An official website of the United States government
As text generated by large language models proliferates, it becomes vital to understand how humans engage with such text, and whether or not they are able to detect when the text they are reading did not originate with a human writer. Prior work on human detection of generated text focuses on the case where an entire passage is either human-written or machine-generated. In this paper, we study a more realistic setting where text begins as human-written and transitions to being generated by state-of-the-art neural language models. We show that, while annotators often struggle at this task, there is substantial variance in annotator skill and that given proper incentives, annotators can improve at this task over time. Furthermore, we conduct a detailed comparison study and analyze how a variety of variables (model size, decoding strategy, fine-tuning, prompt genre, etc.) affect human detection performance. Finally, we collect error annotations from our participants and use them to show that certain textual genres influence models to make different types of errors and that certain sentence-level features correlate highly with annotator selection. We release the RoFT dataset: a collection of over 21,000 human annotations paired with error classifications to encourage future work in human detection and evaluation of generated text.
more »
« less
Topological Analysis of Contradictions in Text
Automatically finding contradictions from text is a fundamental yet under-studied problem in natural language understanding and information retrieval. Recently, topology, a branch of mathematics concerned with the properties of geometric shapes, has been shown useful to understand semantics of text. This study presents a topological approach to enhancing deep learning models in detecting contradictions in text. In addition, in order to better understand contradictions, we propose a classification with six types of contradictions. Following that, the topologically enhanced models are evaluated with different contradictions types, as well as different text genres. Overall we have demonstrated the usefulness of topological features in finding contradictions, especially the more latent and more complex contradictions in text.
more »
« less
- Award ID(s):
- 1910696
- PAR ID:
- 10358350
- Date Published:
- Journal Name:
- Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’22)
- Page Range / eLocation ID:
- 2478 to 2483
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
As text generated by large language models proliferates, it becomes vital to understand how humans engage with such text, and whether or not they are able to detect when the text they are reading did not originate with a human writer. Prior work on human detection of generated text focuses on the case where an entire passage is either human-written or machine-generated. In this paper, we study a more realistic setting where text begins as human-written and transitions to being generated by state-of-the-art neural language models. We show that, while annotators often struggle at this task, there is substantial variance in annotator skill and that given proper incentives, annotators can improve at this task over time. Furthermore, we conduct a detailed comparison study and analyze how a variety of variables (model size, decoding strategy, fine-tuning, prompt genre, etc.) affect human detection performance. Finally, we collect error annotations from our participants and use them to show that certain textual genres influence models to make different types of errors and that certain sentence-level features correlate highly with annotator selection. We release the RoFT dataset: a collection of over 21,000 human annotations paired with error classifications to encourage future work in human detection and evaluation of generated text.more » « less
-
Privacy policies disclose how an organization collects and handles personal information. Recent work has made progress in leveraging natural language processing (NLP) to automate privacy policy analysis and extract data collection statements from different sentences, considered in isolation from each other. In this paper, we view and analyze, for the first time, the entire text of a privacy policy in an integrated way. In terms of methodology: (1) we define PoliGraph , a type of knowledge graph that captures statements in a privacy policy as relations between different parts of the text; and (2) we develop an NLP-based tool, PoliGraph-er , to automatically extract PoliGraph from the text. In addition, (3) we revisit the notion of ontologies, previously defined in heuristic ways, to capture subsumption relations between terms. We make a clear distinction between local and global ontologies to capture the context of individual privacy policies, application domains, and privacy laws. Using a public dataset for evaluation, we show that PoliGraph-er identifies 40% more collection statements than prior state-of-the-art, with 97% precision. In terms of applications, PoliGraph enables automated analysis of a corpus of privacy policies and allows us to: (1) reveal common patterns in the texts across different privacy policies, and (2) assess the correctness of the terms as defined within a privacy policy. We also apply PoliGraph to: (3) detect contradictions in a privacy policy, where we show false alarms by prior work, and (4) analyze the consistency of privacy policies and network traffic, where we identify significantly more clear disclosures than prior work.more » « less
-
Recent developments in shape reconstruction and comparison call for the use of many different types of topological descriptors (persistence diagrams, Euler characteristic functions, etc.). We establish a framework that allows for quantitative comparisons of topological descriptor types and therefore may be used as a tool in more rigorously justifying choices made in applications. We then use this framework to partially order a set of six common topological descriptor types. In particular, the resulting poset gives insight into the advantages of using verbose rather than concise topological descriptors. We then provide lower bounds on the size of sets of descriptors that are complete discrete invariants of simplicial complexes, both tight and worst case. This work sets up a rigorous theory that allows for future comparisons and analysis of topological descriptor types.more » « less
-
Abstract Linguistic experience plays a clear role in accounting for variability in sentence comprehension behavior across individuals and across sentence types. We aimed to understand how individual differences in reading experience predict reading behavior. Corpus analyses revealed the frequencies with which our experimental items appeared in written and spoken language. We hypothesized that reading experience should affect sentence comprehension most substantially for sentence types that individuals primarily encounter through written language. Readers with more text exposure were faster and more accurate readers overall, but they read sentence types biased to written language particularly faster than did readers with less text exposure. We see clear effects of text exposure on sentence comprehension in ways that allow explicit links between written and spoken corpus statistics and behavior. We discuss theoretical implications of effects of text exposure for experience-based approaches to sentence processing.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1145/3477495.3531881
