Oddballness: universal anomaly detection with language models

TL;DR

This paper introduces 'oddballness', a novel unsupervised metric for anomaly detection in text using language models, which measures how 'strange' a token is, outperforming likelihood-based methods in grammatical error detection.

Contribution

The paper proposes a new anomaly detection metric called oddballness that improves unsupervised grammatical error detection using language models.

Findings

oddballness outperforms likelihood-based methods in grammatical error detection

the method is fully unsupervised and applicable to various data sequences

demonstrates effectiveness in text anomaly detection tasks

Abstract

We present a new method to detect anomalies in texts (in general: in sequences of any data), using language models, in a totally unsupervised manner. The method considers probabilities (likelihoods) generated by a language model, but instead of focusing on low-likelihood tokens, it considers a new metric introduced in this paper: oddballness. Oddballness measures how ``strange'' a given token is according to the language model. We demonstrate in grammatical error detection tasks (a specific case of text anomaly detection) that oddballness is better than just considering low-likelihood events, if a totally unsupervised setup is assumed.