deanhoperobertson/Afrikaans-Named-Entity-Recognition

# Afrikaans Named Entity Recognition This repository explores different techiques used for Named Entity Recognition (NER) on the NCHLT Afrikaans Named Entity Annotated Corpus. The data can be found here: Afrikaans Ner Corpus and was curated by the South African Department of Science and Innnovation. The onyl statistcial models currently built on this data was by Roald Eiselen (Univerity of Potchefstroom). Reults can be found here:aclweb.org attaining an average F1 score of **75.86** with the hightest score achievied of **77**. ## Models ### Feature Engineered - Conditional Random Fields [w] (F1=**75.19**) - Conditional Random Fields [w-1] (F1=**76.19**) - Conditional Random Fields [w-2] (F1=**77.64**) - Conditional Random Fields [w-3] (F1=**76.83**) ### Neural Network - BiLSTM-CRF (FastText Wiki 300D 99k) (F1=**64.0**) - BiLSTM-CRF (FastText Common 300D 800k) (F1=**??**) - BiLSTM-CRF (Casing Features + FastText Wiki 300D) (F1=**??**) ### Word Embeddings: I used the FastText Afrikaans pre-trained word embeddings trained on Wikipedia found at: fasttext.cc. ### Results We managed to attain the same results (if not better) using a Conditional Random Field approach and add more *engineered features* than the orginal paper by Roald Eiselen. ### Error Analysis**: The Fasttext word embedding used in the BiLSTM-CRF is missing 39% of the words use in the NCHLT Afrikaans NER Corpus, leaving a large portion of the words unaccounted for.This is not ideal and i noticed that this is was problem was also due to the inconsistent preparation of the data. Hypenated words such as "Suid-Afrika", should be seperated into 3 seperate tokens. After cleaning up this error, the missing word count dropped to 29%-34% depending on which embedding was used.