Can Your Classifier Detect Boundaries? Adaptation of Artificial Text Detection Methods for the Real Or Fake Text Challenge

Anonymous

TL;DR: We study robustness of topological data analysis-based methods, perplexity-based methods and LM-based detector on artificial text boundary detection task in cross-domain and cross-model setting.

Abstract: Due to the rapid development of text generation models, people increasingly often encounter texts that may start out as written by a human but then continue as AI-generated. Detecting the boundary between human-written and machine-generated parts of such texts is a very challenging problem that has not received much attention in literature. We consider a number of different approaches for artificial text boundary detection, comparing predictors over features of different nature. We show that supervised fine-tuning of the RoBERTa model works well for in-domain detection of a single LLM but fails to generalize in important cross-domain and cross-generator settings, demonstrating a tendency to overfit to spurious features of the data. Then, we adapt perplexity-based approaches and propose novel algorithms based on features extracted from a frozen LLM's embeddings. We show that these approaches outperform the human accuracy level on an extremely hard Real or Fake Text benchmark. Analyzing the robustness of our approaches in cross-domain and cross-model settings, we discover important properties of the data that can hinder the performance of artificial text boundary detection algorithms.

Paper Type: long

Research Area: Interpretability and Analysis of Models for NLP

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: English