For greater than three a long time, IBM Consulting® development teams and data scientists have worked with the United States Tennis Association (USTA) to deliver an exciting digital experience to US Open tennis fans.
Let's take a deep dive into this 12 months's innovations in two generative AI projects that leverage IBM's versatile family of enterprise-grade Granite™ foundational models, amongst others. We'll also have a look at how the team used IBM Watsonx Code Assistant™ to speed up code generation and improve productivity and collaboration.
We spoke with project leader Aaron Baughman, IBM Fellow, Master Inventor and IBM Quantum™ Ambassador, to learn the way IBM is driving learnings on the US Open.
Project 1: The Content Engine
Providing up-to-date coverage of the a whole bunch of matches in men's and girls's singles is a large undertaking. But by leveraging the hybrid cloud architecture with Red Hat® OpenShift® that IBM Consulting has built over years of working with the USTA, the event team can quickly create, test and implement recent automated workflows to deal with such challenges. This 12 months's Content Engine is one such workflow.
The content engine produces three principal outputs: bullet points with descriptive text before and after each individual match, spoken commentary and subtitles for match highlights, and multi-part match reports that provide descriptive summaries and evaluation of accomplished matches.
The system draws on countless data points, including the pre-tournament world rankings and the continued play: 128 games in the primary round, 64 within the second, and so forth through the ultimate two men's and girls's championship games. The win probability predictions for every individual game – a well-liked statistic for discussion amongst fans and media commentators – are also generated using AI evaluation of recent performance.
After that process is complete, the generative AI system creates pre-game bullet points. “Every morning, we process the games scheduled for the day through our agent-based architecture,” says Baughman. “The Granite 13b chat model creates our bullet points. We use a few-shot technique where we give it examples to follow and produce similar results.” The pre-game bullet points provide insights based on rankings, head-to-head comparisons and player bios, provide fans with context for the upcoming game and are published on the web site and app.
When a game is over, the system generates text descriptions of what happened – based on statistics akin to aces, break points won, double faults, winners and shot speed. These descriptions are then converted into natural language bullet points by generative AI models, including IBM Granite, hosted on the IBM® watsonx™ AI and data platform.
Next, match reports are generated. These reports are based on trusted US Open data and leverage the combined power of Granite and other models hosted on IBM® watsonx.ai™ to provide in-depth summaries. “The match reports summarize who played, what happened and help explain why a player won,” says Baughman. The reports are then reviewed, edited and published on the app and website, in addition to by the USTA editorial team.
Before the Content Engine, editors needed to spend hours watching replays and interpreting scores and statistics before they may write longer articles to publish on US Open media channels. Match reports provide at-a-glance storylines and highlights in order that they can start writing immediately. And for the primary time ever, the USTA's editorial team will give you the chance to publish a match report for each men's and girls's singles match this 12 months.
Increased development speed and improved collaboration with Watsonx Code Assistant
IBM® watsonx™ Code Assistant™ delivers enterprise-grade code generation and provides snippets and functions to speed up application modernization, automation, and scaling. Trained on Granite foundation models, the assistant provides AI-generated recommendations based on existing source code and responds to natural language queries.
Watsonx Code Assistant accelerated development of key parts of the content engine. Using a code plugin of their integrated development environment, developers could chat with the assistant from a sidebar. They could ask it questions, akin to find out how to randomly select text from an array, after which receive a really useful snippet of code that they may copy and adapt to the info they were using.
In addition to code generation, the watsonx Code Assistant provides priceless insights for collaborating teams trying to know and construct on existing code.
“For example, if someone wrote a function or method, I didn't need to read through all of the code to determine what it did,” says Baughman. “I could highlight the block and the wizard would summarize what the code did. It also helped us create comment blocks that described at a high level what each variable represented. We could more easily judge what the methods would return and find out how to use them.”
If the output remains to be unclear, the assistant also answers follow-up questions. The application also allows immediate feedback on the output via thumbs up or thumbs down, which improves the performance of the tool.
As collaboration and conversation with watsonx Code Assistant progresses, users can either construct on the present chat history or reset it for brand new contexts and questions.
“Let's say you're writing a generative AI fact checker. Some of the outcomes you want, some you don't,” says Baughman. “You're on top of things and may select the code that works best for you. If you don't understand a snippet of code, you may ask an issue about it. It's almost like selecting your individual development adventure.”
Project 2: Audio Commentary
Introduced last 12 months, AI-generated audio commentary provides automated voiceovers and captions for every single-match highlight video shown on the US Open website and app. This feature uses a mixture of models, including Granite 13b chat models, to create complex tennis language to support the generated commentary.
A key goal this 12 months was to make the audio commentary more natural and human-like. The team experimented with two variables: Top k, a parameter that controls the variety of possible answers the model should consider, and temperature sampling, which adjusts the probability distribution of possible answers. These levers help the model generate a more human number of phrases, reasonably than the probably and most repetitive ones.
The team was also capable of influence personality through prompts, that are system-level instructions in natural language that tell the model what to do. “We told the model, 'Create an interesting commentary about this game,'” says Baughman.
During the testing phase, the teams reviewed and tweaked the commentary. “It's concerning the balance between artistry and control – not a straightforward decision,” says Baughman. “The more control, the more development effort. The less control, the less development effort, however the riskier it’s. The use case tells you where you find yourself.”
The next step is converting text into speech. It is significant that the voices sound convincingly human. Through experimentation and many alternative iterations, the teams ensured that the voices were clear and had the precise prosody – that’s, that the pitch and speed match the essence of what’s being said.
Once these equilibria are found, the inference and output processes are largely unsupervised and nearly real-time. Five events trigger the creation of commentary: end of game, start of game, end of set, end of game point, and end of match point – a complete of about 9,000 commentary events across games. When an event occurs, the commentary systems receive a message and execute the generation of commentary phrases. The sound files are then integrated into video highlight reels.
Experience as a basis for innovation
Through its long-standing close collaboration with the USTA, the IBM Consulting team is exploring recent ways to create engaging fan experiences and improve the productivity of the US Open digital team. It can be a possibility to showcase powerful recent tools akin to the Watsonx Code Assistant and the Granite family of foundation models.
Discover how IBM AI Consulting might help your small business. Discover Granite, IBM's flagship family of enterprise-ready LLMs.
