Diagnosing Semantic Properties in Distributional Representations of Word Meaning

When we communicate with each other, a large chunk of what we express is conveyed by the words we use. Computational models of language often rely on data-derived vector representations of words. Such representations are based on large text corpora and capture the many different contexts in which a word is used. For instance, the word lemon is represented by all instances of the word lemon in a large text corpus. It can be argued that the meaning of a word can be best characterized in terms of how the word is used. However, it is difficult to determine what aspects of word meaning the text corpora underlying the representations contain. Furthermore, it is unclear how the computational models used to create the word representations react to different usage examples in the data. Existing computational models of language perform well in some scenarios, but also make silly mistakes humans would never make. Their successes and failures are likely to be caused by what they know (or do not know) about the meaning of words. However, it is unclear what aspects of word meaning data-derived representations capture and what they do not capture. Do computational models know that lemons are yellow and round and have a sour taste? The research presented in this thesis approaches such open questions about language models by means of diagnostic experiments. The core of the research consists of a diagnostic dataset used to `diagnose’ different aspects of word meaning in distributional word representations derived from different types of language models. Word meaning is operationalized in terms of semantic properties (such as having a round shape, a yellow color, and a sour taste). The research is presented in four main parts: Part I contains an outline of existing research about the information captured by distributional representations of meaning. In addition, it presents two use-cases that illustrate the limitations of such representations and the need for a better understanding of what they represent when used to study words in specific texts or collections of texts. Part II presents a theoretical framework of what we can expect to be captured by distributional word representations as well as the methodological considerations underlying the design of the diagnostic dataset. Part III presents the diagnostic dataset. This includes an annotation task used to collect judgments about properties and concepts, an evaluation of the collected judgments, and an analysis of the dataset with respect to its suitability for diagnostic experiments. Part IV focuses on experimental work on the basis of the diagnostic dataset. It presents diagnostic experiments on traditional, static word representations paired with an analysis of corpora underlying such models. In addition, it includes experiments on more advanced, contextualized models. The results of the experiments together with an analysis of corpus data indicate that word representations are unlikely to capture specific semantic properties (e.g. the fact that lemons are yellow). Rather, they seem to reflect information about semantic categories (e.g. the fact that lemons are a type of citrus fruit). These findings confirm tendencies that have already become apparent in previous research from the perspective of model interpretability.