The present work deals with automated emotion recognition in text-based negotiations. As such, a number of possibilities are considered before experiments are conducted using exemplary implementations of applicable methods. The foundation for the corresponding experiments is a given dataset generated by negotiations between two fictitious companies in an experimental setup. Each negotiation message in the dataset comes with values for valence and activation according to Russell's circumplex of affect, which are generated by Multidimensional Scaling. Derived from these two values, class labels for individual document instances (negotiation messages) are generated with respect to radius and location on the bipolar, two-dimensional space. Text analysis is conducted in four major phases based on the framework by Aggarwal and Zhai. Thus, essential preprocessing and document representation aspects are taken into account before, finally, learning methods are chosen. In terms of preprocessing, approaches concerning stopword removal, tokenization, stemming and Part-Of-Speech tagging are explored, while for representation purposes, Bag-Of-Words using Term Frequency/Inverse Document Frequency weighting - also in interaction with Part-Of-Speech tagging - is found to be a promising constellation. In total, 16 experiment settings are put together and applied in combination with supervised learning methods. Particularly, representative algorithms of decision tree, probability-based, Support Vector Machine and proximity-based classifiers are determined for subsequent experiments. Empirical exploration is conducted using the WEKA toolkit, where J48, Naive Bayes Multinomial, Sequential Minimal Optimization, and Instance-Based k Learner are the respective implementations of the classifier families mentioned above. For activities relating to Part-Of-Speech tagging, the Stanford Part-Of-Speech tagger is utilized. To summarize, experiments employing 10-fold cross-validation reveal that the probability-based and the Support Vector Machine approaches are capable of achieving performance measures above 50% in terms of accuracy, precision, recall and F-score, while decision tree and proximity-based variants settled at around 40% in the best case. However, this is still almost double the baseline value of 21.51%, the share of the most frequent class occurring in the training set. In particular, experiment settings considering unigrams and bigrams as features boosted performance of the two better performing learning methods, which delivered the best results when combined with stemming. This, though, turned out to be a general tendency unless a Part-Of-Speech adjusted dataset is used, which is prepared such that it only consists of nouns, verbs and adjectives. Furthermore, stopword removal is generally found to have a negative impact on classifier performance, as features with high discrimination power in terms of Information Gain are neglected. In contrast, the application of Porter's stemmer and bigrams leverage classification results regardless of the particular learning method employed.