ljvmiranda921/ud-tagalog-spacy

Training a POS Tagger and Dependency Parser for a Low-Resource Language (Tagalog) # 🪐 spaCy Project: Training a POS Tagger and Dependency Parser for a Low-Resource Language (Tagalog) This project trains a part-of-speech tagger and dependency parser for a low-resource language such as Tagalog. We will be using the TRG and Ugnayan treebanks for this task. Since the number of sentences in each corpus is small, we'll need to evaluate our model using 10-fold cross validation. How to implement this split will be demonstrated in this project (`scripts/kfold.py`). The cross validation results can be seen below. ### Monolingual evaluation Consists of k-fold cross validation and inter-treebank evaluation. **TRG Treebank** | | TOKEN_ACC | POS_ACC | MORPH_ACC | TAG_ACC | DEP_UAS | DEP_LAS | |---------|-----------|---------|-----------|---------|---------|---------| | 10-fold | 1.000 | 0.843 | 0.749 | 0.833 | 0.846 | 0.554 | | Ugnayan | 0.997 | 0.563 | 0.364 | 0.538 | 0.472 | 0.240 | **Ugnayan Treebank** | | TOKEN_ACC | POS_ACC | MORPH_ACC | TAG_ACC | DEP_UAS | DEP_LAS | |---------|-----------|---------|-----------|---------|---------|---------| | 10-fold | 0.998 | 0.819 | 0.995 | 0.810 | 0.667 | 0.409 | | TRG | 1.000 | 0.789 | 0.424 | 0.779 | 0.793 | 0.572 | ### Cross-lingual evaluation Evaluating models trained from other typologically similar languages against the two `tl` treebanks. **TRG Treebank** | | TOKEN_ACC | POS_ACC | MORPH_ACC | TAG_ACC | DEP_UAS | DEP_LAS | |-----------|-----------|---------|-----------|---------|---------|---------| | id-gsd | 1.000 | 0.374 | 0.320 | 0.000 | 0.342 | 0.151 | | vi-vtb | 1.000 | 0.306 | 0.423 | 0.000 | 0.309 | 0.143 | | ro-rrt | 0.999 | 0.392 | 0.198 | 0.000 | 0.304 | 0.098 | | uk-iu | 1.000 | 0.185 | 0.177 | 0.000 | 0.539 | 0.188 | | ca-ancora | 0.999 | 0.284 | 0.057 | 0.015 | 0.261 | 0.081 | **Ugnayan Treebank** | …