SynK-MA: code and experiment outputs for knowledge-graph-grounded multimodal conversation.

Code and experiment outputs for the paper "Grounding Multimodal Conversation in a Knowledge Graph: Referent Resolution with Association Statistics and Typed Context Memory" (P. R. Venkatesh and P. Radha Krishna, National Institute of Technology Warangal). SynK-MA grounds dialogue understanding over text, images and speech in a generated OWL ontology and knowledge graph. It resolves referring expressions with association statistics between words and knowledge-base attributes, tracks context with a typed memory, and compiles frames to SPARQL. It runs on CPU. Files:- synkma-code-v1.0.zip: the SynK-MA system, the baselines (neural tracker, fine-tuned RoBERTa, LLM parsers), every experiment script and the scripts that regenerate the paper's tables. Setup and data sources are described in experiments/README.md. The datasets (SIMMC 2.1, SpokenWOZ) are not included.- synkma_outputs.tar.gz: every result behind the paper's tables and figures, including all runs (metrics with confidence intervals, tuned thresholds, timings, per-turn predictions), the aggregated analyses, cached LLM responses (Claude Sonnet 5.5, Haiku 4.5) with token usage, SIMMC-Spoken ASR transcripts and vision-agent predictions.- README_outputs.txt: contents of the outputs archive and how to use it. To regenerate every table and number of the paper without the datasets or any training: unzip synkma-code-v1.0.zip cd synkma-code-v1.0/experiments pip install -r requirements.txt tar xzf /path/to/synkma_outputs.tar.gz python scripts/make_tables.py SHA-256 of synkma_outputs.tar.gz: 52f9300752c40dfd0f1ce296c5111d24addde46b0b4a74e25284e1234a75ab6dSHA-256 of synkma-code-v1.0.zip: fd0b48a2e21957db518ffde563c5dcb8eaf28f916b249fb805547b68a5ae08a6 Licences: synkma-code-v1.0.zip is under the PolyForm Noncommercial License 1.0.0 (free for research, teaching and other non-commercial use; commercial use needs the authors' permission). synkma_outputs.tar.gz is under CC BY-NC-SA 4.0, as it is derived from SIMMC 2.1 (CC BY-NC-SA 4.0) and SpokenWOZ (CC BY-NC 4.0).