Abstarct: Coreference resolution is one of the important steps in text processing and one of the basic strategies for improving the performance of text-baxsed systems such as information extraction, summarization, machine translation, and question answering. Coreference resolution deals with identifying all mentions or exxpressions that refer to the same real-world entity. Mentions may be pronouns, noun phrases and named entities. Numerous studies have focused on the issue of coreference resolution over the past four decades but its accuracy is still not acceptable for using in text comprehension applications, mention detection, candidate antecedent detection, and identifying coreference chains. One of the reasons for difficulty of this issue is requires various knowledge resources, including lexical knowledge, syntactic knowledge, world knowledge, discourse structure, and semantic knowledge. In fact, each type of coreference case requires one or more knowledge resources to be solved. Also, selecting the appropriate reference for existing mentions in the text is a function of the various rules that ensure the matching of gender, number, string structure, and other coreference parameters. In this thesis the problem is modeled in the form of effective parameters of the performance in order to improve precision, recall, and other coreference resolution goals using these features. For this purpose, using newest word embedding methods, recurrent neural network, and attention mechanism, span representation and their features were extracted. Then, mentions and candidate antecedent were extracted with high accuracy using knowledge resource representation, choosing appropriate knowledge and, entities and cluster level information. In addition, candidate antecedents were ranked and coreference chains were extracted using multi-criteria ranking baxsed on Perceptron and Kohonen neural networks. According to the simulation results, we achieved 83.9% and 97.3% Avg. F1 on the English CoNLL-2012 shared task and i2b2 medical dataset respectively. Also, the named entity recognition rate was improved 7.04% on English Gigaword dataset.