Training and Prediction Data Discrepancies: Challenges of Text Classification with Noisy, Historical Data

TL;DR

This paper investigates how noisy, historical datasets impact the evaluation of text classification models and explores the effectiveness of training on such data for future, cleaner inputs.

Contribution

It provides an analysis of the challenges posed by noisy, historical data in text classification and evaluates the utility of training on such data for future predictions.

Findings

Training on noisy data can still produce effective models for cleaner future data.

Performance metrics on historical data may not accurately reflect future model performance.

Dirty training datasets can be useful despite their noise levels.

Abstract

Industry datasets used for text classification are rarely created for that purpose. In most cases, the data and target predictions are a by-product of accumulated historical data, typically fraught with noise, present in both the text-based document, as well as in the targeted labels. In this work, we address the question of how well performance metrics computed on noisy, historical data reflect the performance on the intended future machine learning model input. The results demonstrate the utility of dirty training datasets used to build prediction models for cleaner (and different) prediction inputs.