Representation learning, the process whereby representations are modelled from data, has recently become a central part of Natural Language Processing (NLP). Among the most widely used learned representations are word embeddings trained on large corpora of unannotated text, where the learned embeddings are treated as general representations that can be used across multiple NLP tasks. Despite their empirical successes, word embeddings learned entirely from data can only capture patterns of language usage from the particular linguistic domain of the training data. Linguistic knowledge, which does not vary among linguistic domains, can potentially be used to address this limitation. The vast sources of linguistic knowledge that are readily available nowadays can help train more general word embeddings (i.e. less affected by distance between linguistic domains) by providing them with such information as semantic relations, syntactic structure, word morphology, etc.
In this research, I investigate the different ways in which word embedding models capture and encode words’ semantic and contextual information. To this end, I propose two approaches to integrate linguistic knowledge into the statistical learning of word embeddings. The first approach is based on augmenting the training data for a well-known Skip-gram word embedding model, where synonym information is extracted from a lexical knowledge base and incorporated into the training data in the form of additional training examples. This data augmentation approach seeks to enforce synonym relations in the learned embeddings. The second approach exploits structural information in text by transforming every sentence in the data into its corresponding dependency parse trees and training an autoencoder to recover the original sentence. While learning a mapping from a dependency parse tree to its originating sentence, this novel Structure-to-Sequence (Struct2Seq) model produces word embeddings that contain information about a word’s structural context. Given that the combination of knowledge and statistical methods can often be unpredictable, a central focus of this thesis is on understanding the effects of incorporating linguistic knowledge into word representation learning. Through the use of intrinsic (geometric characteristics) and extrinsic (performance on downstream tasks) evaluation metrics, I aim to measure the specific influence that the injected knowledge can have on different aspects of the informational composition of word embeddings.