiments were run on machines with 2 CPUs, 16GB memory, and one GTX 2080 Ti GPU.
Software Dependencies No The paper mentions that the model is 'implemented in Py Torch [29]' but does not provide a specific version number for PyTorch or any other software dependency.
Experiment Setup Yes Model parameters are optimized using RMSProp [40] with a batch size of 32. We decrease the learning rate by a factor of 2 when the best validation F1 score plateaus. The hyperparameters for each model are in the supplementary materials.