{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/props-probabilistic-personalization-of-black","title":"PROPS: Probabilistic personalization of black-box sequence models","arxiv_id":"1903.02013","date":"2019-03-05","proceeding":null,"authors":["Michael Thomas Wojnowicz","Xuan Zhao"],"abstract":"We present PROPS, a lightweight transfer learning mechanism for sequential\ndata. PROPS learns probabilistic perturbations around the predictions of one or\nmore arbitrarily complex, pre-trained black box models (such as recurrent\nneural networks). The technique pins the black-box prediction functions to\n\"source nodes\" of a hidden Markov model (HMM), and uses the remaining nodes as\n\"perturbation nodes\" for learning customized perturbations around those\npredictions. In this paper, we describe the PROPS model, provide an algorithm\nfor online learning of its parameters, and demonstrate the consistency of this\nestimation. We also explore the utility of PROPS in the context of personalized\nlanguage modeling. In particular, we construct a baseline language model by\ntraining a LSTM on the entire Wikipedia corpus of 2.5 million articles (around\n6.6 billion words), and then use PROPS to provide lightweight customization\ninto a personalized language model of President Donald J. Trump's tweeting. We\nachieved good customization after only 2,000 additional words, and find that\nthe PROPS model, being fully probabilistic, provides insight into when\nPresident Trump's speech departs from generic patterns in the Wikipedia corpus.\nPython code (for both the PROPS training algorithm as well as experiment\nreproducibility) is available at\nhttps://github.com/cylance/perturbed-sequence-model.","url_abs":"http://arxiv.org/abs/1903.02013v1","url_pdf":"http://arxiv.org/pdf/1903.02013v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"props-probabilistic-personalization-of-black","repo_url":"https://github.com/cylance/perturbed-sequence-model","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"none","reach":null}],"tasks":[{"task_slug":"articles","task_name":"Articles"},{"task_slug":"language-modeling","task_name":"Language Modeling"},{"task_slug":"language-modelling","task_name":"Language Modelling"},{"task_slug":"transfer-learning","task_name":"Transfer Learning"}],"methods":[{"method_slug":"lstm","method_name":"LSTM"},{"method_slug":"sigmoid-activation","method_name":"Sigmoid Activation"},{"method_slug":"tanh-activation","method_name":"Tanh Activation"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}