📄 SAVI: Web-based Multilayered Semantic Annotation Validation Interface

Summary

Universal Semantic Representation (USR), inspired by Pāṇini and the Indian Grammatical Tradition, describes a document's meaning through five layers that depend on one another: lexico-conceptual, constructional, relational, discourse, and co-reference. Because an error in one layer can spread to the others, common annotation tools that handle single layers or text spans cope poorly with validating USR. The authors built SAVI, a web platform for checking and correcting USR annotations. It gives each layer its own tab and enforces structural constraints on the backend so that the layers stay consistent with each other. A multilingual concept dictionary controls sense-level consistency, and new concepts must pass reviewer approval before they are added. To help validators understand an analysis, SAVI renders sentence- and discourse-level graphs and can generate Hindi text from the USR. The authors evaluated the tool with dictionary-governance statistics, measurements of validation efficiency, and usability studies with annotators. They report that it governs concept proposals well and speeds up multilayer correction. SAVI therefore offers infrastructure for building larger USR corpora.

Key points

• Separate editing tabs: Chapter and Segment tabs handle the input text, with segments split so that each contains one finite verb. Five further tabs map to the USR layers. The Relational tab requires every dependent to have exactly one head, and the Construction tab keeps Complex Concept components aligned with token boundaries.
• Four user roles: Admin, Annotator, Reviewer, and Dictionary Validator. The Dictionary Validator approves or rejects proposed concepts before they enter the Concept Dictionary.
• Technical stack: Flutter frontend, Python/Flask backend built from modular blueprints, PostgreSQL accessed through SQLAlchemy, and Gunicorn. Graphviz produces dependency, discourse, and coreference graphs, which are delivered as SVG.
• Comparison with other tools: The authors compare SAVI with BRAT, INCEpTION, WebAnno, TrEd (PDT), and UMR-Writer. They claim SAVI is the only one that combines strong constraint enforcement with language generation (Hindi).
• Context: The work builds on USR Bank 1.0 and belongs to the "Sanskrit Knowledge Accessor" project, funded by MeitY under NLTM: BHASHINI. The provided text did not include the specific evaluation numbers.