AU2020104254A4

AU2020104254A4 - Healthcare question answering (qa) method and system based on contextualized language model and knowledge embedding

Info

Publication number: AU2020104254A4
Application number: AU2020104254A
Authority: AU
Inventors: Kunhui LIN; Feng Luo; Xiaoli Wang; Qingfeng Wu
Original assignee: Xiamen University
Current assignee: Xiamen University
Priority date: 2020-04-23
Filing date: 2020-12-23
Publication date: 2021-03-11
Anticipated expiration: 2028-12-23
Also published as: CN111524593A; CN111524593B

Abstract

The present disclosure discloses a healthcare question answering (QA) method and system based on a contextualized language model and knowledge embedding. The method includes: acquiring a healthcare question and a candidate healthcare answer in a healthcare answer database; using a trained contextualized language model to generate contextualized embedding of the healthcare question and contextualized embedding of the candidate healthcare answer; using a trained knowledge representation model to generate knowledge embedding of the healthcare question and knowledge embedding of the candidate healthcare answer; fusing the contextualized embedding and knowledge embedding of the healthcare question to obtain a feature representation of the healthcare question; fusing the contextualized embedding and knowledge embedding of the candidate healthcare answer to obtain a feature representation of the candidate healthcare answer; calculating a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer; and determining a healthcare answer to the healthcare question according to correlations of various candidate healthcare answers in the healthcare answer database with the healthcare question. The present disclosure can realize effective retrieval of answers and improve retrieval efficiency. 13 DRAWINGS 101 Acquire a healthcare question 102 Use a trained contextualized language model to extract contextualized embedding of the10 healthcare question 103 Use a trained knowledge representation model to extract knowledge embedding of the healthcare question 104 Fuse the contextualized embedding of the healthcare question and the knowledge embedding of the healthcare question to obtain a feature representation of the healthcare question 105 Acquire a healthcare answer in a healthcare answer database and mark the healthcare answer as a candidate healthcare answer 106 Use the trained contextualized language model to extract contextualized embedding of the candidate healthcare answer 1 107 Use the trained knowledge representation model to extract knowledge embedding of the candidate healthcare answer 1 108 Fuse the contextuaized embedding ot the candidate healthcare answer and the knowledge108 embedding of the candidate healthcare answer to obtain a feature representation of the candidate healthcare answer 109 Calculate a correlation ot the candidate healthcare answer with the healthcare question accordmg to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer No Calculate correlations of all candidate healthcare answers in the healthcare answer database with the ealthcare questio Yes 110 Determine a healthcare answer to the healthcare question according to the correlations FIG. 1 Page 1 of2

Description

DRAWINGS

101

Acquire a healthcare question

102 Use a trained contextualized language model to extract contextualized embedding of the10 healthcare question

103 Use a trained knowledge representation model to extract knowledge embedding of the healthcare question

104 Fuse the contextualized embedding of the healthcare question and the knowledge embedding of the healthcare question to obtain a feature representation of the healthcare question

105 Acquire a healthcare answer in a healthcare answer database and mark the healthcare answer as a candidate healthcare answer

106 Use the trained contextualized language model to extract contextualized embedding of the candidate healthcare answer

1 107 Use the trained knowledge representation model to extract knowledge embedding of the candidate healthcare answer

1 108 Fuse the contextuaized embedding ot the candidate healthcare answer and the knowledge108 embedding of the candidate healthcare answer to obtain a feature representation of the candidate healthcare answer

109 Calculate a correlation ot the candidate healthcare answer with the healthcare question accordmg to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer

No Calculate correlations of all candidate healthcare answers in the healthcare answer database with the ealthcare questio

Yes 110

Determine a healthcare answer to the healthcare question according to the correlations

FIG. 1

Page 1 of2

HEALTHCARE QUESTION ANSWERING (QA) METHOD AND SYSTEM BASED ON CONTEXTUALIZED LANGUAGE MODEL AND KNOWLEDGE EMBEDDING TECHNICAL FIELD The present disclosure relates to the technical field of healthcare question answering (QA), and in particular, to a healthcare QA method and system based on a contextualized language model and knowledge embedding. BACKGROUND With the development of question answering (QA) platforms, much more users tend to obtain first-hand relevant healthcare information through healthcare QA websites. Moreover, the development of healthcare QA websites greatly facilitates people's lives and alleviates the difficulty in seeking adequate healthcare caused by traditional medical treatment. However, since the Internet is flooded with a large number of information resources, how to efficiently and accurately search for a relevant answer to a question challenges various healthcare websites. In terms of the healthcare QA technology, how to find a relevant answer to a question in massive text information is a huge challenge, and improving the performance of QA retrieval is still the focus of research in the field of healthcare QA. Although many QA retrieval methods have been used in healthcare QA, such as traditional text retrieval and QA retrieval methods based on deep learning, these methods still have great shortcomings in accuracy and other properties of healthcare QA. Traditional QA retrieval methods are based on keyword search, but ignore the semantic information in QA. The QA retrieval methods based on deep learning utilize the lexical, syntactic and semantic features, but ignore the external knowledge information. In addition, most healthcare QA technologies are used to solve factoid questions, but in reality, users often ask very abstract questions, which are non-factoid questions. This makes the existing retrieval performance unsatisfactory. SUMMARY The present disclosure is intended to provide a healthcare QA method and system based on a contextualized language model and knowledge embedding for improving retrieval performance. To achieve the above purpose, the present disclosure provides the following technical solutions. A healthcare QA method based on a contextualized language model and knowledge embedding includes: acquiring a healthcare question; using a trained contextualized language model to generate contextualized embedding of the healthcare question; using a trained knowledge representation model to generate knowledge embedding of the healthcare question; fusing the contextualized embedding of the healthcare question and the knowledge embedding of the healthcare question to obtain a feature representation of the healthcare question; acquiring a healthcare answer in a healthcare answer database and marking the healthcare answer as a candidate healthcare answer; using the trained contextualized language model to generate contextualized embedding of the candidate healthcare answer; using the trained knowledge representation model to generate knowledge embedding of the candidate healthcare answer; fusing the contextualized embedding of the candidate healthcare answer and the knowledge embedding of the candidate healthcare answer to obtain a feature representation of the candidate healthcare answer; calculating a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer; repeating the steps of "acquiring a healthcare answer in a healthcare answer database and marking the healthcare answer as a candidate healthcare answer" to "calculating a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer" to obtain correlations of various candidate healthcare answers in the healthcare answer database with the healthcare question; and determining a healthcare answer to the healthcare question according to the correlations. Optionally, the trained knowledge representation model is a knowledge representation model based on a knowledge graph. Optionally, the knowledge representation model is trained by a method including: acquiring a knowledge graph; extracting a corresponding entity relationship triple from a resource description framework (RDF) file corresponding to the knowledge graph, and converting the entity relationship triple into a knowledge graph encoding file, where, the entity relationship triple includes: entity, relationship, and entity-relationship pair, and the knowledge graph encoding file includes an entity number file, a relationship number file, and an entity-relationship pair number file; and training the knowledge representation model based on the knowledge graph encoding file to obtain the trained knowledge representation model. Optionally, the contextualized language model is a BERT model. Optionally, the knowledge representation model is a TransE model. Optionally, the calculating a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer specifically includes: inputting the feature representation of the healthcare question and the feature representation of the candidate healthcare answer into a trained deep neural network (DNN) model to obtain the correlation of the candidate healthcare answer with the healthcare question. Optionally, the DNN model is a PACRR model, a KNRM model, or a DRMMTKS model. Optionally, the determining a healthcare answer to the healthcare question according to the correlations specifically includes: sorting the candidate healthcare answers according to the correlations, and outputting sorted candidate healthcare answers. Optionally, the determining a healthcare answer to the healthcare question according to the correlations specifically includes: outputting a specified number of candidate healthcare answers with a relatively-high correlation as a healthcare answer to the healthcare question. A healthcare QA system based on a contextualized language model and knowledge embedding includes: a healthcare question acquisition module, configured to acquire a healthcare question; a candidate healthcare answer acquisition module, configured to acquire a healthcare answer in a healthcare answer database and mark the healthcare answer as a candidate healthcare answer; contextualized embedding generation module, configured to generate contextualized embedding of the healthcare question and contextualized embedding of the candidate healthcare answer using a trained contextualized language model; knowledge embedding generation module, configured to generate knowledge embedding of the healthcare question and knowledge embedding of the candidate healthcare answer using a trained knowledge representation model; a feature representation generation module for healthcare questions, configured to fuse the contextualized embedding of the healthcare question and the knowledge embedding of the healthcare question to obtain a feature representation of the healthcare question; a feature representation generation module for candidate healthcare answers, configured to fuse the contextualized embedding of the candidate healthcare answer and the knowledge embedding of the candidate healthcare answer to obtain a feature representation of the candidate healthcare answer; a correlation calculation module, configured to calculate a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer; and a healthcare answer determination module, configured to determine a healthcare answer to the healthcare question according to the correlations. Based on specific embodiments provided by the present disclosure, the present disclosure discloses the following technical effects: The healthcare QA method and system provided by the present disclosure combines a contextualized language model and knowledge embedding to generate contextualized embeddings and knowledge embeddings of a healthcare question and a candidate healthcare answer, fuses the contextualized embedding and knowledge embedding of the healthcare question to obtain a feature representation of the healthcare question and fuses the contextualized embedding and knowledge embedding of the candidate healthcare answer to obtain a feature representation of the candidate healthcare answer, and finally determines a final healthcare answer according to the correlation of the feature representation of the candidate healthcare answer with the feature representation of the healthcare question. The present disclosure utilizes contextualized semantic information and external knowledge information to assist in improving the performance of healthcare QA retrieval, and improves the adaptability to healthcare questions. BRIEF DESCRIPTION OF DRAWINGS To describe the technical solutions in the embodiments of the present disclosure or in the prior art more clearly, the following briefly introduces the accompanying drawings required for describing the embodiments. Apparently, the accompanying drawings in the following description show merely some embodiments of the present disclosure, and a person of ordinary skill in the art may still derive other drawings from these accompanying drawings without creative efforts. FIG. 1 is a schematic flowchart of the healthcare QA method based on a contextualized language model and knowledge embedding provided in embodiment 1 of the present disclosure; and FIG. 2 is a structure diagram of the healthcare QA system based on a contextualized language model and knowledge embedding provided in embodiment 2 of the present disclosure. DETAILED DESCRIPTION The technical solutions in the embodiments of the present disclosure are clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are merely a part rather than all of the embodiments of the present disclosure. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure. To make the foregoing objectives, features, and advantages of the present disclosure more comprehensible, the present disclosure is further described in detail below with reference to the accompanying drawings and specific implementations. A first aspect of the present disclosure provides a healthcare QA method based on a contextualized language model and knowledge embedding. FIG. 1 is a schematic flowchart of the healthcare QA method based on a contextualized language model and knowledge embedding provided in embodiment 1 of the present disclosure. As shown in FIG. 1, the healthcare QA method provided in this embodiment includes the following steps: step 101: acquiring a healthcare question; step 102: using a trained contextualized language model to generate contextualized embedding of the healthcare question; step 103: using a trained knowledge representation model to generate knowledge embedding of the healthcare question; step 104: fusing the contextualized embedding of the healthcare question and the knowledge embedding of the healthcare question to obtain a feature representation of the healthcare question; step 105: acquiring a healthcare answer in a healthcare answer database and marking the healthcare answer as a candidate healthcare answer; step 106: using the trained contextualized language model to generate contextualized embedding of the candidate healthcare answer; step 107: using the trained knowledge representation model to generate knowledge embedding of the candidate healthcare answer; step 108: fusing the contextualized embedding of the candidate healthcare answer and the knowledge embedding of the candidate healthcare answer to obtain a feature representation of the candidate healthcare answer; step 109: calculating a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer; repeating steps 105 to 109 to obtain correlations of various candidate healthcare answers in the healthcare answer database with the healthcare question; step 110: determining a healthcare answer to the healthcare question according to the correlations. In the embodiment, a healthcare question q is first acquired, and all candidate healthcare answers {di, d2,...,dn} in the system database are read. After text descriptions corresponding to the question q and the candidate answers {di, d2,...,dn} are acquired, a text content of a specified candidate answer di was selected as an initial input content of the method. The contextualized language model may be a BERT model. The process of generating contextualized embeddings of the healthcare question q and the candidate answer di by the BERT model is as follows: 1. Wordpiece word segmentation is performed on texts of q and di. 2. SequencesWqandWdiobtained from word segmentation of q and di are connected to a sequence W, and a connected sequence W is converted into embedding information of three parts: word segmentation embedding, position embedding, and token embedding, which is input into a multi-layer bidirectional Transformer network for calculation. 3. Output of each layer of Transformer is acquired, and contextualized embeddingsCqand Cdi respectively corresponding to q and di are extracted from the output. The knowledge representation model may be a knowledge representation model based on a knowledge graph. As an implementation, the knowledge representation model may be trained by a method including the following steps: acquiring a large number of knowledge graphs; extracting a corresponding entity relationship triple from a resource description framework (RDF) file corresponding to the knowledge graph, and converting the entity relationship triple into a knowledge graph encoding file, where, the entity relationship triple includes: entity, relationship, and entity-relationship pair, and the knowledge graph encoding file includes an entity number file, a relationship number file, and an entity-relationship pair number file; and training the knowledge representation model based on the knowledge graph encoding file to obtain a trained knowledge representation model. The knowledge representation model here may be a traditional knowledge representation model, TransE model. The generated knowledge graph encoding file may be input into the TransE model to obtain a knowledge representation E corresponding to the knowledge graph. With named entity recognition (NER), such as the NER module in the spacy toolkit and tagme tool, entities included in the healthcare question q and candidate healthcare answer di are recognized and linked to corresponding entities in the knowledge graph to generate entity sequences Eq and Edi, separately. According to the entity sequences Eq and Edi recognized from q and di, vectors indexed to corresponding entities in the knowledge representation E are used as knowledge embeddings Kgand Kai. As an implementation, the method of fusing the contextualized embedding and knowledge embedding in steps 104 and 108 may be as follows: the contextualized embedding and knowledge embedding are used as different matrices to fuse the contextualized embedding and knowledge embedding of the healthcare question and the contextualized embedding and knowledge embedding of the candidate healthcare answer, respectively. For example, given (C,K) and (C,,Ka) transformation parameter matrices ME Rc and NE R*d (C and k are vector space dimensions of contextualized embedding and knowledge embedding, respectively) are used to respectively project the contextualized embedding and knowledge embedding into the same vector space Rd to obtain new feature representations CKq and CK"i. A detailed calculation process is as follows: CKq =Cq* M@ Kq* N

CKd = Cd* M D Kdl* N

As an implementation, step 109 may specifically include: the feature representation of the healthcare question and the feature representation of the candidate healthcare answer are input into the trained DNN model, and the DNN model calculates a correlation matrix R (R =(CKq* CKdi) of the word segmentation granularity; and the correlation abstract feature vector in the correlation matrix R is extracted using a multi-kernel convolution of a convolutional neural network (CNN) and a Gaussian filter and input into a multilayer perceptron (MLP) to obtain a question-answer correlation S i. The DNN model may be an existing PACRR model, KNRM model or DRMMTKS model. As an implementation, step 110 may be: sorting the candidate healthcare answers according to the correlations, and outputting sorted candidate healthcare answers. As another implementation, step 110 may be: outputting a specified number of candidate healthcare answers with a relatively-high correlation as a healthcare answer to the healthcare question. The healthcare QA method provided in this embodiment is written in python, combines the pytorch deep learning framework in the field of deep learning and Nvidia GPU server computing power, and learns healthcare QA tasks through end-to-end supervised learning, which can achieve effective retrieval of answers to healthcare questions raised by users and thus improve the efficiency of healthcare QA retrieval. The contextualized model in the present disclosure learns a large amount of text semantic information through unsupervised pretraining of a large amount of text, in combination with two tasks of text content prediction and continuous sentence prediction. Moreover, in combination with a knowledge representation model obtained by unsupervised pretraining with the knowledge graph, more external knowledge is input so that potential correlative information in the question and answer can be discovered, further improving the performance of the model. The question-answer correlation of the present disclosure is more explanatory, andfine-grained correlation calculations such as word segmentation and entity are recognized. The method first acquires embedding representations of word segmentation and entity granularity in QA through the WordPiece word segmentation and NER technology. On this basis, a correlation matrix is obtained through similarity calculation, and by visualizing the correlation matrix, the question-answer correlation can be easily explained. The effect of the method provided in the present disclosure is verified, and the verification experiment results are shown in Table 1. Table 1: Experimental method MAP Precision@3 Recall@3 NDCG@3

(average (accuracy) (recall) (normalized loss accuracy) cumulativegain) feature-based DRMMTKS 0.7676 0.3010 0.9030 0.7868 knowledge-embedding DRMMTKS-W 0.7762 0.3019 0.9056 0.7944 context-based DRMMTKS-BC 0.9421 0.3310 0.9929 0.9540

CK-HQA DRMMTKS-BC-W 0.9430 0.3310 0.9929 0.9573

Notes: feature-based, knowledge-embedding, and context-based are benchmark models. It can be seen from the experimental results that the method CK-HQA (a healthcare QA method based on a contextualized language model and knowledge embedding) provided by the present disclosure is significantly superior to other benchmark models, and has achieved the optimal results in all experimental indicators. A second aspect of the present disclosure provides a healthcare QA system based on a contextualized language model and knowledge embedding. FIG. 2 is a structure diagram of the healthcare QA system based on a contextualized language model and knowledge embedding provided in embodiment 2 of the present disclosure. As shown in FIG. 2, the healthcare QA system provided in this embodiment includes: a healthcare question acquisition module 201, configured to acquire a healthcare question; a candidate healthcare answer acquisition module 202, configured to acquire a healthcare answer in a healthcare answer database and mark the healthcare answer as a candidate healthcare answer; contextualized embedding generation module 203, configured to generate contextualized embedding of the healthcare question and contextualized embedding of the candidate healthcare answer using a trained contextualized language model, where, the contextualized language model may be a BERT model; knowledge embedding generation module 204, configured to generate knowledge embedding of the healthcare question and knowledge embedding of the candidate healthcare answer using a trained knowledge representation model, where, the knowledge representation model may be a TransE model; a feature representation generation module 205 for healthcare questions, configured to fuse the contextualized embedding of the healthcare question and the knowledge embedding of the healthcare question to obtain a fused feature representation of the healthcare question; a feature representation generation module 206 for candidate healthcare answers, configured to fuse the contextualized embedding of the candidate healthcare answer and the knowledge embedding of the candidate healthcare answer to obtain a fused feature representation of the candidate healthcare answer; a correlation calculation module 207, configured to calculate a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer, where, the correlation may be calculated by using a trained DNN model, and the DNN model may be a PACRR model, a KNRM model, or a DRMMTKS model; and a healthcare answer determination module 208, configured to determine a healthcare answer corresponding to the healthcare question according to the correlations. The healthcare QA system provided in this embodiment can achieve effective retrieval of answers to healthcare questions raised by users and improve the retrieval efficiency of healthcare QA. Each embodiment of the present specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same and similar parts among the embodiments may refer to each other. For a system disclosed in the embodiments, since the system corresponds to the method disclosed in the embodiments, the description is relatively simple, and reference can be made to the method description. Specific examples are used herein for illustration of the principles and implementations of the present disclosure. The description of the embodiments is used to help understand the method and its core principles of the present disclosure. In addition, those skilled in the art can make various modifications to specific implementations and application scope in accordance with the teachings of the present disclosure. In conclusion, the content of the present specification shall not be construed as a limitation to the present disclosure.

Claims

What is claimed is: 1. A healthcare question answering (QA) method based on a contextualized language model and knowledge embedding, comprising: acquiring a healthcare question; using a trained contextualized language model to generate contextualized embedding of the healthcare question; using a trained knowledge representation model to generate knowledge embedding of the healthcare question; fusing the contextualized embedding of the healthcare question and the knowledge embedding of the healthcare question to obtain a feature representation of the healthcare question; acquiring a healthcare answer in a healthcare answer database and marking the healthcare answer as a candidate healthcare answer; using the trained contextualized language model to generate contextualized embedding of the candidate healthcare answer; using the trained knowledge representation model to generate knowledge embedding of the candidate healthcare answer; fusing the contextualized embedding of the candidate healthcare answer and the knowledge embedding of the candidate healthcare answer to obtain a feature representation of the candidate healthcare answer; calculating a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer; repeating the steps of "acquiring a healthcare answer in a healthcare answer database and marking the healthcare answer as a candidate healthcare answer" to "calculating a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer" to obtain correlations of various candidate healthcare answers in the healthcare answer database with the healthcare question; and determining a healthcare answer to the healthcare question according to the correlations.
2. The healthcare QA method based on a contextualized language model and knowledge embedding according to claim 1, wherein, the trained knowledge representation model is a knowledge representation model based on a knowledge graph; wherein, the knowledge representation model is trained by a method comprising: acquiring a knowledge graph; extracting a corresponding entity relationship triple from a resource description framework

(RDF) file corresponding to the knowledge graph, and converting the entity relationship triple into a knowledge graph encoding file, wherein, the entity relationship triple comprises: entity, relationship, and entity-relationship pair, and the knowledge graph encoding file comprises an entity number file, a relationship number file, and an entity-relationship pair number file; and training the knowledge representation model based on the knowledge graph encoding file to obtain the trained knowledge representation model.
3. The healthcare QA method based on a contextualized language model and knowledge embedding according to claim 1, wherein, the contextualized language model is a BERT model; wherein, the knowledge representation model is a TransE model; wherein, the calculating a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer specifically comprises: inputting the feature representation of the healthcare question and the feature representation of the candidate healthcare answer into a trained deep neural network (DNN) model to obtain the correlation of the candidate healthcare answer with the healthcare question; wherein, the DNN model is a PACRR model, a KNRM model, or a DRMMTKS model.
4. The healthcare QA method based on a contextualized language model and knowledge embedding according to claim 1, wherein, the determining a healthcare answer to the healthcare question according to the correlations specifically comprises: sorting the candidate healthcare answers according to the correlations, and outputting sorted candidate healthcare answers; wherein, the determining a healthcare answer to the healthcare question according to the correlations specifically comprises: outputting a specified number of candidate healthcare answers with a relatively-high correlation as a healthcare answer to the healthcare question.
5. A healthcare QA system based on a contextualized language model and knowledge embedding, comprising: a healthcare question acquisition module, configured to acquire a healthcare question; a candidate healthcare answer acquisition module, configured to acquire a healthcare answer in a healthcare answer database and mark the healthcare answer as a candidate healthcare answer; contextualized embedding generation module, configured to generate contextualized embedding of the healthcare question and contextualized embedding of the candidate healthcare answer using a trained contextualized language model; knowledge embedding generation module, configured to generate knowledge embedding of the healthcare question and knowledge embedding of the candidate healthcare answer using a trained knowledge representation model; a feature representation generation module for healthcare questions, configured to fuse the contextualized embedding of the healthcare question and the knowledge embedding of the healthcare question to obtain a feature representation of the healthcare question; a feature representation generation module for candidate healthcare answers, configured to fuse the contextualized embedding of the candidate healthcare answer and the knowledge embedding of the candidate healthcare answer to obtain a feature representation of the candidate healthcare answer; a correlation calculation module, configured to calculate a correlation of the candidate healthcare answer with the healthcare question according to the feature representation of the healthcare question and the feature representation of the candidate healthcare answer; and a healthcare answer determination module, configured to determine a healthcare answer to the healthcare question according to the correlations.