ausblenden:
Schlagwörter:
-
Zusammenfassung:
Knowledge extraction from text is a key task in natural language processing, which involves many sub-tasks, such as taxonomy induction, named entity recognition and typing, relation extraction, knowledge canonicalization and so on. By constructing structured knowledge from natural language text, knowledge extraction becomes a key asset for search engines, question answering and other downstream applications. However, current knowledge extraction methods mostly focus on prominent real-world entities with Wikipedia and mainstream news articles as sources. The constructed knowledge bases, therefore, lack information about long-tail domains, with fiction and fantasy as archetypes. Fiction and fantasy are core parts of our human culture, spanning from literature to movies, TV series, comics and video games. With thousands of fictional universes which have been created, knowledge from fictional domains are subject of search-engine queries - by fans as well as cultural analysts. Unlike the real-world domain, knowledge extraction on such specific domains like fiction and fantasy has to tackle several key challenges: - Training data: Sources for fictional domains mostly come from books and fan-built content, which is sparse and noisy, and contains difficult structures of texts, such as dialogues and quotes. Training data for key tasks such as taxonomy induction, named entity typing or relation extraction are also not available. - Domain characteristics and diversity: Fictional universes can be highly sophisticated, containing entities, social structures and sometimes languages that are completely different from the real world. State-of-the-art methods for knowledge extraction make assumptions on entity-class, subclass and entity-entity relations that are often invalid for fictional domains. With different genres of fictional domains, another requirement is to transfer models across domains. - Long fictional texts: While state-of-the-art models have limitations on the input sequence length, it is essential to develop methods that are able to deal with very long texts (e.g. entire books), to capture multiple contexts and leverage widely spread cues. This dissertation addresses the above challenges, by developing new methodologies that advance the state of the art on knowledge extraction in fictional domains. - The first contribution is a method, called TiFi, for constructing type systems (taxonomy induction) for fictional domains. By tapping noisy fan-built content from online communities such as Wikia, TiFi induces taxonomies through three main steps: category cleaning, edge cleaning and top-level construction. Exploiting a variety of features from the original input, TiFi is able to construct taxonomies for a diverse range of fictional domains with high precision. - The second contribution is a comprehensive approach, called ENTYFI, for named entity recognition and typing in long fictional texts. Built on 205 automatically induced high-quality type systems for popular fictional domains, ENTYFI exploits the overlap and reuse of these fictional domains on unseen texts. By combining different typing modules with a consolidation stage, ENTYFI is able to do fine-grained entity typing in long fictional texts with high precision and recall. - The third contribution is an end-to-end system, called KnowFi, for extracting relations between entities in very long texts such as entire books. KnowFi leverages background knowledge from 142 popular fictional domains to identify interesting relations and to collect distant training samples. KnowFi devises a similarity-based ranking technique to reduce false positives in training samples and to select potential text passages that contain seed pairs of entities. By training a hierarchical neural network for all relations, KnowFi is able to infer relations between entity pairs across long fictional texts, and achieves gains over the best prior methods for relation extraction.
