In this work we analyze the language signals of influence. We focus on their contribution to early prediction of the negotiation outcomes. Language expressions of influence show attitudes and emotional involvement, thus directly relate to subjectivity and sentiment analysis. The empirical evidence for our studies comes from machine learning experiments ran on the first part of negotiations. To evaluate the results, we extend the standard group of machine learning measures (accuracy, F-score, ROS) by introduction of positive and negative likelihoods.