Introducing GLTR: Your Guardian Against Artificially Generated Text
Enter GLTR, a pioneering creation by the collaborative efforts of the MIT-IBM Watson AI lab and HarvardNLP. Designed to harness forensic analysis, GLTR serves as an unparalleled tool in the realm of text detection, expertly sniffing out automatically generated content. Its mechanism revolves around a meticulous examination of the text’s origins, gauging the likelihood of its creation by a language model.
GLTR’s ingenious approach revolves around visual scrutiny of the output churned by OpenAI’s GPT-2 117M language model. This allows GLTR to categorize each word within the text based on its propensity to have originated from the model. The outcome is an insightful ranking, with words being color-coded as green, yellow, red, or purple based on their likelihood of being model-generated. This visual breakdown serves as a direct, intuitive indicator of the text’s authenticity.
Adding to its arsenal, GLTR offers three histograms that synthesize information across the entire text. The first histogram elegantly presents the distribution of words across different categories. The second offers a glance into the ratio between the probabilities of the top-predicted word and its subsequent counterpart. Lastly, the third histogram casts light on the distribution across prediction entropies.
This compilation of histograms stands as corroborating evidence, elevating GLTR’s efficacy in discerning the artificial origins of a text. GLTR’s potential is profound – it serves as an unfailing sentinel against fabricated reviews, comments, and news articles generated by expansive language models. These models hold the capacity to craft content that seamlessly mingles with human-crafted text, rendering it virtually indistinguishable to the untrained eye.
Embracing accessibility, GLTR opens its doors through a live demo, inviting users to witness its prowess firsthand. For those inclined towards the technical realm, its source code resides on GitHub. In academia, GLTR’s impact resonates – its ACL 2019 demo track paper, an epitome of excellence, earned a nomination for the best demo.