Semi-Supervised Variational Reasoning for Medical Dialogue Generation

Abstract

Medical dialogue generation aims to provide automatic and accurate responses to assist physicians to obtain diagnosis and treatment suggestions in an efficient manner. In medical dialogues two key characteristics are relevant for response generation: patient states (such as symptoms, medication) and physician actions (such as diagnosis, treatments). In medical scenarios large-scale human annotations are usually not available, due to the high costs and privacy requirements. Hence, current approaches to medical dialogue generation typically do not explicitly account for patient states and physician actions, and focus on implicit representation instead.We propose an end-to-end variational reasoning approach to medical dialogue generation. To be able to deal with a limited amount of labeled data, we introduce both patient state and physician action as latent variables with categorical priors for explicit patient state tracking and physician policy learning, respectively. We propose a variational Bayesian generative approach to approximate posterior distributions over patient states and physician actions. We use an efficient stochastic gradient variational Bayes estimator to optimize the derived evidence lower bound, where a 2stage collapsed inference method is proposed to reduce the bias during model training. A physician policy network composed of an action-classifier and two reasoning detectors is proposed for augmented reasoning ability. We conduct experiments on three datasets collected from medical platforms. Our experimental results show that the proposed method outperforms state-of-the-art baselines in terms of objective and subjective evaluation metrics. Our experiments also indicate that our proposed semi-supervised reasoning method achieves a comparable performance as state-of-the-art fully supervised learning baselines for physician policy learning.

A Benchmark for Automatic Medical Consultation System: Frameworks, Tasks and Datasets

“…MedDialog (Zeng et al, 2020) is a large scale medical dialogue dataset that contains a Chinese dataset with 3.4 million conversations covering 172 specialties of diseases and an English dataset with 0.26 million conversations covering 96 specialties of diseases. KaMed (Li et al, 2021) is a knowledge aware medical dialogue dataset that contains over 60,000 medical dialogue sessions and is equipped with external medical knowledge from Chinese medical knowledge platform. The tasks built on these corpus are usually response generation in dialogue systems, on which researchers can build automated medical chatbots.…”

Natural Language Processing for Smart Healthcare

“…Text generation (TG) automatically generates text with given inputs while pursuing the goal of appearing indistinguishable from human-written text. Specifically, there are 3 kinds of inputs and corresponding subtasks in smart healthcare: text inputs (e.g., routine reports) associated with text summarization [65]- [67], question generation [68]- [70], dialogue generation [71]- [73], and etc. ; data inputs (e.g., neonatal intensive care data) connected with data-to-text [74]; and image inputs (e.g., medical images) related to image captioning [75], [76], visual question answering (VQA) [77]- [79], and etc.…”

Medical Dialogue Response Generation with Pivotal Information Recalling

“…Medical dialogue system (MDS) has received much attention due to its high practical value. Previous works [5,15,21] usually model the dialogue history as sequential text and employ the sequenceto-sequence (Seq2Seq) models that built on large-scale pretrained text encoder and decoder to generate medical responses.…”

“…This issue may be caused by the fact that the cross-attention mechanism is not trained with explicit supervision signals when recalling pivotal information. Recent works [8,15,26,32] proposed to extract the medical key phrases and sentences from the dialogue history and incorporate them into response generation via the cross-attention mechanism as well. However, these works bypass the fundamental problem of utilizing medical relations between different utterances, and fail to fully exploit the pivotal information from dialogue history during response generation.…”

“…2) We conduct extensive experiments on large-scale medical dialogue datasets MedDG [21] and MedDialog [5].The experimental results show that our proposed model achieves new state-of-the-art results by outperforming previous strong baselines VRBot [15] and BERT-GPT-Entity [21] on BLEU and medical entities F1 metrics.…”