The main principle of the merging method was to use the ME pre-diction and the other predictions as backup. Meaning, if the ME system didn't assign any name class to a word other predictions are taken into consideration. This principle had an excep-tion in case of a prediction of the name class "time".
Combined system
Features with non-binary value are later trans-formed to binary features.Structural Features: A set of binary features: wis the first word in the article; w is the first word in the sentence; w is the last word in the sen-tence. Lexicon Features: w itself is used as a feature. This set contains one feature for each distinct
NER MEMM Features
The Maximum Entropy (ME) probabilistic mod-eling technique has proved to be well adapted to cases where the model includes a large number of features. As apposed to the HMM, a ME model treats each feature separately. It gives each feature a weight according to its impact on the name class prediction.
features which represent regular expres-sions; a feature which indicates whether the word appears within quotes; dictionary features for the current word and a window of ±2 words around it; features which indicate if the current word is in or a part of an expression in one of the pre-processed lists.
HMM architecture with multiple features per state
The disadvantage of this definition is that the features are considered as a union and not separately. The simple HMM only emits one symbol to each state. A Limitation of this model is the use of the inde-pendence assumption which is not always correct in the NER problem.
States are defined as a product of the set of pos-sible name classes and the set of possible Part of Speech tags. For example, there would be a state for PERSON + NOUN, PERSON + VERB etc. We also define special states for the beginning and end of a sentence. Overall we get a set of 212 states. The intuition for this state definition came from the fact that the syntactic structure of the sentence has great impact on the prediction of name classes. Defining the Part of Speech tags as part of the HMM states emphasizes the structure of the sentence through the transition probabili-ties.
HMM Architecture
Transition and emission probabilities were calculated using maximum likelihood esti-mators
HMM architecture training