Source-linked AI summary
End-to-End Knowledge-Routed Relational Dialogue System for Automatic Diagnosis
Lin Xu, Qixian Zhou, Ke Gong, Xiaodan Liang, Jianheng Tang, Liang Lin
TL;DR
Medical diagnosis dialogue systems need medically rational symptom questioning, but existing data-driven approaches do not encode expert knowledge graphs. KR-DS integrates medical knowledge into end-to-end dialogue management and beats state-of-the-art methods by more than 8% in diagnostic accuracy.
Problem
Automatic diagnosis dialogue systems require symptom questioning consistent with medical knowledge and symptom-disease relations, which existing data-driven systems cannot directly support.
Method
KR-DS integrates medical knowledge graphs and symptom-disease relations into end-to-end dialogue management, using KR-DQN to guide topic transitions alongside NLU and NLG.
Results
More than 8% higher diagnostic accuracy than state-of-the-art methods was achieved across two medical dialogue system datasets.
Takeaways & Limitations
KR-DS generated more reasonable symptom inquiries and higher matching rates alongside higher diagnostic accuracy on the medical diagnosis task.
Takeaways & Limitations
The preceding DQN-based system produced intricate, repeated results and relied on template-based natural language models, limiting its fit to real-world diagnosis.
Abstract
from arXiv · showhide
Beyond current conversational chatbots or task-oriented dialogue systems that have attracted increasing attention, we move forward to develop a dialogue system for automatic medical diagnosis that converses with patients to collect additional symptoms beyond their self-reports and automatically makes a diagnosis. Besides the challenges for conversational dialogue systems (e.g. topic transition coherency and question understanding), automatic medical diagnosis further poses more critical requirements for the dialogue rationality in the context of medical knowledge and symptom-disease relations. Existing dialogue systems (Madotto, Wu, and Fung 2018; Wei et al. 2018; Li et al. 2017) mostly rely on data-driven learning and cannot be able to encode extra expert knowledge graph. In this work, we propose an End-to-End Knowledge-routed Relational Dialogue System (KR-DS) that seamlessly incorporates rich medical knowledge graph into the topic transition in dialogue management, and makes it cooperative with natural language understanding and natural language generation. A novel Knowledge-routed Deep Q-network (KR-DQN) is introduced to manage topic transitions, which integrates a relational refinement branch for encoding relations among different symptoms and symptom-disease pairs, and a knowledge-routed graph branch for topic decision-making. Extensive experiments on a public medical dialogue dataset show our KR-DS significantly beats state-of-the-art methods (by more than 8% in diagnosis accuracy). We further show the superiority of our KR-DS on a newly collected medical dialogue system dataset, which is more challenging retaining original self-reports and conversational data between patients and doctors.
Introduction
The paper proposes an end-to-end Knowledge-routed Relational Dialogue System (KR-DS) that integrates medical knowledge and symptom-disease relations into dialogue management while coordinating NLU and NLG. Across two medical dialogue datasets, KR-DS significantly outperforms state-of-the-art methods by more than 8% in diagnostic accuracy.
- Automatic medical diagnosis dialogue can collect additional symptoms, reduce information-collection costs, and provide reports that help doctors diagnose patients more efficiently.
- Medical diagnosis dialogue systems must ask symptoms related to underlying diseases and maintain medical dialogue rationality, requirements that data-driven topic management struggles to satisfy.
- KR-DS integrates a medical knowledge graph and symptom-disease relations into topic transitions while coordinating dialogue management, natural language understanding, and natural language generation.
- Its Knowledge-routed Relational Deep Q-network combines a knowledge-routed graph branch with a relational refinement branch that learns relations among symptoms and symptom-disease pairs.
- The authors construct a new dataset from online medical forums that preserves patients’ original self-reports and doctor-patient interaction utterances for end-to-end training.
- More than 8% higher diagnostic accuracy is achieved by KR-DS than by state-of-the-art methods across two medical dialogue system datasets.
Related Work
Prior work applies recurrent and sequence-to-sequence architectures to dialogue, incorporates knowledge bases into generation, and uses deep reinforcement learning for automatic diagnosis. However, diagnosis systems relied on limited dialogue-management targets and simulated or simplified data, motivating richer relation modeling and graph reasoning.
- Neural Dialogue Systems: RNN and sequence-to-sequence architectures support dialogue systems by learning latent representations without artificial state labels and by modeling task-oriented dialogue.The cited works include Wen et al., Serban et al. 2016, Zhao et al. 2017, Sutskever, Vinyals, and Le 2014, Vinyals and Le 2015, Eric and Manning 2017, and Madotto, Wu, and Fung 2018.
- Knowledge-Based Dialogue: Knowledge-based dialogue research introduces commonsense knowledge, neural knowledge diffusion, and knowledge-base augmentation to improve dialogue generation and task-oriented systems.These directions are represented by Young et al. 2017, Liu et al. 2018, and Eric et al. 2017.
- Reinforcement Learning for Diagnosis: Deep reinforcement learning has been applied to automatic diagnosis, but prior systems targeted dialogue-state tracking and policy learning using simulated or simplified data.The related diagnosis studies are Tang et al. 2016 and Wei et al. 2018; their data does not reflect real diagnosis situations.
Proposed Method
KR-DS is an end-to-end task-oriented medical dialogue system that combines NLU, dialogue management, template-based NLG, and a user simulator for reinforcement-learning training. Its novel KR-DQN integrates basic action prediction, relational refinement, and medical knowledge-routed graph reasoning to select diagnosis-oriented dialogue actions.
- System Architecture: KR-DS combines Bi-LSTM NLU, dialogue management, template-based NLG, and a user simulator conditioned on generated patient goals for end-to-end reinforcement-learning training.NLU extracts intents and slots into semantic frames, while dialogue management requests symptoms or informs diseases.
- Natural Language Understanding: For Chinese medical dialogue, NLU jointly recognizes BIO tags and sentence intent, normalizes symptoms and diseases, and uses dialogue context to fill missing requested slots.The model is supervised-pretrained from labeled intents and symptoms before joint reinforcement-learning training with the rest of KR-DS.
- Knowledge-Routed Relational DQN: KR-DQN extends a basic DQN with relational refinement and knowledge-routed graph branches that use medical knowledge and action dependencies to generate reasonable dialogue actions.The basic branch produces rough actions; the relational branch refines them, and the graph branch performs medical graph reasoning and rule-based decision-making.
- Relational Refinement: The relational branch applies a learnable asymmetric relation matrix to action predictions, with dataset-derived conditional-probability initialization guiding dependency learning.Each refined action is a weighted sum of initially predicted actions, and experiments report that this initialization outperforms random initialization.
- Medical Knowledge Routing: The knowledge-routed graph models diseases and symptoms as nodes connected by conditional-probability edges, representing candidate disease probabilities from observed positive, negative, and unmentioned symptoms.This module simulates doctors’ process of narrowing candidate diseases by asking about significant symptoms.
- Dialogue Management Training: Deep Q-learning trains dialogue management with target networks and experience replay, using rewards for successful diagnosis and penalties for failure, lengthy dialogue, and repeated symptom requests.The policy selects actions to maximize future reward in a sequential diagnosis task with large action spaces and explicit success or failure outcomes.
Experiments
Experiments evaluate KR-DS on the MZ benchmark and a newly collected DX dataset preserving patient self-reports and doctor–patient interactions. KR-DS achieves higher diagnosis accuracy and symptom matching, with ablations, human ratings, and qualitative examples supporting its knowledge-guided dialogue behavior.
- Datasets: The DX dataset preserves original patient self-reports and doctor–patient utterances, covering five diseases, while evaluation uses 423 training dialogues and 104 test dialogues.The diseases are allergic rhinitis, upper respiratory infection, pneumonia, pediatric hand-foot-mouth disease, and pediatric diarrhea.
- Benchmark results: KR-DS achieves higher diagnosis accuracy and matching rate than the compared methods on the MZ and DX medical-dialogue evaluations.The paper context reports that KR-DS beats state-of-the-art methods by more than 8% in diagnosis accuracy; the matching rate assesses whether acquired symptoms are reasonable.
- Component analysis: Ablations show that both the knowledge-routed graph branch and relation refinement branch improve performance, while conditional-probability relation initialization outperforms random initialization.The MZ ablation study targets the two principal KR-DS components and compares relation-matrix initialization strategies.
- Reward analysis: Using smaller reward magnitudes yields similar accuracy while producing a more stable training process.The reward-magnitude experiment evaluates alternative success, failure, and penalty settings on MZ.
- Human evaluation: Human evaluators rate KR-DS highest for diagnosis validity, symptom rationality and relevance, and dialogue fluency, while Basic DQN frequently requests unrelated symptoms.Three medically trained users evaluated all 104 test dialogues across these criteria.
- Qualitative analysis: Qualitative DX examples show KR-DS using prior knowledge and symptom–disease relations to acquire discriminatory symptoms rather than immediately producing a potentially misdiagnosed result.The examples compare KR-DS with pure DQN and display the related knowledge graph for each diagnosis process.
Conclusions
The paper develops KR-DS, an end-to-end medical dialogue system whose dialogue management, natural language understanding, and natural language generation cooperatively optimize through reinforcement learning. Its KR-DQN manages topic transitions by extending DQN with relational refinement for symptom and symptom–disease relations.
- KR-DS jointly optimizes dialogue management, natural language understanding, and natural language generation through reinforcement learning.
- KR-DQN extends a basic DQN to manage topic transitions in the dialogue system.
- Its relational refinement branch encodes relations among different symptoms and symptom–disease pairs.