Faith Demo Shows Transparent Temporal Question Answering Across Heterogeneous Sources
Grafi Research Lab released Faith, a temporal QA system that turns implicit time constraints into explicit conditions and visualizes evidence from KBs.
Source: Grafy Research LabWhat is Faith and why does it matter?
Faith is a newly demonstrated temporal question answering (TQA) system that works over heterogeneous data sources—including Wikidata, Wikipedia text, tables, and infoboxes—while providing traceable, explainable answers. According to the Grafi Research Lab paper, the system addresses two long‑standing gaps in TQA: the inability of most deep‑learning models to handle implicit temporal constraints precisely, and the reliance on a single source of information, which limits coverage.
How does Faith understand temporal questions?
The first stage converts a natural‑language query into a structured representation called the Time‑aware Structured Frame (TSF). The TSF contains six slots: question entity, relation, expected answer type, temporal category (implicit or explicit), temporal signal (e.g., before, after, overlap), and temporal value. A fine‑tuned BART seq2seq model fills the first five slots, while temporal values are extracted with SUTime and regular‑expression matching. When a question contains an implicit time constraint, Faith invokes an implicit question resolver that recursively calls the TQA system to generate intermediate questions (e.g., “When did Cristiano Ronaldo play for Juventus FC start date?”). The answers to those intermediates become explicit date spans such as , which are then inserted into the TSF.
How does Faith retrieve and rank evidence?
With the TSF in hand, Faith performs entity linking to Wikidata using the Clocq tool, then pulls matching facts from the knowledge base and fetches the corresponding Wikipedia page from a locally hosted 2024‑06 dump. All retrieved items—sentences, table rows, infobox fields—are verbalized into uniform textual statements. Temporal expressions inside each statement are normalized to date‑granular spans, and any evidence that conflicts with the TSF’s temporal signal and value is pruned. The remaining candidates are re‑ranked, first with a BM25 scorer and optionally with a cross‑encoder trained on pairs. The top‑100 pieces of evidence are passed to the final answering stage.
How does Faith produce explainable answers?
The answer generation relies on the Explaignn framework, which builds a heterogeneous graph where nodes represent entities (including dates) and evidence statements, and edges connect entities to the evidence that mentions them. A graph neural network (GNN) iteratively scores and trims the graph, gradually discarding low‑relevance nodes until the most probable answer emerges. Because the graph and its reduction steps are visualized in the user interface, users can follow exactly how each piece of evidence contributed to the final answer, satisfying the demand for faithful, traceable reasoning.
Using Faith with Grafy tools
While Faith itself is a standalone Python/Flask application, Grafy users can explore its underlying concepts without writing code. The Reader app lets you upload the demonstration paper (PDF) and ask follow‑up questions, effectively reproducing Faith’s question‑understanding pipeline on‑the‑fly. For developers interested in the recursive resolver or the GNN‑based answer graph, the Code studio provides a sandbox where you can clone the public GitHub repository (https://github.com/zhenjia2017/FAITH-DEMO) and run the Flask backend directly within Grafy’s cloud environment. This hands‑on approach lets you experiment with different source combinations or re‑ranking models, mirroring the configuration options showcased in the demo UI.
Frequently asked questions
Does Faith handle both explicit and implicit temporal constraints?
Yes. Explicit dates are extracted with SUTime, while implicit constraints are resolved by generating intermediate questions and feeding them back into the system, as described in the paper.
Which data sources does Faith draw from to answer a question?
Faith aggregates information from Wikidata knowledge graph, Wikipedia article text, tables, and infoboxes, normalizing all of them into a unified textual format before evidence ranking.
Can I try Faith’s visual reasoning interface without installing anything?
Grafy’s Reader app can load the demo PDF and let you query the system interactively; for the full UI you would need to run the open‑source code in Grafy Code, but the visual steps are also captured in the demonstration video linked in the repository.
Source: https://arxiv.org/abs/2611.19926
Source: Grafy Research Lab — A Demonstration System for Faithful Temporal Question Answering


