Abstract
Our ability to combine simple constituents into more complex conceptual combinations is a fundamental aspect of cognition. Gradable adjectives (e.g., 'tall' and 'light') are a critical example of this process, as their meanings vary depending on the noun with which they are combined. For example, a dark diamond is less dark than dark charcoal. Here, we investigate how a neural network encodes the flexible nature of gradable adjectives in adjective-noun pairs, using the perceptual feature of brightness as a test case. We trained a neural network to predict human brightness ratings for unmodified nouns and adjective-noun pairs and assessed its ability to generalize to untrained combinations (e.g., 'light paint' vs. 'dark paint'). We also explored how this information is encoded. We found that flexible learning of gradable adjectives was possible, with neural networks first making predictions based on the adjective alone, and then modulating these with information from the noun later in learning. We also found that model outputs mimicked the kind of non-additive feature modulation present in human data. Our results have implications for understanding how semantic composition occurs and generate testable predictions for future work.