In this repository you will find the code for our paper on joint segmenting, tagging, and parsing of incorporating languages (namely, we've been using Chukchi). While developed with these languages in mind, the code works for any data in the Universal Dependencies format.
We've decided to use the AllenNLP library for model creation because of its versatility and configurability, as well as because it uses pyTorch as its backend deep learning framework. To run training, create or choose a configuration file that encodes the model's hyperparameters such as the tasks you want the model to solve (any of the three tasks can be switched on and off), the number of parameters in the layers, and regularization. Then open the repository's directory in the Terminal and type:
allennlp train <CONFIG_PATH> -s <SERIALIZATION_DIR> --include-package modules.
Furthermore, you can use the training_example.ipynb notebook as reference.
Our approach is inspired by A Graph-based Model for Joint Chinese Word Segmentation and Dependency Parsing, a 2019 paper by Hang Yan, Xipeng Qiu, and Xuanjing Huang. Some of the core parts of the model's code were adapted from their open-sourced code for the paper, available on https://github.com/fastnlp/JointCwsParser. The resulting architecture is similar to the one implemented in the paper An improved neural network model for joint POS tagging and dependency parsing by Dat Quoc Nguyen and Karin Verspoor (the code is available on https://github.com/datquocnguyen/jPTDP, it's written in a deep learning framework called dynet).
The important feature that distinguishes our work from those related works is that it appears to be the first annotation of incorporation in a language using the Universal Dependencies. The problem of dependency annotation in polysynthetic languages is discussed in detail in the article, that can be consulted on https://www.aclweb.org/anthology/2020.udw-1.22/.
The dataset was developed on the corpus of spoken Chukchi in the Amguema variant, available on https://chuklang.ru/. Its volume is 65 texts, 6112 tokens, and 1004 sentences. The type-token ratio (TTR) is 50.26%. The original morphological annotation was revised, corrected and converted to the CoNLL-U format with the addition of POS tags. The dataset was enriched with syntactic dependency annotation.
Team members: Francis Tyers, Nikita Sykhrannov, Elizaveta Ezhergina, Karina Mishchenkova