Unsupervised Methods for Learning and Using Semantics of Natural Language

Teaching the computer to understand language is the major goal in the field of natural language processing. In this thesis we introduce computational methods that aim to extract language structure — e.g. grammar, semantics or syntax — from text, which provides the computer with information in order to understand language. During the last decades, scientific efforts and the increase of computational resources made it possible to come closer to the goal of understanding language. In order to extract language structure, many approaches train the computer on manually created resources. Most of these so-called supervised methods show high performance when applied to similar textual data. However, they perform inferior when operating on textual data, which are different to the one they are trained on. Whereas training the computer is essential to obtain reasonable structure from natural language, we want to avoid training the computer using manually created resources. In this thesis, we present so-called unsupervised methods, which are suited to learn patterns in order to extract structure from textual data directly. These patterns are learned with methods that extract the semantics (meanings) of words and phrases. In comparison to manually built knowledge bases, unsupervised methods are more flexible: they can extract structure from text of different languages or text domains (e.g. finance or medical texts), without requiring manually annotated structure. However, learning structure from text often faces sparsity issues. The reason for these phenomena is that in language many words occur only few times. If a word is seen only few times no precise information can be extracted from the text it occurs. Whereas sparsity issues cannot be solved completely, information about most words can be gained by using large amounts of data. In the first chapter, we briefly describe how computers can learn to understand language. Afterwards, we present the main contributions, list the publications this thesis is based on and give an overview of this thesis. Chapter 2 introduces the terminology used in this thesis and gives a background about natural language processing. Then, we characterize the linguistic theory on how humans understand language. Afterwards, we show how the underlying linguistic intuition can be operationalized for computers. Based on this operationalization, we introduce a formalism for representing words and their context. This formalism is used in the following chapters in order to compute similarities between words. In Chapter 3 we give a brief description of methods in the field of computational semantics, which are targeted to compute similarities between words. All these methods have in common that they extract a contextual representation for a word that is generated from text. Then, this representation is used to compute similarities between words. In addition, we also present examples of the word similarities that are computed with these methods. Segmenting text into its topically related units is intuitively performed by humans and helps to extract connections between words in text. We equip the computer with these abilities by introducing a text segmentation algorithm in Chapter 4. This algorithm is based on a statistical topic model, which learns to cluster words into topics solely on the basis of the text. Using the segmentation algorithm, we demonstrate the influence of the parameters provided by the topic model. In addition, our method yields state-of-the-art performances on two datasets. In order to represent the meaning of words, we use context information (e.g. neighboring words), which is utilized to compute similarities. Whereas we described methods for word similarity computations in Chapter 3, we introduce a generic symbolic framework in Chapter 5. As we follow a symbolic approach, we do not represent words using dense numeric vectors but we use symbols (e.g. neighboring words or syntactic dependency parses) directly. Such a representation is readable for humans and is preferred in sensitive applications like the medical domain, where the reason for decisions needs to be provided. This framework enables the processing of arbitrarily large data. Furthermore, it is able to compute the most similar words for all words within a text collection resulting in a distributional thesaurus. We show the influence of various parameters deployed in our framework and examine the impact of different corpora used for computing similarities. Performing computations based on various contextual representations, we obtain the best results when using syntactic dependencies between words within sentences. However, these syntactic dependencies are predicted using a supervised dependency parser, which is trained on language-dependent and human-annotated resources. To avoid such language-specific preprocessing for computing distributional thesauri, we investigate the replacement of language-dependent dependency parsers by language-independent unsupervised parsers in Chapter 6. Evaluating the syntactic dependencies from unsupervised and supervised parses against human-annotated resources reveals that the unsupervised methods are not capable to compete with the supervised ones. In this chapter we use the predicted structure of both types of parses as context representation in order to compute word similarities. Then, we evaluate the quality of the similarities, which provides an extrinsic evaluation setup for both unsupervised and supervised dependency parsers. In an evaluation on English text, similarities computed based on contextual representations generated with unsupervised parsers do not outperform the similarities computed with the context representation extracted from supervised parsers. However, we observe the best results when applying context retrieved by the unsupervised parser for computing distributional thesauri on German language. Furthermore, we demonstrate that our framework is capable to combine different context representations, as we obtain the best performance with a combination of both flavors of syntactic dependencies for both languages. Most languages are not composed of single-worded terms only, but also contain many multi-worded terms that form a unit, called multiword expressions. The identification of multiword expressions is particularly important for semantics, as e.g. the term New York has a different meaning than its single terms New or York. Whereas most research on semantics avoids handling these expressions, we target on the extraction of multiword expressions in Chapter 7. Most previously introduced methods rely on part-of-speech tags and apply a ranking function to rank term sequences according to their multiwordness. Here, we introduce a language-independent and knowledge-free ranking method that uses information from distributional thesauri. Performing evaluations on English and French textual data, our method achieves the best results in comparison to methods from the literature. In Chapter 8 we apply information from distributional thesauri as features for various applications. First, we introduce a general setting for tackling the out-of-vocabulary problem. This problem describes the inferior performance of supervised methods according to words that are not contained in the training data. We alleviate this issue by replacing these unseen words with the most similar ones that are known, extracted from a distributional thesaurus. Using a supervised part-of-speech tagging method, we show substantial improvements in the classification performance for out-of-vocabulary words based on German and English textual data. The second application introduces a system for replacing words within a sentence with a word of the same meaning. For this application, the information from a distributional thesaurus provides the highest-scoring features. In the last application, we introduce an algorithm that is capable to detect the different meanings of a word and groups them into coarse-grained categories, called supersenses. Generating features by means of supersenses and distributional thesauri yields an performance increase when plugged into a supervised system that recognized named entities (e.g. names, organizations or locations). Further directions for using distributional thesauri are presented in Chapter 9. First, we lay out a method, which is capable of incorporating background information (e.g. source of the text collection or sense information) into a distributional thesaurus. Furthermore, we describe an approach on building thesauri for different text domains (e.g. medical or finance domain) and how they can be combined to have a high coverage of domain-specific knowledge as well as a broad background for the open domain. In the last section we characterize yet another method, suited to enrich existing knowledge bases. All three directions might be further extensions, which induce further structure based on textual data. The last chapter gives a summary of this work: we demonstrate that without language-dependent knowledge, a computer can learn to extract useful structure from text by using computational semantics. Due to the unsupervised nature of the introduced methods, we are able to extract new structure from raw textual data. This is important especially for languages, for which less manually created resources are available as well as for special domains e.g. medical or finance. We have demonstrated that our methods achieve state-of-the-art performance. Furthermore, we have proven their impact by applying the extracted structure in three natural language processing tasks. We have also applied the methods to different languages and large amounts of data. Thus, we have not proposed methods, which are suited for extracting structure for a single language, but methods that are capable to explore structure for “language” in general.