Deep Learning for Grounded Compositional Meaning

Great progress has been made in natural language processing thanks to many different algorithms, each often specific to one application. Most learning algorithms force language into simplified representations such as bag-of-words or fixed-sized windows or require human-designed features. I will introduce models based on recursive neural networks that can learn linguistically plausible representations of language and reason over knowledge bases. These methods jointly learn compositional features and grammatical sentence structure for parsing or phrase level sentiment predictions. They can also be used to represent the visual meaning of a sentence which can be used to find images based on query sentences or to describe images with a more complex description than single object names.

Besides the state-of-the-art performance, the models capture interesting phenomena in language such as compositionality. For instance, people easily see that the "with" phrase in "eating spaghetti with a spoon" specifies a way of eating whereas in "eating spaghetti with some pesto" it specifies the dish. I show that my model solves these prepositional attachment problems well thanks to its distributed representations. In sentiment analysis, a new tensor-based recursive model learns different types of high level negation and how they can change the meaning of longer phrases with many positive words. They also learn that when contrastive conjunctions such as "but" are used the sentiment of the phrases following them usually dominates.