Quantifying the Conceptual Combination Effect on Word Meanings

How do people understand concepts such as dog, aggressive dog, dog house or house dog? The meaning of a concept depends crucially on the concepts around it. While this hypothesis has existed for a long time, only recently it has become possible to test it based on neuroimaging and quantify it using computational modeling. In this paper, a neural network is trained with backpropagation to map attributebased semantic representations to fMRI images of subjects reading everyday sentences. Backpropagation is then extended to the attributes, demonstrating how word meanings change in different contexts. Across a large corpus of sentences, the new attributes are more similar to the attributes of other words in the sentence than they are to the original attributes, demonstrating that the meaning of the context is transferred to a degree to each word in the sentence. Such dynamic conceptual combination effects could be included in natural language processing systems to encode rich contextual embeddings to mirror human performance more accurately.