metadata

sdk: gradio
python-version: 3.8

Coding-Free Automated Anecdotal Discourse Classification

Generalized versus Anecdotal Contexts

People may offer generalized or anecdotal contexts when making statements. Anecdotal context includes narratives, personal anecdotes, case histories, personal stories, personal assertions and testimonies. An example is when someone argues for legalization of marijuana based on what happened to their cousin who was jailed for possession. Generalized context, on the other hand, involves broad assertions about the world, including logical conclusions, reliable causal links and claims about entire categories of people or things. An example is saying marijuana should be legalize because it would improve the economy.

Generalized versus Anecdotal Statements in Language

You can think of anecdotal contexts as statements focused on more limited physical and temporal spaces. I argue in this doctoral dissertation that more or less anecdotal statements can be identified using the combination of four clause-level linguistic features:

Genericity: Whether the entity the clause is about is a generic category (e.g., humanity; generic) or specific instances of one (e.g., one's friends; specific).
Eventivity: (a.k.a. fundamental semantic aspect): Whether the verb complex describes a state (e.g., God is benevolent; stative) or an event (e.g., I went to Nebraska; dynamic).
Boundedness: Whether any events presented have a temporal boundary (e.g., I ate this morning; episodic) or do not (e.g., God loves us; static).
Habituality: Events can also be repetitive (e.g., I went to school there for years; habitual) or not (I stopped by; episodic). I have combined this feature with boundedness given their close relation. The most anecdotal content is theorized to involve specific entities with bounded, non-habitual events. But different mixtures of the dimensions can provide gradations of anecdotal focus.

AI-based Automatic Identification of Anecdotal Statements in Natural Text

This repository contains a coding-free tool for deriving labels for the above features from any text. It uses the transformer neural networks developed in the dissertation discussed above. Note that this tool is best used for analysis. This other tool would be more appropriate for demonstration purposes. Links and scripts for the corpus it was trained on can be found in this repository.