Understanding and generating language with abstract meaning representation

Abstract

Abstract Meaning Representation (AMR) is a semantic representation for natural language that encompasses annotations related to traditional tasks such as Named Entity Recognition (NER), Semantic Role Labeling (SRL), word sense disambiguation (WSD), and Coreference Resolution. AMR represents sentences as graphs, where nodes represent concepts and edges represent semantic relations between them. Sentences are represented as graphs and not trees because nodes can have multiple incoming edges, called reentrancies. This thesis investigates the impact of reentrancies for parsing (from text to AMR) and generation (from AMR to text). For the parsing task, we showed that it is possible to use techniques from tree parsing and adapt them to deal with reentrancies. To better analyze the quality of AMR parsers, we developed a set of fine-grained metrics and found that state-of-the-art parsers predict reentrancies poorly. Hence we provided a classification of linguistic phenomena causing reentrancies, categorized the type of errors parsers do with respect to reentrancies, and proved that correcting these errors can lead to significant improvements. For the generation task, we showed that neural encoders that have access to reentrancies outperform those who do not, demonstrating the importance of reentrancies also for generation. This thesis also discusses the problem of using AMR for languages other than English. Annotating new AMR datasets for other languages is an expensive process and requires defining annotation guidelines for each new language. It is therefore reasonable to ask whether we can share AMR annotations across languages. We provided evidence that AMR datasets for English can be successfully transferred to other languages: we trained parsers for Italian, Spanish, German, and Chinese to investigate the cross-linguality of AMR. We showed cases where translational divergences between languages pose a problem and cases where they do not. In summary, this thesis demonstrates the impact of reentrancies in AMR as well as providing insights on AMR for languages that do not yet have AMR datasets.