a happy movieabout war
One request. Two constraints a retrieval system needs to hold together.
KRAG · Presented at NGEN-AI 2026
Best Research Award, Ramapo College
When a relevant answer misses the intent
A request for a happy movie about war contains two different constraints. A retrieval system can match the topic and still miss the experience the person wanted. My research formalizes this as intent-content conflict.
I built Knowledge-graph Retrieval with Affective Grounding (KRAG), combining semantic retrieval with graph relationships and an explicit affective constraint. Across controlled movie-retrieval trials, it improved intent fidelity in all six conflict scenarios while retaining topical relevance.
In conflict trials, KRAG found a jointly relevant top-ten result in 521 of 800 cases, compared with 263 for Vector-RAG.
Read the method and results
Results from the controlled benchmark | Evaluation | 1,600 trials: 800 agreement, 800 conflict |
|---|
| Conflict trials with a jointly relevant top-ten result | 521 / 800 for KRAG
263 / 800 for Vector-RAG |
|---|
| Affective displacement error at 10, under conflict | 0.384 → 0.347 with affective weighting |
|---|
| Rank impact of removing emotion-edge evidence | 2.3× greater under conflict than agreement |
|---|
KRAG represents affect as graph evidence, aligns graph representations with Sentence-BERT, and reranks semantic candidates using topical, relational, and affective signals. Edge-removal experiments test how much the rankings depend on emotion relationships.
I presented Resolving Intent-Content Conflict in Retrieval: An Emotion-Aware Knowledge Graph Framework at NGEN-AI 2026 in September, following presentations at the Ramapo DMC Fair and Scholar’s Day. The conference proceedings are forthcoming in Springer’s CCIS series.
