A transition-based approach for AMR parsing using LSTM networks

We present a procedure for generating Abstract Meaning Representation (AMR) structures from English sentences based on a transition-based system. Our proposed solution makes use of Long Short Term Memory networks to learn the action sequence that needs to be applied on the sentence in order to obtain the AMR graph. The action set is an extension of the arc-standard dependency parser, with several actions being altered to support action labels - necessary in the AMR graph. We apply several pre-processing steps on the sentences, to handle special language constructs - such as named entities, dates or monetary values - as well as a series of post-processing actions to restore the appropriate AMR nodes.