For a successful automated negotiation, a vital issue is how well the agent can learn the latent preferences of opponents. Opponents however in most practical cases would be unwilling to reveal their true preferences for exploitation reasons. Existing approaches tend to resolve this issue by learning opponents through their observations during negotiation. While useful, it is hard because of the indirect way the target function can be observed as well as the limited amount of experience available to learn from. This situation becomes even worse when it comes to negotiation problems with large outcome space. In this work, a new model is proposed in which the agents can not only negotiate with others, but also provide information (e.g., labels) about whether an offer is accepted or rejected by a specific agent. In particular, we consider that there is a crowd of agents that can present labels on offers for certain payment; moreover, the collected labels are assumed to be noisy, due to the lack of expert knowledge and/or the prevalence of spammers, etc. Therefore to respond to the challenges, we introduce a novel negotiation approach that (1) adaptively sets the aspiration level on the basis of estimated opponent concession; (2) assigns labeling tasks to the crowd using online primal-dual techniques, such that the overall budget can be both minimized with sufficiently low errors; (3) decides, at every stage of the negotiation, the best possible offer to be proposed.