In this paper we use named entity recognition model on electronic messages with project information. We want to recognize following named entities SUBJECT, SKILL, DATE, QUANTITY, PERCENT, ORG, LOC in POSITION. For named entity recognition we first create word embeddings. For embeddings we use stacked embeddings which contains GloVe embeddings and contextual string embeddings. On embedded words we use bidirectional long short-term memory with conditional random fields for named entity recognition. We train our model with multiple optimization algorithms: gradient descent, gradient descent with weight decay, Adam optimizer, Adam optimizer with weight decay and Adam optimizer with parameters set to mimic gradient descent. Model was evaluated with $F_1$ score. Model with best score on test data was the one trained with gradient descent.
|
