From Stanford researchers AI Open-Source Diffusion-LM: a new, controllable language model based on continuous broadcasts, which enables new forms of complex, fine-grained control tasks

Language models often behave in unprecedented ways. Additionally, natural language generation continues to face significant challenges in controlling the behavior of language models without relearning. Gaining control over simple sentence properties (like sentiment) has led to progress in this area. However, progress in complex and precise checks like syntactic structure has been minimal to non-existent. Academics at Stanford University have developed an open-source broadcast language model to solve this problem. This language model uses a plug-and-play control strategy, where the language model is fixed, and the generated text is controlled by a third-party classifier that judges how well an output matches expected parameters. Many elements of the desired output, such as necessary parts of speech, syntax tree, or sentence length, are controllable by the user.

When evaluated on five controlled text creation tasks, this non-autoregressive generative language model outperforms previous approaches such as GPT-3. Several existing generative language models are autoregressive. They can anticipate the word that will come next in a sequence, append it to the existing sequence, and then use the modified sequence as input for future predictions. These models can produce language identical to that written by humans, and they can produce text to address a variety of issues, from interactive conversation to question-and-answer scenarios. However, the user has little or no control over the output generated in these legacy systems with respect to factors such as desired sentence length, structure, or sentiment. Fine-tuning the LM to allow it to accept additional control input has been suggested as a potential solution to this problem. However, this update can be computationally demanding and may not be generalized to accommodate many control settings. The alternative strategy uses a plug-and-play method, which holds the parameters of the LM at a fixed value and directs the generation using an external classifier that measures how closely the output aligns with the expected parameters. Piloting the autoregressive model nevertheless proved difficult. Instead of attempting to drive an autoregressive LM, Stanford researchers broke new ground for language development by creating a diffusion model.


Although these models have demonstrated excellent performance in computer vision and other continuous domains, their power has not been examined in domains such as text production. Diffusion-LM, in the opinion of the researchers, is the original diffusion model for text production. The model has undergone two adjustments. The first consists of an integration function that converts words into vectors in the continuous latent space of the diffusion model. After that, a “rounding” technique was created to convert these vectors back into discrete words. The model creates text by treating a random vector in latent space as a noisy interpretation of the output sentence embedding. The denoised embedding is then fed to an outside classifier at each step, which creates a gradient update of the embedding for the next step in the iteration. The rounding method converts the final embedding to text after completing the iterations. The team found that the Diffusion-LM is slower than other models for performance training and decoding. Researchers are excited about the additions the community can make. Therefore, they opened the code currently accessible on Github.

This Article is written as a summary article by Marktechpost Staff based on the paper 'Diffusion-LM Improves Controllable Text Generation'. All Credit For This Research Goes To Researchers on This Project. Checkout the paper, github and reference post.

Please Don't Forget To Join Our ML Subreddit

Khushboo Gupta is an intern consultant at MarktechPost. She is currently pursuing her B.Tech from Indian Institute of Technology (IIT), Goa. She is passionate about the fields of machine learning, natural language processing and web development. She likes to learn more about the technical field by participating in several challenges.

James G. Williams