Tried a CNN BLSTM CRF model

Format might be better in a .txt file: (It is in a public repo, should I use a private one?)

• NLP project notes: 2019-10-05 to 2019-10-18

Content:

Ian Ma, Oct 10 - Oct 18, 2019:

Repos: Model: https://github.com/guillaumegenthial/tf_ner Dataset: https://github.com/synalp/NER Paper: End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF (https://arxiv.org/abs/1603.01354)

------ Results: (Relatively short training, ~ 30 min, training accuracy: ~0.97, precision: ~0.89)

Labels: PER: Person ORG: Organization LOC: Location MISC: everything other than the first 3 classes

1. 002fcbe4-503c-400d-8cc0-77a395570ade.txt words: A military source who spoke to THISDAY, disclosed that the battle to uproot the insurgents from the town was led by the General Officer Commanding (GOC), 7 Division, General Lamidi Adeosun. preds: O O O O O O B-ORG O O O O O O O O O O O O O O O B-ORG I-ORG I-ORG I-ORG O O O B-PER I-PER

2. 0b4ea4e0-fade-4d35-b449-c36e265795d3.txt words: Jos—General Officer Commanding 3 Armoured Division of the Nigerian Army, Major General Jack Nwaogbo, has again re-assured Nigerians that the Boko Haram insurgency would soon be contained. preds: B-ORG I-ORG I-ORG O B-ORG I-ORG O O B-MISC I-MISC O O B-PER I-PER O O O B-PER O O B-ORG I-ORG O O O O O

3. 017326a6-80d1-44de-ad81-a72f47318254.txt words: They Chief of Army Staff who was represented by the General Officer Commanding (GOC), 1 Mechanised Division of the Nigerian Army, Kaduna, Maj-Gen. Adeniyi Oyebade however said that, all hands must be on deck to ensure proper training and upbringing of Nigerian children, believing that such will reduce to the barest minimal, security threats against the country. preds: O O O B-ORG I-ORG O O O O O B-ORG I-ORG I-ORG I-ORG O B-ORG I-ORG O O B-MISC I-MISC I-MISC O B-PER I-PER O O O O O O O O O O O O O O O O B-MISC O O O O O O O O O O O O O O O

------ Observations:

1. Works poorly when the organizations's name starts with a number. E.g., 1 Mechanised Division of the Nigerian Army (Result 3.)
2. This model only have lables: PER ORG LOC MISC (MISC is for everything other than the first 3 classes). Therefore, it cannot classify titles, ranks, etc. The dataset is CONLL2003. If we switch to our dataset, it might improve.