We investigate the Perceptron HMM algorithm, an instance of the averaged perceptron approach, which incorporates discriminative training into the traditional Hidden Markov Model (HMM) approach. We demonstrate the efficiency of the algorithm by applying it to the biomedical term recognition problem. We show that the Perceptron HMM overcomes the limited expressiveness of the traditional, generative HMMs by incorporating additional, potentially overlapping features. This simple and elegant learning method produces performance that is comparable to the current state-of-the-art, while using only straightforward features derived from the provided training data. Our experiments illustrate the relative value of competing techniques that employ more complex learning algorithms and semantic features constructed from external resources.