Generative AI language models such as OpenAI's GPT-2 impressively generate coherent and grammatical text. The control of the attributes of this text – such. Subject matter or mood – but requires a change in architecture or adaptation to specific data. For this reason, a team of scientists from Uber, Caltech, and the Hong Kong University of Science and Technology has developed the so-called Plug-and-Play Language Model (PPLM), which combines a pre-made language model with one or more attribute classifiers as a guide to Novel Text
Preliminary results in a preprinted paper show that PPLM is able to control a number of themes and mood styles, especially without compromising speech intelligibility and maintaining the flexibility that is required in any combination of differentiable models Text generation controls.
Their research builds on the published by Google and the University of Michigan late last year, in which an architecture was studied that generates sentences from a given sample and changes the mood, complexity, tension or even voice, while meaning of the original text. And it could impact work on Plato, Uber's conversation-developing and testing platform, released in July with connectors that integrate with existing machine learning and model optimization frameworks.
PPLM uses two or more small and easy-to-use AI. According to the researchers, construct attribute models: a word bag or a representation that describes the occurrence of words in a text snippet; and a linear discriminator or a method that characterizes two or more classes of objects or events. The optimization is done later, and the control is fine-grained. A strength parameter determines how strong the attribute influence should be.
"PPLM can be viewed as a generic machine for mirroring literary criticism models that can judge the value of texts that they read into author models … that can write something equally valuable. In this sense, the machinery of PPLM is quite general and allows the use of arbitrary differentiable attribute models, which allow beyond the examples given here, diverse and creative applications, "said researchers Uber Al Labs, Rosanne Liu, Sumanth Dathathri, Andrea Madotto, Piero Molino and Jason Yosinski added in a blog post that any number of attribute controllers can be combined with a state-of-the-art generative model. "With this research, we present … an instant configurable method for controlled speech generation that allows the flexible combination of a large, pre-learned language model with one or more … models."
In a series of experiments, a wordbox with attribute model and the subject "space" with different prefixes ("before", "an illustration", "the relationship between"), PPLM generated coherent paragraph-length samples. More impressively, the framework has successfully transitioned from "space" to topics such as "military" and "science," while cleverly handling prefixes that did not launch the sentence near the intended topic. In addition, PPLM steered the generation towards a mixed theme about "winter", "politics" and "kitchen". In a separate test using a sentiment discriminator trained on a popular open source dataset, PPLM controlled both positive and negative text generation.
PPLM is applicable to a number of scenarios, including scenarios with biased or offensive language. Previous studies have shown that counter-attacking language models can lead to racist outcomes. In a test with 10 triggers and evaluators to indicate the toxicity of 500 samples, the PPLM reduced the proportion of generated toxic speech from 63.6% to 4.6%.
Of course, PPLM can be used by malicious actors to create a more venomous language that researchers claim is a risk inherent in such general-purpose technologies. "We believe that overall this work will produce a positive net result by driving the discussion on model capabilities, properties, and potential solutions to the risks faced by current models," wrote Liu, Dathathri, Madotto, Molino, and Yosinski. "Recently, there has been considerable debate about ethical language models, both in terms of their potential to recapitulate problematic social prejudices, and to directly abuse them for social harm."
The Uber research team has published the code accompanying their article, a programming notebook in Google Colab, and an online demo developed in collaboration with Hugging Face. The publicly available demo contains a dozen attribute models.