Evidence-linked NLP Early research prototype

Follow the request.
Find the evidence.

A complaint tells a story. The requested resolution can be harder to find. RemedyTrace is being developed to connect explicit requests to the words that support them.

Explore an example

Built around English financial complaints and human review.

From narrative to evidenceIllustrative example

Synthetic complaint excerpt

“Please refund the duplicate fee and send me a written explanation.”
01 · Monetary redress

“refund the duplicate fee”

02 · Information request

“send me a written explanation”

Manually prepared to show the intended output. This is not a live model prediction.

Developed for customer operations and complaint-review teams.

English-language research · Human judgment at the center

The subject is only
part of the story.

A single complaint may describe earlier conversations, several current requests, and conditional alternatives. Our research focuses on making those distinctions visible to a reviewer.

01 / REQUEST

What is being asked for?

Represent multiple explicit remedies in one narrative, from a refund or record correction to restored access or a request for information.

02 / EVIDENCE

Where does the text say it?

Pair each request with an exact source excerpt so a reviewer can inspect the language behind the interpretation.

03 / CONTEXT

What needs a closer look?

Account for past requests, conditions, and uncertainty. An unclear passage should remain open for human review.

Where we are today

Research first.
Claims backed by evidence.

RemedyTrace is an early research project originating in an NLP course. We are developing the dataset, evaluation workflow, and model components for requested-remedy analysis.

In place
A six-category remedy taxonomy, multilabel annotation tools, source-excerpt checks, and component tests using synthetic examples.
In development
Training and evaluation on the real requested-remedy task, followed by an integrated reviewer interface. Predictive performance is not yet established.
Intended role
Support a human reviewer. The project does not determine whether an allegation is true or which resolution an institution should provide. It is not yet a production service.

Help shape the research

How do you review requests?

We welcome conversations about complaint-review workflows and what makes evidence useful. Please use synthetic examples and avoid sending personal or confidential complaint data.