OpenAI has developed a new model to study the alignment problem of machine learning. This model can summarize books of any length by creating summaries of each chapter. Yes, you heard it right; OpenAI’s new machine learning model can summarize the entire book.
The proposed machine learning model summarizes a small part of the book and then summarizes these summaries to obtain a higher-level overview. This research has been done as an empirical study on scaling correspondence problems which is usually tricky for AI algorithms because they require complex input text or numbers that have not yet been trained.
The research team have fine-tuned the GPT-3 language model to generate an abstract for a book, with even some of its qualities equaling those found in manually written works. When performing the book summary of the BookSum data set, the proposed model performs really good. The model by OpenAI can also be applied with other models, such as the zero-sample question and answer model, to achieve better book question and answer results.
Open AI is trying to solve the alignment problem of machine learning. The most critical challenge before the researchers can safely deploy general artificial intelligence with any degree of reliability will be ensuring that these models behave precisely how you want them to. This problem is called a correspondence problem.
It is hard to evaluate the output of a machine learning model because humans can’t always tell if it’s correct or not. So OpenAI researchers want to create scalable versions that summarize large amounts when teaching them new topics, but they’ll need more data first.
The OpenAI team combined human feedback and recursive task classification to create an effective machine learning model for summarizing books. They found that large-scale pre-training models are not very good at this type of summary because it requires judging the entire work without having had enough time to read every page, which humans need several hours so that they can get through all their information quickly.
This is how OpenAI researchers solved the problem of summarizing long texts. They used a method called “recursive task decomposition” which breaks down complex and difficult tasks into simple actions within programs like Brevabet or Summarizer Demonstrator to make them easier for humans who need more time evaluate models’ abstracts while still being able to summarize books without limits by context length in converter language.
This study is essential for improving the ability of humans to evaluate models. As model complexity increases, it is difficult for humans to assess model outputs. This may lead them to negative consequences if they do not have better evaluations capabilities as a result. Future research will need more complex tasks and an understanding on how we should interact with these computers because there are still many unknowns about AI.
The proposed solution will allow humans to obtain the assistance of models for evaluating machine model output. By generating summaries on individual chapters in the book, It can save people’s time reading through text themselves and also know if they need more information or not before delving into an entire chapter’s worth.
BookSum dataset: https://github.com/salesforce/booksum
Other Source: https://www.ithome.com.tw/news/146876