SAN FRANCISCO, Dec. 3, 2020 /PRNewswire/ -- Today, Searchable.ai releases KitanaQA, a suite of innovative tools to train question-answering models to respond more reliably and accurately when they encounter noisy inputs like misspellings, incorrect grammar and foreign words. These types of inputs can very easily render leading language models—including those developed by Google, Facebook and Microsoft—completely ineffective because even though they exhibit high accuracy on research benchmarks, these models are often brittle and unpredictable in the real world. As a result, a gap has emerged between cutting edge research and industry use cases.
Searchable.ai Labs is working to bridge this gap by training models that are more robust and resistant to the types of noise that occur in the wild. "Robustness is key to building trust in NLP-based technologies, and KitanaQA can help harden question answering models before they are ever deployed in production," says Searchable.ai Chief Scientist Aaron Sisto.
In order to provide researchers with a complete set of tools to accomplish this task, KitanaQA focuses on three main features: adversarial training, data augmentation and workflow automation. Together, these techniques enable language models to generalize to noisy user queries or reference text even if they have never encountered these exact inputs before. Much of this benefit arises from the multifaceted approach to training provided by KitanaQA. While adversarial training prepares the model to produce more reliable answers in the presence of small amounts of noise, data augmentation exposes the model to more extreme types of noise at the same time, helping the model to understand the true meaning behind a user's questions, no matter the specific instantiation.
KitanaQA makes these tools accessible to the broader research community by providing automated pipelines that can be integrated with many different models and proprietary datasets. The creators point out that no model is robust to every type of user input, even with these tools on hand. In fact, some questions are simply too vague or challenging to answer and a question answering system should inform the user of this, instead of trying to answer the question incorrectly. For many question answering applications involving PDFs, webpages or MS Office documents, KitanaQA can provide meaningful improvements to model performance and reach a level of robustness that is acceptable to critical enterprise applications.
Searchable is a new kind of Conversational Search engine for your work—leveraging patent-pending Neural Indexing technology to cut through the clutter and find the exact file, information, or answer you need, no matter where it's stored. The company is based out of San Francisco, CA. For more information, please visit http://searchable.ai or email [email protected].