Question answering forums like Reddit have been quite effective in improving social interaction and disseminating useful information. Community members ask a variety of questions related to a subject which are answered by other community members. The answers are given ratings by other members. In this thesis we study the problem of learning to recognize good answers using the community ratings as supervision. We design an attentive clustering neural network architecture to discriminate good answers from bad answers to a question. Taking advantage of the problem setting where there are usually many answers to the question that need to be scored, we also develop a collective classification model which clusters similar answers together, and biases the learner so that the answers in the same cluster have similar scores. The proposed solution uses a wide convolutional neural network to learn the text representation and computes a normalized score based on the relationship between the question and the answer and the similarity of the answer to other answers in the same cluster. Empirical results demonstrate that our collective classification model outperforms the baseline models and achieves the state-of-the-art performance in multiple benchmark domains.