Evolutionary Data Measures: Understanding the Difficulty of Text Classification Tasks

Abstract

Classification tasks are usually analysed and improved through new model architectures or hyperparameter optimisation but the underlying properties of datasets are discovered on an ad-hoc basis as errors occur. However, understanding the properties of the data is crucial in perfecting models. In this paper we analyse exactly which characteristics of a dataset best determine how difficult that dataset is for the task of text classification. We then propose an intuitive measure of difficulty for text classification datasets which is simple and fast to calculate. We show that this measure generalises to unseen data by comparing it to stateof-the-art datasets and results. This measure can be used to analyse the precise source of errors in a dataset and allows fast estimation of how difficult a dataset is to learn. We searched for this measure by training 12 classical and neural network based models on 78 real-world datasets, then use a genetic algorithm to discover the best measure of difficulty. Our difficulty-calculating code 1 and datasets 2 are publicly available.

Learning From Multiple Experts: Self-paced Knowledge Distillation for Long-Tailed Classification

“…N is the class probability for class i, and q i = 1 C denotes the uniform probability. Imbalance Absolute Deviation [4] is defined as the sum of aboslute distance between each long-tailed and uniform probability:…”

Not All Negatives are Equal: Label-Aware Contrastive Loss for Fine-grained Text Classification

“…For instance, "coarse" sentiment classification involves distinguishing negative and positive sentiments in text, and fine-grained sentiment classification involves further distinguishing the positive class into very positive and positive. This problem is challenging because the classes are semantically similar, which makes it difficult for the model to learn the labels (Collins et al, 2018).…”

“…This involves distinguishing between some closely confusable pairs of emotions, such as "sad" and "devastated", or "furious" and "annoyed". Fine-grained classification tasks are challenging precisely due to the presence of class interference amongst closely confusable classes (Collins et al, 2018;Zhao et al, 2017).…”

Quantifying the Task-Specific Information in Text-Based Classifications

“…Understanding the datasets In machine learning and NLP literature, several works studied the "difficulty" of datasets (Blache and Rauzy, 2011;Gupta et al, 2014;Collins et al, 2018;Jain et al, 2020), but they did not consider factoring out the impacts of shortcuts. D'Amour et al ( 2020) framed the shortcut learning issue as an underspecification problem: There is not enough information in training set to distinguish between spurious artifacts and the inductive biases (or rather, the linguistic knowledge).…”

“…Metrics for cross-task comparison. Consider reporting the performance on a unified scale of "task-specific informativeness", rather than relying on average model performance metrics (Collins et al, 2018). Designing metrics with grounds in linguistic knowledge is an interesting direction of future work.…”