This paper develops a novel Bayesian natural language processing framework to recover the division of labour in inventor teams directly from patent texts. The method estimates the probability that each inventor contributed specific knowledge components, informed by their prior collaborations. Applying this framework to approximately 4 million patents, I validate the approach and introduce a novel measure of individual contributions within teams. The results show that as inventors gain experience, they collaborate on more patents per year. However, within projects, senior colleagues contribute relatively less than their junior collaborators. Importantly, concentrating contributions in the hands of lower-performing team members significantly reduces patent value, highlighting how the structure of collaboration shapes the social returns to innovation.