AWS for M&E Blog
PGA TOUR automates play-by-play commentary with Amazon Bedrock
A professional golf tournament spans multiple days, with up to 156 athletes playing 18 holes daily across 200 acres of fairway, bunkers, greens, water features, and more. The super-sized, sprawling nature of each event makes it inherently challenging for the PGA TOUR to keep fans apprised of action taking place across an extensive field of play. PGA TOUR coverage of golf tournaments over four days is unlike other mainstream sports like football, baseball, or hockey, which generally take place in a single arena or stadium for only a few hours.
The PGA TOUR has long partnered with Amazon Web Services (AWS) to help scale and deepen its coverage to provide more real-time updates to its fans. The latest advancement to develop from this partnership is set to revolutionize the golf fan experience through generative artificial intelligence (AI) and machine learning (ML). In collaboration with NLP Logix, an AWS Partner with specialization in AI/ML systems and automation solutions, the PGA TOUR recently developed a solution to generate context-driven commentary leveraging Amazon Bedrock, a fully managed service from AWS, to build and scale generative AI applications.
“We’re still barely scratching the surface of what’s possible in terms of personalizing the fan experience with generative AI. AWS is helping us transform how we broadcast the sport of golf, and fans are loving it. We’re unlocking new ways to bring audiences closer into the action and take our coverage to new heights,” said Scott Gutterman, SVP of Digital Operations, PGA TOUR.
TOUR Championship debut
The inaugural run of the PGA TOUR’s automated play-by-play commentary powered by Amazon Bedrock is set for the TOUR Championship, taking place August 29 – September 1, 2024 at East Lake Golf Club in Atlanta, Georgia.
A private demonstration for its first outing, the solution will display text play-by-play commentary on an internal build of TOURCAST, the PGA TOUR’s next-generation experience that tracks players in real-time and allows fans to follow their favorite players live or to dive deep into a specific golfer’s play following completion of their round.
TOURCAST allows fans to see radar data and shot shapes, view players’ progress around the course from different angles, view video highlights, and other in-depth features. By integrating generative AI-powered play-by-play commentary with TOURCAST, the PGA TOUR can offer fans a deeper contextual analysis of every shot by every player for each event.
Taking AI commentary to the next level
Along with the upcoming rollout of text-based AI commentary in TOURCAST, the PGA TOUR already has its sights set on the solution’s next evolution: audio commentary. Because golf is an international sport with top players hailing from around the world, golf fans are similarly global. The PGA TOUR already delivers “Every Shot Live” during THE PLAYERS Championship, held annually in March, which includes 48 different streams of live video such that fans can watch every shot of their favorite players via their local broadcast app. Currently, only the coverage that goes to television audiences features live commentary, as it historically has been cost-prohibitive to run commentary on all 48 streams.
Using this new automated play-by-play commentary solution powered by Amazon Bedrock, the PGA TOUR can not only provide contextual audio commentary on Every Shot Live video streams but also generate commentary in multiple languages, all in real time. The PGA TOUR is working on this live commentary solution with a goal of demonstrating or releasing it in 2025.
How it works
Play-by-play commentary comprises two components – Facts and Context. Facts simply convey shot details in terms of the who, what, when, and where. Context then explains why the shot is interesting. The PGA TOUR’s solution uses PGA TOUR data to build a Fact Prompt, describing who, what, when and where. The Fact Prompt is then passed to Amazon Bedrock to generate a Fact-based commentary. The Fact-based commentary and shot data are combined to build the Context prompt, which is again passed to Amazon Bedrock to generate Factual and Contextual commentary. The following diagram illustrates this process with an example.
This approach leads to the creation of smart AI-generated commentary with context, such as “Joel Dahmen one-putted on every green in his round today,” rather than delivering only basic facts, “Joel Dahmen: shot 1 hit 225 yards, 3 feet to the hole.” It also enables a predictive element; “Joel Dahmen could take the lead if he converts a 20-foot putt from his tee shot on the par-3 seventeenth hole.”
As the development of the AI play-by-play commentary solution progresses, the PGA TOUR plans to integrate the ability to send personalized push notifications to viewers based on their favorite player selections.
Technical architecture
PGA TOUR data is distributed via the distribution system into the “Shot Feed” Amazon SNS Topic. The “Shot Queue” Amazon SQS queue subscribes to the “Shot Feed” Topic. Shot events in Amazon SQS queue triggers an AWS Lambda function, which makes an API call via Amazon API Gateway into GenAI Logic that runs in an Amazon ECS container and creates a prompt based on received TOUR data. Amazon Bedrock LLMs generate context-based narrative commentary using the created prompt as input.
Under the hood, commentary is generated leveraging multiple models in Amazon Bedrock using multiple prompts. Currently the TOUR is using Claude 3.5 Sonnet, which was selected to balance cost, latency, and quality across the various prompts specific to this use case. Generated commentary publishes to an Amazon SNS topic that can be subscribed to by other systems, such as TOURCAST.
Data, including shot history, leaderboard history, and truth deck, are stored in Amazon DynamoDB through data refreshes from TOUR data stores. This data is used both in contextual prompt generation and validation of generated commentary for accuracy. Amazon CloudWatch monitors all the services and alerts system engineers to any issues in generating or validating commentary.
Maintaining the PGA TOUR gold standard
Accuracy and quality are very important to the PGA TOUR, so generated commentary is thoroughly validated in multiple steps, whether used for text or audio output. With around 32,000 shots taken at each tournament, it’s not possible for a staff member to validate each output. Establishing a strong validation methodology is key to ensuring that proposed narratives align with PGA TOUR guidelines, and are both factually and contextually accurate before delivery to fans.
TOUR data is used to generate the commentary using an LLM model. Next, facts and context are extracted from generated commentary using an LLM model. The extracted data is compared with the TOUR data to calculate accuracy score based on identified discrepancies.
The PGA TOUR employs a multi-factor validation process to assess the accuracy and quality of live commentary generations. If the combined validation scores meet the predetermined threshold, the commentary is published. Otherwise, it is withheld. This process ensures high-quality content while minimizing the risk of inaccurate information being disseminated.
Personalizing the PGA TOUR fan experience at scale
The PGA TOUR orchestrates nearly 50 golf tournaments a year. By embracing generative AI across its operations, the PGA TOUR continues to innovate and elevate the fan experience with new features, such as this automated play-by-play commentary, while continuously improving operational efficiencies and employee productivity. Learn more about how PGA TOUR creates fan-driven experiences at scale using AWS.