![]() We also tried a couple different approaches of passing the values directly in as parameters. We experimented with taking each row of columns and turning it into a sentence, and then passing that into the model as a parameter. We attempted this in a few different ways. Our goal then became to force the models to use the document provided and not the pretrained data in the model itself. The result being that we could use GPT directly if we did not need 100% factual or more likely, up to date data. Upon review of the models, it would appear that the answers were being genereated from GPT itself, and that either it had been trained on this WHO data or on data that contained similar infomration and that the data was simply old. Unfortunately when we examined the numerical output, the values did not corrospond to the current data in the visualizatoins. Surprisingly this generated excellent outputs. In our initial testing, we were simiply asking for a summary or asking specific questions of the data, and specifying the WHO URL for the intended data set. GPT is great at document summaries and great at language and conversation understanding, however most of this data is column data in the simplest form. The biggest technical challange with this solution was in using OpenAI services for the generation of the data summary. Technical deals of each individual solution are contained in the readme files in the folders Python and C#. Most of the calls to cognitive services are just a few cents per API call, so the cost of Azure would likely be derived more from the Azure infrastructure to support integration between this summarization engine and the front end appliation. ![]() Once this solution is in production, the cost of all of these services are based on the number of calls to the system. Most of these services have a free tier, so development of this technology should be relatively cheap for customers. The Python solution is currently using the following technology.Īzure Cognitive Services - This is currently on free tier pricing.Īzure Function or Logic Apps - To execute the C#Īzure Storage - To facilitate data movementĪzure App Services - This could be used to host a front end web app. This will be further discussed in the technical details below.įor both solutions, cost for our customers was a consideration. ![]() There are certain limitations due to data size, however our data scientists did provide recommendations and sample code for extensions. These ML models will ingest data and allow for text summarization and user Q&A functionality regardless of industry. This use case not unique to nonprofit or the WHO. The output however is a fully accessible replacement to PowerBI visualizations and "Ask a question of the data" functionality for both visually impaired as well as mobile or remote scenarios. While our solutions would both require some modification by customers prior to implementation, this technological lift should be small. This extends further as users are also able to simply ask additional questions of the data.Īs this code is easily called by serverless services such as Azure Functions or Logic Apps, our code is very extensible and easily embedded behind web apps or chat bots or by referencing voice captured through telephony solutions. We focused on the “summarization engine” portion of the project and built solutions to record customer data requests, ingest data, summarize key points with GTP and read this back to the user.Īt a very high level this is a tool to replace visualizations for PowerBI with text summaries enabling visually impaired and mobile users. Our team built two different solutions, one based on Python and the other on C#. A stretch goal was to allow the users to ask questions of the data in a more ad-hoc manner and have the analysis read back. The goal of this challenge was to use OpenAI to provide text and audio summarizations of PowerBI or other visualizations, for visually impaired or mobile data users. The data in the WHDH is provided as primarily as reports with visualizations although there are additionally API endpoints to access the data directly. ![]() From collection to use the platform provides a world class experience leveraging innovative technology to address challenges and pain points. The platform brings together an ambitious product stack to deliver an end-to-end solution for WHO data processes. WHO's World Health Data Hub(WHDH) provides an interactive digital platform and trusted source for global health data, fulfilling WHO’s commitment to provide health data as a public good. WHO/WHDH Chart Data Text Summarization - Extracting Insights from Data for Visually Impaired ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |