MT-Telescope lifts the hood on quality performance of machine translation (MT) systems, building on COMET’s predictive quality measurement. This open-source tool delivers a zoomed-in view of performance, so the best MT system can be deployed, for the right reasons.
Greater visibility into performance drivers
MT-Telescope further enhances our ability to evaluate machine translation performance. Where COMET offers a way to predict quality measures of human-executed analysis (for example, MQM), MT-Telescope gives us a range of comparative metrics to help us look deeper into the quality performance of the system. This much-needed context helps in decision-making in system deployment.
Visualized system comparison
MT-Telescope provides three visualizations that compare two MT systems:
- disparity in quality scores for specific translations
- side-by-side error analysis of each overall system
- a general evaluation of the distribution of quality scores.
- This can all be viewed from within a browser.
Filters that zoom into the detail
In addition to the overall comparison, MT-Telescope allows a filtered view in several categories:
- keywords (named entities): specific words or names
- phrases (terminology): important phrases or descriptions (for example, a product description
- length of text (segment length): useful for different channels, for example, email and chat
- repetitive phrases (duplication): conversational or etiquette style phrases like “how are you?”
These granular, filtered comparisons give organizations the decision-making tools in system deployment.
MT-Telescope helps our LangOps specialists and development teams make smarter decisions on behalf of our customers about which MT system is best suited to their needs, and enables the MT research community to easily use best practice analysis tools to rigorously benchmark their advances.
VP of Language Technologies Unbabel