3 min read
Using ChatGPT to Navigate Microsoft Bing
Microsoft's Bing has received a turbocharged upgrade with the infusion of OpenAI's ChatGPT technology. This transformation has elevated the...
1 min read
Writing Team : Sep 14, 2023 12:43:23 PM
Microsoft Research recently unveiled a groundbreaking conversational question answering model that promises to redefine the future of web search.
This model, known as Generative Retrieval For Conversational Question Answering (GCoQA), has demonstrated impressive capabilities in answering questions swiftly and accurately, all while consuming significantly fewer resources than existing methods.
At the heart of GCoQA lies a revolutionary approach to ranking passages from web content.
Instead of relying on traditional keyword-based algorithms, GCoQA leverages generative retrieval techniques.
This innovative approach employs "identifier strings," which essentially serve as representations of passages within a document. These identifiers are derived from page titles (to identify the page's topic) and section titles (to identify specific passage content).
In practice, GCoQA functions much like using a webpage's title to understand its overall subject matter and using headings to discern the content of its sections.
This hierarchical approach helps organize information efficiently, first by page topic and then by individual passages within the page, as determined by section headings.
What sets GCoQA apart from other methods is its remarkable efficiency. The model consumes only a fraction of the memory resources required by current models, making it not only faster but also more practical for real-world applications.
Microsoft's researchers conducted extensive comparisons with other commonly used methods and found that GCoQA consistently outperformed them.
It addresses limitations and bottlenecks that have plagued existing techniques, offering a more convenient and efficient solution for conversational question answering.
Despite its promising potential, GCoQA does face certain limitations. For instance, its effectiveness outside of well-structured sources like Wikipedia remains uncertain, as many webpages lack meaningful section headings.
Additionally, handling ambiguous questions in real-world scenarios is a challenge that requires further investigation.
While GCoQA represents a significant leap forward in conversational question answering, there are challenges to address.
The research suggests two promising areas for further exploration: extending the use of generative retrieval in broader web search scenarios and integrating passage retrieval and answer prediction within a single, generative model.
The research paper "Generative Retrieval for Conversational Question Answering" is available on GitHub, offering valuable insights into this cutting-edge technology.
While it may not be available on mainstream search engines anytime soon, GCoQA serves as a glimpse into the future of web search, showcasing how researchers are harnessing generative models to revolutionize the way we access information online.
3 min read
Microsoft's Bing has received a turbocharged upgrade with the infusion of OpenAI's ChatGPT technology. This transformation has elevated the...
A Google patent was published on May 30, 2023, titled "Generating Query Variants Using A Trained Generative Model."
On June 10, 2023, Google introduced the highly anticipated Perspectives feature within its mobile search results, a development initially teased at...