There is a growing interest in conversation agents and robots which conduct attentive listening. However, the current systems always generate the same or limited forms of backchannels every time, giving a monotonous impression. This study investigates the generation of a variety of backchannel forms appropriate for the dialogue context, using the corpus of counseling dialogue. At first, we annotate all acceptable backchannel form categories considering the permissible variation in backchannels. Second, we analyze how the morphological form of backchannels relates to linguistic features of the preceding utterance such as the utterance boundary type and the linguistic complexity. Based on this analysis, we conduct machine learning to predict backchannel form from the linguistic and prosodic features of the preceding context. This model outperformed a baseline which always outputs the same form of backchannels and another baseline which randomly generates backchannels. Finally, subjective evaluations by human listeners show that the proposed method generates backchannels more naturally and gives a feeling of understanding and empathy.