Microsoft Research Focus: Hybrid Work, Data Tables, LLMs, and Pronunciation Assessment
Research Focus: Week of August 14, 2023
This post highlights notable publications, events, code/datasets, new hires, and other milestones from Microsoft Research. It covers advancements in hybrid work solutions, data transformation, in-context learning for large language models, and pronunciation assessment.
NEW RESEARCH
HyWay: Enabling Mingling in the Hybrid World
As remote work becomes more prevalent, traditional videoconferencing tools like Teams are effective for structured meetings. However, for informal interactions like hallway conversations, newer spatial tools have emerged, but they are limited to virtual-only settings. Many organizations now adopt hybrid work models, creating a challenge for remote participants to connect with in-person colleagues. Existing tools often fall short in supporting unstructured interactions or hybrid settings.
Microsoft researchers have developed HyWay, a system designed to facilitate informal interactions between physical and virtual participants. HyWay utilizes large displays in physical zones to allow remote users to see and be seen by in-person attendees. Remote users can navigate between these zones using a map-based interface, while in-person users can move freely. The paper details user survey findings from multiple deployments.
- Key Contribution: HyWay supports informal interactions in hybrid work environments.
- Technology: Uses large displays in physical zones and a map-based interface for navigation.
- Benefit: Bridges the gap between remote and in-person participants for better mingling.
Microsoft Research Podcast
Collaborators: Silica in space with Richard Black and Dexter Greene
This episode features Dexter Greene, a college freshman, and Richard Black, a Microsoft research manager, discussing how technology that stores data in glass is supporting students in their efforts to communicate human experiences to extraterrestrials.
NEW RESEARCH
Auto-Tables: Synthesizing Multi-Step Transformations to Relationalize Tables without Using Examples
Relational tables, where each row is an entity and each column an attribute, are standard in databases. However, a significant percentage of real-world tables (over 30%) are not relational, requiring complex transformations for analysis. Manually programming these transformations is challenging for both technical and non-technical users.
Microsoft researchers have introduced Auto-Tables, a system that automatically synthesizes multi-step transformation pipelines to convert non-relational tables into relational forms. This system eliminates the need for manual programming, making data preparation more accessible.
- Problem: Non-relational tables hinder data analysis.
- Solution: Auto-Tables automates complex data transformations.
- Features: Supports Python and other languages, requires no user input for transformations.
- Evaluation: Successfully transforms over 70% of test cases at interactive speeds.
- Dataset: Includes a benchmark of 244 real-world test cases.
NEW RESEARCH
Learning to Retrieve In-Context Examples for Large Language Models
In-context learning allows large language models (LLMs) to perform tasks with few-shot examples without parameter updates. The effectiveness of this method depends heavily on the quality of the selected examples.
Microsoft researchers propose a novel framework to train dense retrievers for identifying high-quality in-context examples for LLMs. The framework involves training a reward model based on LLM feedback and then using knowledge distillation to train a dense retriever. Experiments show significant performance enhancements in in-context learning across various tasks and LLM sizes.
- Concept: In-context learning relies on quality examples.
- Framework: Trains dense retrievers using LLM feedback and knowledge distillation.
- Results: Demonstrates significant improvements in LLM performance on various tasks.
- Analysis: Retrieves examples with similar patterns for better generalization.
NEW RESEARCH
End-to-End Word-Level Pronunciation Assessment with MASK Pre-training
Computer-Aided Pronunciation Training (CAPT) systems help users improve language skills. Word-level pronunciation assessment is a key challenge, often limited by alignment accuracy in current methods.
Researchers at Microsoft propose Masked pre-training for Pronunciation Assessment (MPA), an end-to-end method that uses a mask-predict strategy to overcome misalignment issues. MPA enables assessment in unsupervised and supervised settings. Experiments on the SpeechOcean762 dataset show MPA outperforms previous methods without explicit alignment, though it requires more inference time and reference text.
- Challenge: Accurate word-level pronunciation assessment in CAPT.
- Method: MPA uses mask-predict for end-to-end training, avoiding alignment issues.
- Capabilities: Supports unsupervised and supervised assessment.
- Performance: Outperforms previous methods on the SpeechOcean762 dataset.
- Limitations: Requires more inference time and reference text.
Related Publications
- End-to-End Word-Level Pronunciation Assessment with MASK Pre-training
- Auto-Tables: Synthesizing Multi-Step Transformations to Relationalize Tables without Using Examples
- Learning to Retrieve In-Context Examples for Large Language Models
- HyWay: Enabling Mingling in the Hybrid World
Read More
- Research Focus: Week of August 26, 2024
- Research Focus: Week of December 18, 2023
- Research Focus: Week of November 22, 2023
- Research Focus: Week of November 8, 2023
Research Areas
- Artificial intelligence
- Audio and Acoustics
- Data platforms and analytics
- Human language technologies
- Human-computer interaction
- Systems and networking
Research Groups
- Mobility, Networks, and Systems
- General Artificial Intelligence
- Microsoft Research Asia - Shanghai
Related Projects
- HyWay
In Relation
- Microsoft Research Lab - Asia
- Microsoft Research Lab - India
Follow Us
- Follow on X
- Like on Facebook
- Share on LinkedIn
- Subscribe on YouTube
- Follow on Instagram
- Subscribe to our RSS feed
Share this page
- Share on X
- Share on Facebook
- Share on LinkedIn
- Share on Reddit
What's New
- Surface Pro
- Surface Laptop
- Surface Laptop Studio 2
- Surface Laptop Go 3
- Microsoft Copilot
- AI in Windows
- Discover Microsoft products
- Windows 11 apps
Microsoft Store
- Account Profile
- Download Center
- Microsoft Store Support
- Returns
- Order Tracking
- Virtual Workshops and Training
- The Microsoft Store Promise
Education
- Microsoft Education
- Devices for Education
- Microsoft Teams for Education
- Microsoft 365 Education
- Office Education
- Educator training and development
- Student and parent deals
- Azure for Students
Business
- Microsoft Cloud
- Microsoft Security
- Azure
- Dynamics 365
- Microsoft 365
- Microsoft Advertising
- Microsoft 365 Copilot
- Microsoft Teams
Developers & IT
- Microsoft Developer
- Microsoft Learn
- AI Marketplace App Support
- Microsoft Technical Community
- Azure Marketplace
- AppSource
- Microsoft Power Platform
- Visual Studio
Company
- Careers
- About Microsoft
- Microsoft Privacy
- Investors
- Accessibility
- Sustainability
Privacy & Cookies
- Communicate with Microsoft
- Privacy
- Manage Cookies
- Terms of Use
- Trademarks
- About our Ads
- © Microsoft 2025
Original article available at: https://www.microsoft.com/en-us/research/blog/research-focus-week-of-august-14-2023/?lang=fr_ca&locale=fr-ca