23.2 Discourse Structure Parsing
What is discourse parsing?
This is the task of automatically determining the coherence between sentences.
Describe RST parsing and the shift-reduce parser method for building RST coherence structure.
The first step is to perform EDU segmentation, where we extract the start and end of each EDU. The shift-reduce parser uses a stack and a queue and produces the structure by taking a series of actions on the states. An example of this process is included below. The actions include:
shift – pushes the first EDU in the queue onto the stack creating a single-node subtree
reduce(l, d) – merges the top two subtrees on the stack, where l is the coherence relation and d is the nuclearity direction
pop – remove final tree from stack
The method is implemented using an encoder-decoder architecture, where the encoder is use to encode the input words and EDUs using a hierarchical biLSTM. The encoded representation is then feed into the decoder (feed forward) that outputs an action.
What is shallow discourse parsing?
Shallow discourse parsing is used to describe PDTB discourse parsing as it only involves flat relationships between text spans rather than the full trees. There are four subtasks:
Find the discourse connectives – connectives disambiguation
Find the two spans for each connective – can used the same models as finding RST EDUs
Label the relationship between these spans
Assign a relation between every adjacent pair of sentences
Subtask 1 – 3 is used for explicit connectives whereas subtask 4 is used for implicit connectives. A simple effective algorithm for subtask 4 is to represent the two spans using BERT contextualised embeddings and feeding the output of the last hidden layer (from CLS token) into a single tanh layer with softmax for classification.