Parsing of Lexicalised Linear Context-Free Rewriting Systems via Supertagging

Alex IvlievAlex Ivliev

Alex Ivliev
Parsing of Lexicalised Linear Context-Free Rewriting Systems via Supertagging
Project Thesis, TU Dresden, May 2020

Details
Bibtex

KurzfassungAbstract
Many natural language processing tasks require a syntactic representation of the input sentence. This can be achieved via dependency trees, which describe grammatical relations between words as directed edges. Non-projective trees, i.e. dependency trees with "crossing" edges, cannot be expressed by context-free grammars. Using context-sensitive grammars, however, is not feasable especially in real-time applications since parsing them is PSPACE-complete. Therefore, a number of grammar formalisms were proposed that can represent non-projective structures occuring in natural languages while still allowing the sentence to be parsed in polynomial time. These formalisms are summarized under the term mildly-context-sensitive grammars which includes for example head grammars, tree-adjoining and linear indexed grammars. Linear context-free rewriting systems (LCFRS) were introduced to represent this class of grammars in a uniform way. To further optimize the parsing of such grammars, a method called Supertagging has been developed in the context of tree-adjoining grammars. This method aims to reduce the number of rules considered during parsing by pre-selecting a small number of so-called elementary trees. In this work, we adapt this approach to lexicalised LCFRS grammars. As a way to pre-select rules during parsing we use BERT, which is a pre-trained neural network developed by Google AI. To train the network, we apply an embedding algorithm that assigns each rule to a small-dimensional vector of real numbers based on syntactic similarities between the rules. We further measure the impact of this method on the time and accuracy of parsing using established corpura.
Bemerkung: Note: Supervised by Heiko Vogler and Richard Mörbitz
Forschungsgruppe:Research Group: Wissensbasierte SystemeKnowledge-Based Systems

@misc{I2020,
  author = {Alex Ivliev},
  title  = {Parsing of Lexicalised Linear Context-Free Rewriting Systems via
            Supertagging},
  year   = {2020},
  month  = {May}
}