WikiPath: Explainable Wikipedia-Grounded Dialogue via Explicit Knowledge Selection and Entity-Path Planning

Abstract

Knowledge-grounded open-domain dialogue systems avoid hallucination by anchoring responses in external sources. This paper presents WikiPath, a lightweight and fully explainable pipeline that treats dialogue generation as (i) explicit Wikipedia sentence selection and (ii) explicit entity-path planning over a turn-local entity graph. WikiPath ranks candidate Wikipedia sentences with BM25 and then constructs a graph whose nodes are the candidate page titles plus a source entity from the dialogue state. Candidates receive an additional shortest-path bonus that prefers knowledge connected to the current entity focus, yielding an explicit and auditable entity chain for every turn. The response generator copies the selected knowledge sentence to provide strict grounding and transparent provenance. We conduct full empirical evaluations on the Wizard of Wikipedia benchmark and report knowledge selection F1, entity grounding accuracy, and response groundedness. On the test-seen split (965 dialogues, 3865 turns), WikiPath improves knowledge selection F1 from 0.1868 to 0.1888, entity grounding accuracy from 0.2732 to 0.2828, and response groundedness F1 from 0.1379 to 0.1386 compared to retrieval-only BM25. On test-unseen, improvements persist (KnowF1 0.1845→0.1858; EntityAcc 0.2650→0.2714). Correlation analyses reveal a consistent negative association between entity diversity and an automatic user-score proxy, highlighting the need to control new-entity introduction for coherent grounded responses. All reported results are reproducible given the dataset files and hyperparameters in this paper.