English | 2021 | ASIN : B09BDD12TY | 343 pages | PDF, AZW3, EPUB, MOBI | 25.5 MB
Before the advent of deep learning, traditional natural language processing (NLP) approaches had been widely used in tasks such as spam filtering, sentiment classification, and part of speech (POS) tagging. These classic approaches utilized statistical characteristics of sequences such as word count and co-occurrence, as well as simple linguistic features. However, the main disadvantage of these techniques was that they could not capture complex linguistic characteristics, such as context and intra-word dependencies.
Recent developments in neural networks and deep learning have given us powerful new tools to match human-level performance on NLP tasks and build products that deal with natural language. Deep learning for NLP is centered around the concept of word embeddings or vectors, also known as Word2vec, which encapsulate the meanings of words and phrases as dense vector representations. Word vectors, which are able to capture semantic information about words better than traditional one-hot representations, allow us to handle the temporal nature of language in an intuitive way when used in combination with a class of neural networks known as recurrent neural networks (RNNs). While RNNs can capture only local word dependencies, recently proposed vector-based operations for attention and alignment over word vector sequences allow neural networks to model global intra-word dependencies, including context