One application area where deep learning methods have shown remarkable success is natural language processing (NLP). The speech recognition and language understanding of virtual agents like Alexa and Siri, automatic translation between languages, and the sentiment analysis of texts and other applications have seen significant improvements over the last decade. The underlying neural network architectures for deep learning rely on inputs being presented as vectors; this is straightforward for images, instrument readings, and many other data sources. Natural language, however, is structured as a linear sequence of words, which is not directly amenable to translation into a vector. Vector space models list words and their occurrences in documents in a large matrix, making them in principle suitable for neural networks. Initial approaches listing all words against all documents result in huge structures, leading to dimensionality reduction methods like latent semantic analysis. Another problem with the vector representation of words is the need to distinguish between different meanings of the same word.
展开▼
