Large generative language models such as GPT-2 are well-known for their ability to generate text as well as their utility in supervised downstream tasks via fine-tuning. Its prevalence on the web, however, is still not well understood - if we run GPT-2 detectors across the web, what will we find? Our work is twofold: firstly we demonstrate via human evaluation that classifiers trained to discriminate between human and machine-generated text emerge as unsupervised predictors of "page quality", able to detect low quality content without any training. This enables fast bootstrapping of quality indicators in a low-resource setting. Secondly, curious to understand the prevalence and nature of low quality pages in the wild, we conduct extensive qualitative and quantitative analysis over 500 million web articles, making this the largest-scale study ever conducted on the topic.

This paper seeks to develop a deeper understanding of the fundamental properties of neural text generations models.The study of artifacts that emerge in machine generated text as a result of modeling choices is a nascent research area.Previously, the extent and degree to which these artifacts surface in generated text has not been well studied.In the spirit of better understanding generative text models and their artifacts, we propose the new task of distinguishing which of several variants of a given model generated a piece of text, and we conduct an extensive suite of diagnostic tests to observe whether modeling choices (e.g., sampling methods, top-k probabilities, model architectures, etc.) leave detectable artifacts in the text they generate.Our key finding, which is backed by a rigorous set of experiments, is that such artifacts are present and that different modeling choices can be inferred by observing the generated text alone.This suggests that neural text generators may be more sensitive to various modeling choices than previously thought.

Dan Gillick, Cliff Brunk, Oriol Vinyals, Amarnag Subramanya. Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2016.

With the recent surge in the volume of search queries that explicitly or implicitly express users' geographical interests, to accurately infer users' locality preference becomes an increasingly important yet challenging issue. We study two click-based models of the distribution of such geographical interests by mining the user click stream data in the search engine logs, addressing three important issues in spatial Web search. First, search queries and documents can be classified by the models according to their spatial specificity. Second, the geographic center(s) of interests for queries and documents can be inferred. Finally, the model can be applied to generate relevance features for search ranking. We evaluated our proposals on a large dataset with about 10,000 unique queries sampled from the Yahoo! Search query logs, and about 450 million user clicks on 1.4 million unique Web pages over a six-months period. We report about 90% accuracy and about 3% false positive rate in identifying search queries with or without specific geographical interests, as well as statistically significant improvement in relevance ranking over a strong baseline.

In sparse data environments, greater classification accuracy can be achieved by learning several concept descriptions of the data and combining their classifications. Stochastic searching can be used to generate many concept descriptions (rule sets) for each class in the data. We use a tractable approximation to the optimal Bayesian method for combining classifications from such descriptions. The primary result of this paper is that multiple concept descriptions are particularly helpful in "flat" hypothesis spaces in which there are many equally good ways to grow a rule, each having similar gain. Another result is experimental evidence that learning multiple rule sets yields more accurate classifications than learning multiple rules for some domains.>