Search Results
Improving Zero-Shot Text Classification with Graph-based Knowledge Representations
Insufficient training data is a key challenge for text classification. In particular, long-tail class distributions and emerging, new classes do not provide any training data for specific classes. Therefore, such a zeroshot setting must incorporate additional, external knowledge to enable transfer learning by connecting the external knowledge of previously unseen classes to texts. Recent zero-shot text classifier utilize only distributional semantics defined by large language models and based on class names or natural language descriptions. This implicit knowledge contains ambiguities, is not able to capture logical relations nor is it an efficient representation of factual knowledge. These drawbacks can be avoided by introducing explicit, external knowledge. Especially, knowledge graphs provide such explicit, unambiguous, and complementary, domain specific knowledge. Hence, this thesis explores graph-based knowledge as additional modality for zero-shot text classification. Besides a general investigation of this modality, the influence on the capabilities of dealing with domain shifts by including domain-specific knowledge is explored.
Understanding Class Representations: An Intrinsic Evaluation of Zero-Shot Text Classification
Frequently, Text Classification is limited by insufficient training data. This problem is addressed by Zero-Shot Classification through the inclusion of external class definitions and then exploiting the relations between classes seen during training and unseen classes (Zero-shot). However, it requires a class embedding space capable of accurately representing the semantic relatedness between classes. This work defines an intrinsic evaluation based on greater-than constraints to provide a better understanding of this relatedness. The results imply that textual embeddings are able to capture more semantics than Knowledge Graph embeddings, but combining both modalities yields the best performance.
Improving Language Model Predictions via Prompts Enriched with Knowledge Graphs
Despite advances in deep learning and knowledge graphs (KGs), using language models for natural language understanding and question answering remains a challenging task. Pre-trained language models (PLMs) have shown to be able to leverage contextual information, to complete cloze prompts, next sentence completion and question answering tasks in various domains. Unlike structured data querying in e.g. KGs, mapping an input question to data that may or may not be stored by the language model is not a simple task. Recent studies have highlighted the improvements that can be made to the quality of information retrieved from PLMs by adding auxiliary data to otherwise naive prompts. In this paper, we explore the effects of enriching prompts with additional contextual information leveraged from the Wikidata KG on language model performance. Specifically, we compare the performance of naive vs. KG-engineered cloze prompts for entity genre classification in the movie domain. Selecting a broad range of commonly available Wikidata properties, we show that enrichment of cloze-style prompts with Wikidata information can result in a significantly higher recall for the investigated BERT and RoBERTa large PLMs. However, it is also apparent that the optimum level of data enrichment differs between models.