Method, device, apparatus and computer-readable medium for policy review

Abstract

The present invention discloses a method, device, equipment and computer-readable medium for policy review, and relates to the field of computer technology. A specific implementation of the method includes: inputting a policy to be reviewed into a first prediction model, and if the first prediction result of the policy to be reviewed is a negative sample, determining that the policy to be reviewed is a high-risk policy; if the first prediction result of the policy to be reviewed is a positive sample, inputting the policy to be reviewed into a second prediction model again, and if the second prediction result of the policy to be reviewed is a negative sample, determining that the policy to be reviewed is a medium-risk policy, and if the second prediction result of the policy to be reviewed is a positive sample, determining that the policy to be reviewed is a low-risk policy; based on the risk level of the policy to be reviewed, determining the review method for the policy to be reviewed. This implementation can improve the efficiency of reviewing policies.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, device, equipment and computer-readable medium for insurance policy review. Background Art

[0002] Policy review is the top priority of the insurance business, and intelligence is the key to improving underwriting performance. The "health disclosure" in the policy is a necessary option for users to purchase life insurance and health insurance. The long text description brings considerable trouble to underwriting. Such policies cannot pass the self-review rules and can only be reviewed manually. Since there is no unified format, it is easy to read it wrongly or miss it, and reading the information is time-consuming and laborious.

[0003] In the process of implementing the present invention, the inventors found that there are at least the following problems in the prior art: on the one hand, the rule engine cannot process and judge the text content, and all of it must be transferred to the manual review process, which increases the burden of underwriting; on the other hand, the information of health notification is hidden in a long paragraph of text description, including health status, medical treatment status, physical examination status, surgical treatment, postoperative recovery status, etc., which all need to be understood and refined by underwriters. Therefore, the efficiency of reviewing insurance policies is low. Summary of the invention

[0004] In view of this, embodiments of the present invention provide a method, apparatus, device and computer-readable medium for reviewing insurance policies, which can improve the efficiency of reviewing insurance policies.

[0005] To achieve the above object, according to one aspect of an embodiment of the present invention, a method for reviewing an insurance policy is provided, comprising:

[0006] Input the to-be-reviewed insurance policy into a first prediction model, and if the first prediction result of the to-be-reviewed insurance policy is a negative sample, determine that the to-be-reviewed insurance policy is a high-risk insurance policy, wherein the first prediction model is a model trained by a word list of normal underwriting, a word list of conditional underwriting, and a word list of rejected insurance;

[0007] If the first prediction result of the policy to be reviewed is a positive sample, the policy to be reviewed is input into the second prediction model again, if the second prediction result of the policy to be reviewed is a negative sample, the policy to be reviewed is determined to be a medium-risk policy, and if the second prediction result of the policy to be reviewed is a positive sample, the policy to be reviewed is determined to be a low-risk policy, and the second prediction model is a model trained by a word list of normal underwriting and a word list of conditional underwriting;

[0008] Based on the risk level of the insurance policy to be reviewed, a review method is determined for the insurance policy to be reviewed.

[0009] Before inputting the pending insurance policy into the first prediction model, the method further includes:

[0010] Based on key events, a key event word list for text annotation is constructed, wherein the key events are a collection of health-related information in the text;

[0011] The plurality of key event word lists are divided into the normal underwriting word list, the conditional underwriting word list and the rejected underwriting word list according to the underwriting situation.

[0012] The key event word list for text annotation is constructed based on the key events, including:

[0013] Based on the verbs in key events and the synonyms of the verbs, a key event word list for text annotation is constructed, wherein the synonyms of the verbs include one or more of the following: synonyms of the characters in the verbs, convertible forms of the verbs, similar words to the verbs in the hospital semantic dictionary, and words with the same meaning as the verbs in the hospital semantic dictionary.

[0014] The key event word list for text annotation is constructed based on the key events, including:

[0015] Based on the verbs in the key events, the synonyms of the verbs, the sentence features and the position features in the key events, a key event vocabulary list for text annotation is constructed, the sentence vector is used to identify the global semantic information of the text including the key time, and the position vector is used to identify the semantic information carried by words in different positions.

[0016] The first prediction model is used to obtain the characteristics of the policy that is rejected; the second prediction model is used to obtain the characteristics of the policy that is conditionally covered;

[0017] The first prediction model and the second prediction model are trained based on a neural network model.

[0018] The first prediction model uses a preset method to obtain the characteristics of the rejected insurance policies, and the second prediction model uses the preset method to obtain the characteristics of the conditionally covered insurance policies, and the preset method includes one of the following: attention mechanism, gradient-based sensitivity method, interpretable student model and analysis of hidden layers.

[0019] The step of determining a review method for the insurance policy to be reviewed based on the risk level of the insurance policy to be reviewed includes:

[0020] The risk level of the insurance policy to be reviewed is a low-risk insurance policy, and automatic review is determined for the insurance policy to be reviewed;

[0021] The risk level of the policy to be reviewed is a high-risk policy or a medium-risk policy, and manual review is determined for the policy to be reviewed.

[0022] Determining manual review for the policy to be reviewed includes:

[0023] Generate underwriting case characteristics of the underwriter based on the underwriter's historical underwriting policies;

[0024] The underwriter for manual review is determined based on the characteristics of the insurance policy to be reviewed and the characteristics of the underwriter's underwriting cases.

[0025] According to a second aspect of an embodiment of the present invention, there is provided a device for reviewing an insurance policy, comprising:

[0026] A preliminary screening module, used for inputting the to-be-reviewed insurance policy into a first prediction model, and if the first prediction result of the to-be-reviewed insurance policy is a negative sample, determining that the to-be-reviewed insurance policy is a high-risk insurance policy, wherein the first prediction model is a model trained by a word list of normal underwriting, a word list of conditional underwriting, and a word list of rejected insurance;

[0027] a re-screening module, for inputting the policy to be reviewed into a second prediction model again if the first prediction result of the policy to be reviewed is a positive sample, determining that the policy to be reviewed is a medium-risk policy if the second prediction result of the policy to be reviewed is a negative sample, and determining that the policy to be reviewed is a low-risk policy if the second prediction result of the policy to be reviewed is a positive sample, wherein the second prediction model is a model trained by a word list of normal underwriting and a word list of conditional underwriting;

[0028] The audit module is used to determine the audit method for the insurance policy to be audited based on the risk level of the insurance policy to be audited.

[0029] According to a third aspect of an embodiment of the present invention, there is provided an electronic device for reviewing an insurance policy, comprising:

[0030] one or more processors;

[0031] a storage device for storing one or more programs,

[0032] When the one or more programs are executed by the one or more processors, the one or more processors implement the above method.

[0033] According to a fourth aspect of an embodiment of the present invention, there is provided a computer-readable medium on which a computer program is stored, and when the program is executed by a processor, the method as described above is implemented.

[0034] One embodiment of the above invention has the following advantages or beneficial effects: inputting the to-be-reviewed insurance policy into the first prediction model, and if the first prediction result of the to-be-reviewed insurance policy is a negative sample, then determining that the to-be-reviewed insurance policy is a high-risk insurance policy, and the first prediction model is a model trained by a normal underwriting vocabulary list, a conditional underwriting vocabulary list, and a rejected insurance vocabulary list; if the first prediction result of the to-be-reviewed insurance policy is a positive sample, then inputting the to-be-reviewed insurance policy into the second prediction model again, and if the second prediction result of the to-be-reviewed insurance policy is a negative sample, then determining that the to-be-reviewed insurance policy is a medium-risk insurance policy, and if the second prediction result of the to-be-reviewed insurance policy is a positive sample, then determining that the to-be-reviewed insurance policy is a low-risk insurance policy, and the second prediction model is a model trained by a normal underwriting vocabulary list and a conditional underwriting vocabulary list; based on the risk level of the to-be-reviewed insurance policy, determining the review method for the to-be-reviewed insurance policy. Since the corresponding review method can be determined according to the risk level, the efficiency of reviewing insurance policies can be improved.

[0035] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0037] Figure 1 is a schematic diagram of the main process of the method for policy review according to an embodiment of the present invention;

[0038] Figure 2 is a schematic diagram of a process of constructing a vocabulary list according to an embodiment of the present invention;

[0039] Figure 3 is a schematic diagram of a process for determining a review method for a policy to be reviewed according to an embodiment of the present invention;

[0040] Figure 4 is a schematic diagram of a process for determining an underwriting personnel according to an embodiment of the present invention;

[0041] Figure 5 is a schematic diagram of the use of a nuclear insurance pool according to an embodiment of the present invention;

[0042] Figure 6 is a schematic diagram of an underwriting display page according to an embodiment of the present invention;

[0043] Figure 7 is a schematic diagram of the main structure of a device for policy review according to an embodiment of the present invention;

[0044] Figure 8 is an exemplary system architecture diagram to which embodiments of the present invention may be applied;

[0045] Fig. 9 It is a schematic diagram of the structure of a computer system of a terminal device or a server suitable for implementing an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0047] At present, the following methods are mainly used to improve the intelligence level of the policy review process:

[0048] Method 1: Analyze the core data based on the physical examination report through optical character recognition (OCR) technology, and then input it into the underwriting model. It is only applicable to users with physical examination reports, which account for a small proportion in the insurance policy, have limited help for manual underwriting, and the underwriting model is too simple.

[0049] Method 2: Obtain the diagnosis data of the user in the medical institution, and query the underwriting model through the disease code and insurance type. Since it is necessary to obtain the diagnosis data of the user in the medical institution, the timeliness and applicability are poor, and the underwriting conclusion is only based on the disease code for rule judgment.

[0050] In addition, Method 1 and Method 2 ignore the fact that the risk levels of standard policies, sub-standard policies, extended policies, and rejected policies in the underwriting business are different and cannot be predicted by a single model, resulting in low underwriting efficiency.

[0051] In order to solve the low underwriting efficiency, the following technical solution in the embodiment of the present invention can be adopted.

[0052] See also Figure 1 , Figure 1 Schematic diagram of the main process of the method for policy review according to an embodiment of the present invention, using a preset model to determine the risk of the policy to be reviewed, so as to determine the review method. Figure 1 As shown, the specific steps include:

[0053] S101. Input the policy to be reviewed into the first prediction model. If the first prediction result of the policy to be reviewed is a negative sample, it is determined that the policy to be reviewed is a high-risk policy. The first prediction model is a model trained by a word list of normal underwriting, a word list of conditional underwriting, and a word list of rejected underwriting.

[0054] In an embodiment of the present invention, the risk of the policy to be reviewed is determined by a prediction model. The preset model is obtained by training with training data. The training data can be obtained from the health notice text of the policy.

[0055] Therefore, for different types of insurance policies, since the health notice texts are different, the prediction models used are different. In other words, for insurance policies with the same health notice text, the prediction models used are the same; for insurance policies with different health notice texts, the prediction models used are different.

[0056] In addition, it should be noted that the training data can be obtained from a variety of texts. That is, the training data can be obtained not only from the health notice text in the insurance policy, but also from multiple texts based on actual needs.

[0057] The following takes obtaining training data from a health notification text as an example to illustrate obtaining a vocabulary list including training data.

[0058] See also Figure 2 , Figure 2 1 is a flow chart of constructing a vocabulary table according to an embodiment of the present invention, which specifically includes the following steps:

[0059] S201. Based on key events, a key event vocabulary for text annotation is constructed. Key events are a collection of health-related information in the text.

[0060] There are many contents that can be mentioned in the health notification text. The embodiment of the present invention does not limit the event type. Key event A, key event B, key event C... represent the event type in the health notification text. Including but not limited to health status, illness, medical treatment, physical examination, surgery, etc., the key event set can be determined according to the actual business logic. It can be understood that the key event is a collection of health-related information in the text.

[0061] Based on the key event set, a key event word list for text annotation is constructed, as shown in Table 1. Table 1 is a key time word list. A, B and C in Table 1 represent key event A, key event B and key event C respectively.

[0062] Table 1

[0063] Trigger word dictionary Argument Dictionary A <![CDATA[A1......A i ]]> <![CDATA[a1......a j ]]> B <![CDATA[B1......B i ]]> <![CDATA[b1......b j ]]> C <![CDATA[C1......C i ]]> <![CDATA[c1……c j ]]> ...

[0064] Each key event includes the trigger words that may trigger this type of event, as well as the required argument roles and possible corresponding argument roles under this type of event. The number of words in the argument dictionary is greater than the number of words in the trigger word dictionary.

[0065] The noun collocated with a predicate is called an argument. The argument role is the semantic role assigned by the predicate to these noun phrases according to the semantic relationship between the predicate and the relevant noun phrases, that is, the semantic role played by the argument used with the verb in the action or state expressed by the verb. Each verb has its own argument structure, specifying the required arguments and optional arguments.

[0066] In an embodiment of the present invention, to ensure that the trigger dictionary is controlled within a reasonable range, a key event word list for text annotation is constructed based on the verb and its synonyms in the key event. Among them, the synonyms of the verb include one or more of the following: synonyms of the characters in the verb, transformable forms of the verb, approximate words of the verb in the medical semantic dictionary, and words with the same meaning as the verb in the medical semantic dictionary.

[0067] In specific implementation, the construction process is mainly based on the following points:

[0068] (1) Incorporate the words with the verb part of speech in the training data into the candidate trigger word dictionary, such as "medication, treatment, chemotherapy, radiotherapy, transplantation", etc.

[0069] (2) Add the possible transformable forms of the candidate trigger words, such as "pierce" corresponding to "penetrate" and "perforate".

[0070] (3) According to the medical semantic dictionary, incorporate the approximate words or words with the same meaning of the candidate trigger words. In this way, selecting approximate words and words with the same meaning from a medical perspective can enrich the key event word list and improve the diversity of the word list as training data.

[0071] After constructing the key time word list, it is necessary to generate the key event types and their corresponding trigger words and arguments.

[0072] Specifically, using one-hot encoding or word2vec method, the text is transformed into the word vectors of the sentences. The word vectors of the above sentences are used as the input of the deep learning model. Then, a bidirectional LSTM model (Bidirectional LSTM, Bi-LSTM) is used to automatically learn the abstract features of the key events. Finally, through the CRF (Conditional Random Fields) layer, a softmax classifier is used to predict the features, generating the key event types and their corresponding trigger words and arguments.

[0073] In an embodiment of the present invention, to improve the accuracy of the vector in expressing semantics, a key event word list for text annotation is constructed based on the verb, its synonyms, the sentence features and position features in the key event.

[0074] That is to say, the sentence vector features and position vector features are also added and integrated into the vector. As an example, using the BERT-Base-Chinese pre-trained model, the input is text data, and the sentence vector and position vector are added after embedding.

[0075] Sentence vectors are used to identify the global semantic information of text. Sentence vectors are fused with the semantic information of words through the embedding layer. Position vectors are used to identify the semantic information carried by words at different positions.

[0076] S202. Divide the plurality of key event word lists into a normal underwriting word list, a conditional underwriting word list, and a rejected underwriting word list according to underwriting conditions.

[0077] For multiple key event word lists, they can be divided into: normal underwriting word list, conditional underwriting word list and rejected underwriting word list according to the underwriting situation. As an example, the words that appear frequently in normal underwriting policies are divided into: normal underwriting word list; the words that appear frequently in conditional underwriting policies are divided into: conditional underwriting word list; the words that appear frequently in rejected underwriting policies are divided into: rejected underwriting word list.

[0078] exist Figure 2 In the embodiment, the text is divided into different types of vocabulary lists based on key events in the text to provide training data for the training model.

[0079] In an embodiment of the present invention, the word list is integrated with other underwriting data, including but not limited to the gender, age, customer level, insurance agent credit level, etc. of the insured, and input into the neural network model to realize the training model.

[0080] The risks of the insurance policy review include the following three types: high-risk insurance policy, medium-risk insurance policy and low-risk insurance policy. In the embodiment of the present invention, two separate prediction models are used, a first prediction model and a second prediction model.

[0081] In the embodiment of the present invention, the normal insured standard body insurance policies are classified as group 1, the sub-standard body insurance policies whose underwriting conclusions contain health special mark and health surcharge mark are classified as group 2, and the insurance policies whose underwriting conclusions are extension or rejection are classified as group 3.

[0082] It can be understood that group 1 involves a list of terms for normal coverage; group 2 involves a list of terms for conditional coverage; and group 3 involves a list of terms for rejection of coverage.

[0083] In the training process of the first prediction model, group 1 and group 2 are used as positive samples, and group 3 is used as negative samples for training and continuous optimization. In other words, the first prediction model is a model obtained by training the word list of normal insurance, the word list of conditional insurance, and the word list of rejected insurance.

[0084] Specifically, the first prediction model is obtained by training a first positive sample and a first negative sample, the first positive sample includes a word list of normal underwriting and a word list of conditional underwriting, and the first negative sample includes a word list of rejected underwriting.

[0085] In the training process of the second prediction model, group 1 is used as a positive sample and group 2 is used as a negative sample for training and continuous optimization. In other words, the second prediction model is a model trained by the word list of normal insurance and the word list of rejected insurance.

[0086] Specifically, the second prediction model is trained by a second positive sample and a second negative sample, the second positive sample includes a word list of normal insurance coverage, and the first negative sample includes a word list of rejected insurance.

[0087] It should be noted that the training data of the second prediction model does not involve the word list of rejected insurance. This is because, based on previous business experience and modeling experience, the features of the word list of normal insurance and the word list of conditional insurance are very similar, while the gap between the word list of normal insurance and the word list of rejected insurance is large. Therefore, a better solution is to first use the first prediction model to distinguish the policies of normal insurance and conditional insurance from the rejected policies, and then use the second prediction model to distinguish the policies of normal insurance and conditional insurance. This can greatly improve the accuracy of model prediction.

[0088] The first prediction model and the second prediction model can be trained on the basis of the neural network model. Various neural network models can be trained on real data, and after comparing the model effects, the optimal model can be selected for the first prediction model and the second prediction model respectively. The model selection range includes but is not limited to convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory network (LSTM) and gated recurrent unit (GRU).

[0089] In the process of policy review, in order to determine the risk of the policy to be reviewed, the policy to be reviewed can be first input into the first prediction model to obtain the first prediction result of the policy to be reviewed. The first prediction model is trained by the word list of normal underwriting, the word list of conditional underwriting, and the word list of rejected insurance. Therefore, if the first prediction model can predict the result as a positive sample, the risk of the policy to be reviewed is medium risk or low risk; if the first prediction model can predict the result as a negative sample, the risk of the policy to be reviewed is high risk.

[0090] Specifically, if the first prediction result of the policy to be reviewed is a negative sample, it is determined that the policy to be reviewed is a high-risk policy; if the first prediction result of the policy to be reviewed is a positive sample, it is difficult to directly determine the risk of the policy to be reviewed, and further judgment is required in combination with the second prediction model.

[0091] S102. If the first prediction result of the policy to be reviewed is a positive sample, the policy to be reviewed is input into the second prediction model again. If the second prediction result of the policy to be reviewed is a negative sample, the policy to be reviewed is determined to be a medium-risk policy. If the second prediction result of the policy to be reviewed is a positive sample, the policy to be reviewed is determined to be a low-risk policy. The second prediction model is a model trained by the word list of normal underwriting and the word list of conditional underwriting.

[0092] It is known through the first prediction model that the first prediction result of the policy to be reviewed is a positive sample. In order to clarify whether the policy to be reviewed is a medium-risk policy or a low-risk policy, the policy to be reviewed needs to be input into the second prediction model again.

[0093] Since the second prediction model is trained based on the word list of normal underwriting and the word list of conditional underwriting, if the second prediction model can predict that the result is a positive sample, the risk of the policy to be reviewed is low risk; if the second prediction model can predict that the result is a negative sample, the risk of the policy to be reviewed is medium risk.

[0094] As an example, the risks of the pending policies are marked by labeling. Specifically, the pending policies are input into the first prediction model. If the first prediction result is a negative sample, it is labeled as "high-risk policy". If the first prediction result is a positive sample, the pending policies are input into the second prediction model. If the second prediction result output by the second prediction model is a negative sample, it is labeled as "medium-risk policy". If the second prediction result is a positive sample, it is labeled as "low-risk policy".

[0095] In one embodiment of the present invention, risk thresholds of the first prediction model and the second prediction model may be set separately according to underwriting risk control requirements to distinguish positive samples from negative samples.

[0096] In one embodiment of the present invention, the first prediction model is used to obtain the characteristics of the policy that is rejected, and the second prediction model is used to obtain the characteristics of the policy that is conditionally covered.

[0097] Specifically, according to the fitting process of the neural network model, the policy features are generated. The attention mechanism is introduced into the neural network, and different weights are assigned to the features and hidden features in the network through the Attention vector, and the weights are learned during the training process to obtain the importance of each feature for model learning. The first prediction model allows users to obtain the policy features of rejected insurance. The second prediction model is used for the policy features of conditional underwriting.

[0098] The attention mechanism adjusts the direction of attention and the weighted model, with the goal of improving the model effect. The attention mechanism is used for interpretable analysis of the model operation mode to determine which position's features contribute most to the results. In other words, the attention mechanism is used to enable the first prediction model to obtain the features of rejected policies. The second prediction model is used for the features of conditionally covered policies.

[0099] In addition, a gradient-based sensitivity method, an interpretable student model, or a hidden layer analysis method can be used to enable the first prediction model to obtain the characteristics of the policy that is rejected. The second prediction model is used for the characteristics of the policy that is conditionally covered.

[0100] S103. Determine the review method for the insurance policy to be reviewed based on the risk level of the insurance policy to be reviewed.

[0101] The handling methods for awaiting review of insurance policies with different risks are different. On the one hand, it can reduce the error rate of the review; on the other hand, it can improve the efficiency of the policy review.

[0102] See also Figure 3 , Figure 3 The present invention is a flowchart of determining a review method for a policy to be reviewed according to an embodiment of the present invention, which specifically includes the following steps:

[0103] S301. The risk level of the policy to be reviewed is low-risk policy, and automatic review is determined for the policy to be reviewed.

[0104] The risk level of the policy to be reviewed is low risk, which means that the underwriting risk of the policy to be reviewed is low. The policy review can be achieved through automatic review to improve the review efficiency.

[0105] S302: If the risk level of the policy to be reviewed is a high-risk policy or a medium-risk policy, manual review is determined for the policy to be reviewed.

[0106] The risk level of the policy to be reviewed is high-risk policy or medium-risk policy. In order to avoid losses, the policy to be reviewed is reviewed manually.

[0107] As an example, the management of pending policies based on the underwriting pool. After a new policy is insured, it enters the underwriting pool of different levels according to the risk level of the policy, and multiple underwriters select policies within the same underwriting pool. However, different underwriters have different underwriting experiences, and their proficiency in handling different policy cases varies greatly.

[0108] See also Figure 4 , Figure 4 The following is a flow chart of determining an underwriter according to an embodiment of the present invention, which specifically includes the following steps:

[0109] S401. Generate underwriting case characteristics of the underwriter based on the underwriter's historical underwriting policies.

[0110] Different levels of underwriting pools are pre-set, and each underwriting pool is used to store policies to be reviewed. As an example, the high-risk underwriting pool is used to store high-risk policies to be reviewed; the medium-risk underwriting pool is used to store medium-risk policies to be reviewed.

[0111] For underwriters, the underwriting case characteristics of the underwriters can be generated based on their historical underwriting policies.

[0112] As an example, extract keywords from the underwriter's historical underwriting documents. Figure 2 Then, the first prediction model is used to extract the features of the rejected policies in the historical insurance policies, and the second prediction model is used to extract the features of the conditionally covered policies in the historical insurance policies. The keywords, the features of the rejected policies, and the features of the conditionally covered policies are used as the features of the underwriter.

[0113] As another example, extract the disease types from the underwriter's historical underwriting policies. Accumulate the disease types that appear in each policy, divide the feature frequency by the total number of policies, and use it as the value of the underwriter under this disease type. For example, if underwriter A has 300 historical underwriting policies, and disease A, disease B, and disease C appear 2 times, 10 times, and 90 times, respectively, then the feature values ​​corresponding to disease A, disease B, and disease C are 0.0067, 0.033, and 0.33.

[0114] S402. Determine the underwriter for manual review based on the characteristics of the policy to be reviewed and the characteristics of the underwriter's underwriting cases.

[0115] According to the characteristics of the insurance policy to be reviewed and the characteristics of the underwriter's underwriting cases, a recommendation algorithm is used to determine the underwriter for manual review. The recommendation algorithm includes but is not limited to a collaborative filtering recommendation algorithm and a user behavior serialization recommendation algorithm. Among them, the first prediction model and the second prediction model can be used to obtain the characteristics of the insurance policy to be reviewed.

[0116] See also Figure 5 , Figure 5 Schematic diagram of the use of the core protection pool according to an embodiment of the present invention. Figure 5 It includes multiple underwriting pools, such as underwriting pool A, underwriting pool B, and underwriting pool C. In addition, it also includes multiple underwriters. As an example, underwriters are divided into junior underwriters, intermediate underwriters, and senior underwriters. The policies in the preset underwriting pool A are reviewed by junior underwriters; the policies in the preset underwriting pool B are reviewed by intermediate underwriters; and the policies in the preset underwriting pool C are reviewed by senior underwriters. Specifically, a recommendation algorithm can be used to determine the underwriters for manual review for the policies to be underwritten.

[0117] The following is a specific example to illustrate the technical solution in the embodiment of the present invention.

[0118] After the user purchases insurance, it is first determined whether the insured has any health disclosure content this time or in the past. If it is determined to be an abnormal health disclosure policy, the technical solution in the embodiment of the present invention is executed.

[0119] (1) Generate a vocabulary list

[0120] The trigger words and arguments of four types of events, namely, illness events, diagnosis and treatment events, physical examination events, and surgical events, are extracted from the health notice text of each insurance policy as the characteristics of the insurance policy to be reviewed. That is, key events include illness events, diagnosis and treatment events, physical examination events, and surgical events.

[0121] (2) Determine the policy risk

[0122] Input the characteristics of the insurance policy to be reviewed into the first prediction model; or input the characteristics of the insurance policy to be reviewed into the first prediction model and the second prediction model, thereby obtaining the risk level of the insurance policy to be reviewed.

[0123] (3) Determine the audit method

[0124] The review method is determined based on the risk level. Low-risk policies are transferred to the self-review process in real time; medium-risk or high-risk policies are determined by the underwriter based on the characteristics of the policy to be reviewed and the underwriter's underwriting case characteristics.

[0125] See also Figure 6 , Figure 6 is a schematic diagram of an underwriting display page according to an embodiment of the present invention. Figure 6 The insured person's information and notification are included. Figure 6 The underwriting display page can promptly provide feedback on the underwriting results to all owners of the policies to be reviewed.

[0126] In the above-mentioned embodiment of the present invention, the policy to be reviewed is input into the first prediction model. If the first prediction result of the policy to be reviewed is a negative sample, the policy to be reviewed is determined to be a high-risk policy, and the first prediction model is a model trained by a normal underwriting vocabulary, a conditional underwriting vocabulary, and a rejected underwriting vocabulary; if the first prediction result of the policy to be reviewed is a positive sample, the policy to be reviewed is input into the second prediction model again. If the second prediction result of the policy to be reviewed is a negative sample, the policy to be reviewed is determined to be a medium-risk policy, and if the second prediction result of the policy to be reviewed is a positive sample, the policy to be reviewed is determined to be a low-risk policy, and the second prediction model is a model trained by a normal underwriting vocabulary and a conditional underwriting vocabulary; based on the risk level of the policy to be reviewed, the review method is determined for the policy to be reviewed. Since the corresponding review method can be determined according to the risk level, the efficiency of reviewing the policy can be improved.

[0127] The solution adopted in the embodiment of the present invention has the following advantages: wide application range, serving more underwriters; building multiple models to output prediction results; automatically identifying case features and matching the best underwriter.

[0128] See also Figure 7 , Figure 7 is a schematic diagram of the main structure of the device for policy review according to an embodiment of the present invention. The device for policy review can implement a method for policy review, such as Figure 7 As shown, the device for policy review specifically includes:

[0129] The initial screening module 701 is used to input the to-be-reviewed insurance policy into a first prediction model. If the first prediction result of the to-be-reviewed insurance policy is a negative sample, the to-be-reviewed insurance policy is determined to be a high-risk insurance policy. The first prediction model is a model trained by a normal underwriting word list, a conditional underwriting word list, and a rejected insurance word list;

[0130] A re-screening module 702 is used for inputting the policy to be reviewed into a second prediction model again if the first prediction result of the policy to be reviewed is a positive sample, determining that the policy to be reviewed is a medium-risk policy if the second prediction result of the policy to be reviewed is a negative sample, and determining that the policy to be reviewed is a low-risk policy if the second prediction result of the policy to be reviewed is a positive sample, wherein the second prediction model is a model obtained by training a word list of normal underwriting and a word list of conditional underwriting;

[0131] The review module 703 is used to determine the review method for the insurance policy to be reviewed based on the risk level of the insurance policy to be reviewed.

[0132] In one embodiment of the present invention, the initial screening module 701 is further used to construct a key event word list for text annotation based on key events, wherein the key event is a collection of health-related information in the text;

[0133] The plurality of key event word lists are divided into the normal underwriting word list, the conditional underwriting word list and the rejected underwriting word list according to the underwriting situation.

[0134] In one embodiment of the present invention, the initial screening module 701 is specifically used to construct a key event vocabulary list for text annotation based on the verbs in the key events and the synonyms of the verbs, wherein the synonyms of the verbs include one or more of the following: synonyms of the characters in the verbs, convertible forms of the verbs, similar words to the verbs in the hospital semantic dictionary, and words with the same meaning as the verbs in the hospital semantic dictionary.

[0135] In one embodiment of the present invention, the initial screening module 701 is specifically used to construct a key event vocabulary list for text annotation based on the verbs in the key events, the synonyms of the verbs, the sentence features and the position features in the key events, the sentence vector is used to identify the global semantic information of the text including the key time, and the position vector is used to identify the semantic information carried by words in different positions.

[0136] In one embodiment of the present invention, the first prediction model is used to obtain characteristics of insurance policies that are rejected; the second prediction model is used to obtain characteristics of insurance policies that are conditionally covered;

[0137] The first prediction model and the second prediction model are trained based on a neural network model.

[0138] In one embodiment of the present invention, the first prediction model adopts a preset method to obtain the characteristics of the rejected insurance policy, and the second prediction model adopts the preset method to obtain the characteristics of the conditionally covered insurance policy, and the preset method includes one of the following: attention mechanism, gradient-based sensitivity method, interpretable student model and analysis of hidden layer.

[0139] In one embodiment of the present invention, the review module 703 is specifically used to determine automatic review for the insurance policy to be reviewed when the risk level of the insurance policy to be reviewed is a low-risk insurance policy;

[0140] The risk level of the policy to be reviewed is a high-risk policy or a medium-risk policy, and manual review is determined for the policy to be reviewed.

[0141] In one embodiment of the present invention, the review module 703 is specifically used to generate the underwriting case characteristics of the underwriter based on the underwriter's historical underwriting policies;

[0142] The underwriter for manual review is determined based on the characteristics of the insurance policy to be reviewed and the characteristics of the underwriter's underwriting cases.

[0143] Figure 8 An exemplary system architecture 800 is shown to which the method or apparatus for policy review according to an embodiment of the present invention may be applied.

[0144] like Figure 8 As shown, system architecture 800 may include terminal devices 801, 802, 803, a network 804, and a server 805. Network 804 is used to provide a medium for communication links between terminal devices 801, 802, 803 and server 805. Network 804 may include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

[0145] Users can use terminal devices 801, 802, and 803 to interact with server 805 through network 804 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 801, 802, and 803, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only examples).

[0146] The terminal devices 801 , 802 , and 803 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0147] The server 805 may be a server that provides various services, such as a backend management server (only an example) that provides support for shopping websites browsed by users using the terminal devices 801, 802, and 803. The backend management server may analyze and process the received data such as product information query requests, and feed back the processing results (such as target push information, product information - only an example) to the terminal device.

[0148] It should be noted that the insurance policy review method provided in the embodiment of the present invention is generally executed by the server 805 , and accordingly, the insurance policy review device is generally set in the server 805 .

[0149] It should be understood that Figure 8 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0150] Reference below Fig. 9 , which shows a schematic diagram of the structure of a computer system 900 of a terminal device suitable for implementing an embodiment of the present invention. Fig. 9The terminal device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0151] like Fig. 9 As shown, the computer system 900 includes a central processing unit (CPU) 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage part 908 into a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the system 900 are also stored. The CPU 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0152] The following components are connected to the I / O interface 905: an input section 906 including a keyboard, a mouse, etc.; an output section 907 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, a modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the I / O interface 905 as needed. A removable medium 911, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 910 as needed, so that a computer program read therefrom is installed into the storage section 908 as needed.

[0153] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 909, and / or installed from the removable medium 911. When the computer program is executed by the central processing unit (CPU) 901, the above-mentioned functions defined in the system of the present invention are executed.

[0154] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0155] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0156] The modules involved in the embodiments of the present invention may be implemented by software or hardware. The modules described may also be arranged in a processor. For example, they may be described as follows: a processor includes a primary screening module, a secondary screening module, and an audit module. The names of these modules do not, in some cases, constitute limitations on the modules themselves. For example, the audit module may also be described as "used to determine the audit method for the policy to be audited based on the risk level of the policy to be audited."

[0157] As another aspect, the present invention further provides a computer-readable medium, which may be included in the device described in the above embodiment; or may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device includes:

[0158] Input the to-be-reviewed insurance policy into a first prediction model, and if the first prediction result of the to-be-reviewed insurance policy is a negative sample, determine that the to-be-reviewed insurance policy is a high-risk insurance policy, wherein the first prediction model is a model trained by a word list of normal underwriting, a word list of conditional underwriting, and a word list of rejected insurance;

[0159] If the first prediction result of the policy to be reviewed is a positive sample, the policy to be reviewed is input into the second prediction model again, if the second prediction result of the policy to be reviewed is a negative sample, the policy to be reviewed is determined to be a medium-risk policy, and if the second prediction result of the policy to be reviewed is a positive sample, the policy to be reviewed is determined to be a low-risk policy, and the second prediction model is a model trained by a word list of normal underwriting and a word list of conditional underwriting;

[0160] Based on the risk level of the insurance policy to be reviewed, a review method is determined for the insurance policy to be reviewed.

[0161] According to the technical solution of the embodiment of the present invention, the insurance policy to be reviewed is input into the first prediction model. If the first prediction result of the insurance policy to be reviewed is a negative sample, the insurance policy to be reviewed is determined to be a high-risk insurance policy, and the first prediction model is a model obtained by training the word list of normal insurance, the word list of conditional insurance, and the word list of rejected insurance; if the first prediction result of the insurance policy to be reviewed is a positive sample, the insurance policy to be reviewed is input into the second prediction model again. If the second prediction result of the insurance policy to be reviewed is a negative sample, the insurance policy to be reviewed is determined to be a medium-risk insurance policy, and if the second prediction result of the insurance policy to be reviewed is a positive sample, the insurance policy to be reviewed is determined to be a low-risk insurance policy, and the second prediction model is a model obtained by training the word list of normal insurance and the word list of conditional insurance; based on the risk level of the insurance policy to be reviewed, the review method is determined for the insurance policy to be reviewed. Since the corresponding review method can be determined according to the risk level, the efficiency of reviewing insurance policies can be improved.

[0162] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for reviewing an insurance policy, characterized in that: include: Input the to-be-reviewed insurance policy into a first prediction model. If the first prediction result of the to-be-reviewed insurance policy is a negative sample, the to-be-reviewed insurance policy is determined to be a high-risk insurance policy. The first prediction model is a model obtained by training a word list of normal insurance, a word list of conditional insurance, and a word list of rejected insurance. The first prediction model is obtained by training a first positive sample and a first negative sample. The first positive sample includes a word list of normal insurance and a word list of conditional insurance, and the first negative sample includes a word list of rejected insurance. If the first prediction result of the to-be-reviewed policy is a positive sample, the to-be-reviewed policy is input into the second prediction model again, if the second prediction result of the to-be-reviewed policy is a negative sample, it is determined that the to-be-reviewed policy is a medium-risk policy, and if the second prediction result of the to-be-reviewed policy is a positive sample, it is determined that the to-be-reviewed policy is a low-risk policy, the second prediction model is a model trained by a word list of normal underwriting and a word list of conditional underwriting; the second prediction model is trained by a second positive sample and a second negative sample, the second positive sample includes the word list of normal underwriting, and the second negative sample includes the word list of conditional underwriting; Based on the risk level of the insurance policy to be reviewed, a review method is determined for the insurance policy to be reviewed.

2. The method for policy review according to claim 1, characterized in that: Before inputting the pending insurance policy into the first prediction model, the method further includes: Based on key events, a key event word list for text annotation is constructed, wherein the key events are a collection of health-related information in the text; The plurality of key event word lists are divided into the normal underwriting word list, the conditional underwriting word list and the rejected underwriting word list according to the underwriting situation.

3. The method for policy review according to claim 2, characterized in that: The key event word list for text annotation is constructed based on the key events, including: Based on the verbs in key events and the synonyms of the verbs, a key event word list for text annotation is constructed, wherein the synonyms of the verbs include one or more of the following: synonyms of the characters in the verbs, convertible forms of the verbs, similar words to the verbs in the hospital semantic dictionary, and words with the same meaning as the verbs in the hospital semantic dictionary.

4. The method for policy review according to claim 2, characterized in that: The key event word list for text annotation is constructed based on the key events, including: Based on the verbs in key events, the synonyms of the verbs, the sentence features and the position features in the key events, a key event vocabulary list for text annotation is constructed. The sentence vector is used to identify the global semantic information of the text including the key events, and the position vector is used to identify the semantic information carried by words in different positions.

5. The method for policy review according to claim 1, characterized in that: The first prediction model is used to obtain the characteristics of the policy that is rejected; the second prediction model is used to obtain the characteristics of the policy that is conditionally covered; The first prediction model and the second prediction model are trained based on a neural network model.

6. The method for policy review according to claim 5, characterized in that: The first prediction model uses a preset method to obtain the characteristics of the rejected insurance policies, and the second prediction model uses the preset method to obtain the characteristics of the conditionally covered insurance policies, and the preset method includes one of the following: attention mechanism, gradient-based sensitivity method, interpretable student model and analysis of hidden layers.

7. The method for policy review according to claim 1, characterized in that: The step of determining a review method for the insurance policy to be reviewed based on the risk level of the insurance policy to be reviewed includes: The risk level of the insurance policy to be reviewed is a low-risk insurance policy, and automatic review is determined for the insurance policy to be reviewed; The risk level of the policy to be reviewed is a high-risk policy or a medium-risk policy, and manual review is determined for the policy to be reviewed.

8. A device for reviewing insurance policies, characterized in that: include: The initial screening module is used to input the to-be-reviewed insurance policy into a first prediction model. If the first prediction result of the to-be-reviewed insurance policy is a negative sample, the to-be-reviewed insurance policy is determined to be a high-risk insurance policy. The first prediction model is a model trained by a word list of normal insurance, a word list of conditional insurance, and a word list of rejected insurance. The first prediction model is trained by a first positive sample and a first negative sample. The first positive sample includes a word list of normal insurance and a word list of conditional insurance, and the first negative sample includes a word list of rejected insurance. a re-screening module, for inputting the policy to be reviewed into a second prediction model again if the first prediction result of the policy to be reviewed is a positive sample, determining that the policy to be reviewed is a medium-risk policy if the second prediction result of the policy to be reviewed is a negative sample, and determining that the policy to be reviewed is a low-risk policy if the second prediction result of the policy to be reviewed is a positive sample, wherein the second prediction model is a model trained by a word list of normal underwriting and a word list of conditional underwriting; the second prediction model is trained by a second positive sample and a second negative sample, wherein the second positive sample includes the word list of normal underwriting and the second negative sample includes the word list of conditional underwriting; The audit module is used to determine the audit method for the insurance policy to be audited based on the risk level of the insurance policy to be audited.

9. An electronic device for policy review, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images