What is discourse coherence?

Language consists of sentences that are collocated, structured, and coherent together. This is known as discourse. The coherence refers to the relationship between sentences that makes up real discourses. An example of discourse is news articles where all the sentences are structured to bring you through a story.

What is consider a discourse coherent?

Real discourses has both local and global coherence. This means that documents that are created by concatenating sentences from different sources would not be consider discourse coherent. There are three ways to determine local coherent:

  1. Sentences are related to nearby sentences in a systematic ways

  2. Virtue of being “about” someone or something

  3. Topically coherent

The structured relationships that hold between sentences are known as coherence relations. An example would be REASON, where you state the action of an entity in one sentence followed by the reason of the action in the next sentence. In the second way, a coherent discourse should consists of salient entities and the discourse focuses on them. This is known as entity-based coherence and it tracks salient entities across a discourse. The most influential theory of entity-based coherence is Centering Theory, where we keep track of which entities in the discourse model are salient at any given point. Therefore, Centering Theory can pick up sentences that maintain the same salient and would classify those sentences to be more coherent. This theory can be implemented using the entity grid. In the last way, it makes sense that discourses should have nearby sentences that are generally around the same topic and uses similar vocabulary.

Global coherence refers to a standardised discourse structures. For example, academic papers tend to have many common sections such as Methodology and Conclusion.

Why does discouse coherence matters?

This is the task that can measures the quality of a document by assessing the coherence of sentences. This could be useful in summarisation or in mental health tasks where one of the features is detecting incoherent text.

23.1 Coherence Relations

What are the two most commonly used models for coherence relations?

Rhetorical Structure Theory (RST) and Penn Discourse TreeBank (PDTB)

Describe RST.

RST relations are defined between two spans of text known as nucleus and satellite. The nucleus is the unit that can be interpret independently and it’s more close to the message the writer’s wants to convey. The satellite is less central and it’s generally only interpretable with respect to the nucleus. Examples of RST coherence relations are:

  1. Reason – Nucleus is the action, satellite is the reason

  2. Elaboration – Nucleus is the situation, satellite is additional information supporting the situation

  3. Evidence – Nucleus is the situation, satellite is additional information supporting the situation

  4. Attribution – Nucleus is the event or information, satellite is the source of the information

  5. List – Multinucleus relation

RST is usually represented in a graph where the nucleus and satellite are connected by a directional edge. There exists hierarchical structure between coherence relations as shown below. The leaves in the discourse tree is known as discourse segments.

The RST Discourse TreeBank is the largest available discourse corpus. These coherence information is useful in summarisation where we can choose to keep nucleus since they tend to consists of more important information and can be interpret independently.

Describe PDTB.

PDTB is lexically grounded meaning that PDTB looks into discourse connectives instead of coherence relation, which are words that signal discourse relations, such as “because”, “although”, “since”, etc. The annotation involves mapping the discourse connectives with relevant coherence relations although not all coherence relations are marked by explicit connectives. The figure below showcase the four top-level classes and its subclasses within PDTB.



Data Scientist

Leave a Reply