NL2033995B1

NL2033995B1 - Computer implemented method for insurance underwriting

Info

Publication number: NL2033995B1
Application number: NL2033995A
Authority: NL
Inventors: Te Nicolas; Jouili Salim; Grumiau Christopher; Sethia Jain Amit; Doumen Jan; Thoppan Mohanchandralal Sudaman
Original assignee: Allianz Benelux Nv
Priority date: 2022-01-24
Filing date: 2023-01-20
Publication date: 2023-08-04
Also published as: BE1029657B1

Abstract

Example embodiments describe a computer implemented method for underwriting an insurance policy. The method comprises, receiving an unstructured electronic data message comprising a request for underwriting the insurance policy; extracting structured underwriting data from the unstructured electronic data message by means of natural language processing; and generating at least one query based on the structured underwriting data for querying a knowledge graph. The method further comprises obtaining a selection of entities from the knowledge graph by querying the knowledge graph with the at least one query; identifying one or more peers of the insurance policy from a portfolio of insurance policies based on the selection of entities, wherein a peer is characterized by a set of entities substantially similar to the selection of entities; and determining a loss ratio for underwriting the insurance policy based on the loss ratios of the one or more peers.

Description

COMPUTER IMPLEMENTED METHOD FOR INSURANCE UNDERWRITING

Field of the Invention

[01] The present invention generally relates to, amongst others, insurance underwriting. In particular, the invention relates to automated insurance underwriting of customized insurance policies.

Background of the Invention

[02] Underwriting an insurance policy relates to analysing a risk to be transferred from a prospective policyholder to an insurer and determining an appropriate premium to transfer that risk. Successful underwriting relies on collecting relevant data associated with the risk such that an underwriter can make an informed underwriting decision. The quality of the underwriting decision thus depends on the exhaustiveness of the collected data, the relevance of the data, and the risk analysis.

[03] In common cases, e.g. home or car insurance, the underwriting process is straightforward as the relevant data for assessing and analysing the risk is substantially similar for all cases. As such, the underwriting process for common cases can be automated relatively easily. In specific cases, e.g. customized insurance policies for small and medium-sized enterprises, SMEs, case-specific data is required to analyse the risk. This makes collecting, assessing, and selecting data time-consuming. Additionally, the vast amount of available data and diversification of data formats makes it increasingly difficult for underwriters to collect, interpret, and select all relevant data.

[04] A problem with automating the underwriting process of customized insurance policies is processing the diversity of policy requests as each case is substantially unique. Another problem is the limited amount of available training data as the amount of specific cases that handle customized insurance policies can be substantially limited. Another problem is collecting the relevant underwriting data as the risk-determining factors can be substantially different for each specific case. Yet another problem is that the risk analysis and underwriting decision depend on the experience and expertise of the underwriter. Therefore, simply automating a human workflow and decision taking process is not possible.

Summary of the Invention

[05] It is an object of the present invention, amongst others, to solve or alleviate the above identified challenges and problems by at least partially automating the underwriting of insurance policies, in particular of customized insurance policies.

[06] According to a first aspect, this object is achieved by a computer implemented method for underwriting an insurance policy. The method comprises, receiving an unstructured electronic data message comprising a request for underwriting the insurance policy, extracting structured underwriting data from the unstructured electronic data message by means of natural language processing, comprising at least an identification of a policyholder, an insured object, and a desired coverage; and generating at least one query based on the structured underwriting data for querying a knowledge graph, wherein the knowledge graph is a network of entities comprising one or more labels and their respective relationships. The method further comprises obtaining a selection of entities from the knowledge graph by querying the knowledge graph with the at least one query, wherein an entity of the selection has a relationship with the structured underwriting data and comprises one or more predetermined labels; identifying one or more peers of the insurance policy from a portfolio of insurance policies based on the selection of entities, wherein a peer is characterized by a set of entities substantially similar to the selection of entities; and determining a loss ratio for underwriting the insurance policy based on the loss ratios of the one or more peers.

[07] Underwriting an insurance policy can refer to accepting a new insurance policy or managing an existing insurance policy, e.g. renewing or adjusting an existing insurance policy. The loss ratio can be expressed as the ratio of paid claims to collected premiums for an insurance policy over a time period, e.g. one year. The determined loss ratio can therefore allow to determine a suitable premium for the insurance policy and whether the risk is acceptable, even when the insurance policy relates to a substantially unique case.

[08] To this end, structured underwriting data is extracted by natural language processing from an unstructured electronic data message, e.g. an e-mail, a text message, or an instant message, comprising a request for underwriting an insurance policy. The structured underwriting data can be keywords, tokens, or values associated with an identification of a policyholder, an insured object, or a desired coverage. As such, interpreting a substantially unique request for underwriting an insurance policy can be automated. It is an advantage that interpreting the request can be automated without using standardized forms, as the appropriate underwriting data in each request can vary substantially in customized insurance policies.

[09] Based on this structured underwriting data a knowledge graph is queried comprising a network of real-world entities, e.g. a policyholder, a company, or an asset, and the relationships between them. An entity is further characterized by one or more labels, e.g. a category, an entity type, or a risk level. Querying the knowledge graph obtains entities that have a relationship with the structured underwriting data and that are characterized by one or more predetermined labels. In other words, entities relevant to the risk of an insurance policy are obtained, i.e. the selection of entities. It is an advantage that the knowledge graph makes collecting and selecting relevant data associated with the risk of an insurance policy more efficient and more exhaustive. It is a further advantage that unknown relevant entities can be discovered by exploring the relationships between entities in the knowledge graph, which can result in a better assessment of the underwriting risk and can thus enhance the underwriting process.

[10] By the selection of entities, one or more peers of the insurance policy are identified from a portfolio of insurance policies, i.e. a collection of previous and current insurance policies underwritten by the insurer. A peer of the insurance policy is an insurance policy in the portfolio that is characterized by one or more similar entities. As such, peers of an insurance policy can be identified even tough the insurance policies are substantially unique. The loss ratio for underwriting the insurance policy is then determined based on the loss ratios of the respective peers.

[11] Thus, the above method substantially differs from a mere automatization of the manual underwriting process as performed by an underwriter, in that it provides a machine architecture that allows the underwriting process to be performed in a different and improved way.

[12] According to an embodiment, extracting the structured underwriting data can be performed by a classifier trained to extract structured underwriting data from an unstructured electronic data message.

[13] A natural language processing algorithm, i.e. the classifier, can thus be trained by machine learning to extract structured keywords from the unstructured electronic data message. The classifier can for example be, amongst others, a neural network, a support vector machine, a Bayesian network, a conditional random field, or a maximum entropy model. Such a classifier can be trained by supervised learning, wherein a plurality of unstructured electronic data messages comprising annotated information on structured underwriting data are presented to the classifier.

Alternatively or complementary, the classifier can be pre-trained on a benchmark dataset, trained by unsupervised learning, or trained by reinforcement learning.

[14] According to an embodiment, extracting the structured underwriting data can further comprise correcting errors in the extracted structured underwriting data by comparing erroneous structured underwriting data with other structured underwriting data.

[15] An error in structured underwriting data can be corrected by matching the erroneous underwriting data to correct underwriting data describing the same characteristic feature of an entity, e.g. the name of a policyholder. Alternatively, an error in structured underwriting data can be corrected by comparing the meaning of the erroneous underwriting data to the meaning of correct underwriting data describing the same entity, e.g. by comparing the name and the national registration number of a policyholder. Alternatively or complementary, structured underwriting data can be checked for correctness by searching databases and/or by performing approximate string matching. This correcting can be performed substantially after or during the extracting of structured underwriting data. This has the further advantage that the extracting of underwriting data can be more reliable.

[16] According to an embodiment, one or more labels of the respective entities in the knowledge graph can be determined based on an observed action history of underwriters.

[17] An action history of an underwriter refers to the actions an underwriter performs when collecting and/or selecting data associated with the risk of an insurance policy. These actions can be observed by monitoring and logging the actions of underwriters in a computer application. This allows to label entities in the knowledge graph based on these observed actions, e.g. by adding labels relating to the risk of an entity as perceived by an underwriter. As such, querying the knowledge graph can result in a selection of entities that substantially matches the selection made by underwriters. It is thus an advantage that the experience and expertise of underwriters can be embedded in the knowledge graph.

[18] According to an embodiment, the method can further comprise obtaining underwriter feedback on one or more entities in the selection of entities, and adjusting the labels of the respective entities in the knowledge graph according to the feedback.

[19] The selection of entities can be exposed to an underwriter after querying the knowledge graph, e.g. in a computer application, to enhance and/or redact the selection. The actions performed by the underwriter, e.g. omitting entities from the selection or re-querying the knowledge graph for additional information, can be monitored and logged. The labels of the entities in the knowledge graph can then be adjusted according to the logged actions of the underwriter, i.e. the underwriter feedback. It is thus an advantage that the knowledge graph can be continuously improved by embedding the underwriter feedback.

[20] According to an embodiment, determining the loss ratio for underwriting the insurance policy can be based on a probability distribution of the loss ratios of the one or more peers.

[21] By the probability distribution, the most probable loss ratio for the insurance policy can be determined. It is thus an advantage that the loss ratio for underwriting an insurance policy can be determined even if the prospective policyholder is unknown to the insurer, i.e. has not previously taken out an insurance policy with the insurer.

[22] According to an embodiment, the method can further comprise determining a premium for the insurance policy based on the determined loss ratio for underwriting the insurance policy.

[23] The premium can directly be determined from the determined loss ratio, i.e. the loss ratio for underwriting the insurance policy that is determined based on the loss ratios of the one or more peers. The determined premium can further be expanded with a correction factor, i.e. a safety margin to correct for prediction errors in the determined loss ratio.

[24] According to an embodiment, the method can further comprise predicting a state of the portfolio of insurance policies when underwriting the insurance policy based on the loss ratio for underwriting the insurance policy, previous states of the portfolio, and previous underwritten insurance policies of the portfolio.

[25] In other words, the effect of underwriting the insurance policy on the global insurance portfolio of an insurer can be predicted. The portfolio state can be characterized by one or more performance metrics such as, for example, expense ratio, retention rate, average claim cost, claim frequency, global loss ratio of the portfolio, a gross sum of insured damages, market share, diversification, and profit margins. This can allow to optimize the portfolio by adjusting the insurance policy according to the predicted state of the portfolio. It is a further advantage that the global portfolio of insurance policies is considered in the underwriting process of an individual insurance policy.

[26] According to an embodiment, the method can further comprise extracting structured cost data from previous unstructured claims by natural language processing, and associating the structured cost data to terms and conditions of the insurance policy for which an unstructured claim was filed.

[27] Structured cost data can thus be extracted from a history of claims, i.e. previous unstructured claims. Structured cost data can include claim elements and the cost associated to the respective claim elements. In other words, the structured cost data allows to link elements in filed claims to their respective cost. By comparing the claim elements with the terms and conditions of the respective insurance policies for which the claim was filed, a cost can be associated to elements of the terms and conditions.

[28] According to an embodiment, the method can further comprise determining the terms and conditions for the insurance policy based on the structured cost data and the determined loss ratio for underwriting the insurance policy.

[29] By the cost associated to elements of the terms and conditions, the terms and conditions can be adjusted to optimize the loss ratio of the insurance policy, i.e. reduce the loss ratio. It is an advantage that the exposure of an insurer can be controlled by excluding high-cost events in the terms and conditions of an insurance policy.

[30] According to an embodiment, the method can further comprise parsing structured data and/or scraping unstructured data associated with the insurance policy from at least one data source to encode it in the knowledge graph.

[31] The at least one data source can be an internal data source, e.g. data stored on private servers, or an external data source, e.g. data publicly available on the

Internet. A web crawler can be implemented to systematically browse the World Wide

Web to collect structured and/or unstructured data. The collected data can be encoded in the knowledge graph to improve the selection of entities and thus the underwriting process. This has the further advantage that collecting and interpreting data from a variety of data sources is more efficient.

[32] According to an embodiment, the unstructured data can be structured by means of natural language processing and/or computer vision.

[33] The unstructured data can for example be, amongst others, text, audio, pictures, and video. Natural language processing can be used to structure unstructured text and audio, for example respectively by natural language understanding and speech recognition. Computer vision can be used to structure unstructured pictures and video.

[34] According to a second aspect, the invention relates to a data processing system configured to perform the computer implemented method according to the first aspect.

[35] According to a third aspect, the invention relates to a computer program comprising instructions which, when the program is executed by a computer, cause the computer to perform the computer implemented method according to the first aspect.

[36] According to a fourth aspect, the invention relates to a computer-readable medium comprising instructions which, when executed by a computer, cause the computer to perform the computer implemented method according to the first aspect.

Brief Description of the Drawings

[37] Fig. 1 shows an example of an underwriting process for underwriting an insurance policy;

[38] Fig. 2 shows steps according to a computer implemented method for underwriting an insurance policy in an automated way according to an embodiment;

[39] Fig. 3 shows steps to encode unstructured data and structured data extracted from at least one data source into a knowledge graph according to embodiments;

[40] Fig. 4 shows a supervised training process of a classifier to extract structured underwriting data from an unstructured electronic data message comprising a request for underwriting an insurance policy according to an embodiment; and

[41] Fig. 5 shows an example embodiment of a suitable computing system for performing steps according to example aspects of the invention.

Detailed Description of Embodiment(s)

[42] Fig. 1 shows an example of an underwriting process 100 for an underwriting an insurance policy. Underwriting an insurance policy relates to analysing a risk to be transferred from a prospective policyholder 101 to an insurer 105 and determining an appropriate premium to transfer that risk. Underwriting an insurance policy can refer to accepting a new insurance policy or managing an existing insurance policy, e.g. reviewing, renewing, or adjusting an existing insurance policy. A premium can be a cost that an insurer periodically charges to a policyholder to assume the risk of the insurance policy.

[43] A prospective policyholder 101, e.g. a small and medium-sized enterprise,

SME, can instruct an insurance broker 103 to negotiate an insurance policy for an insured object. The insured object may for example be a building, a machine, a ship, a vehicle, computer equipment, or any other asset. To this end, the prospective policyholder 101 may provide 102 the appropriate underwriting information to the broker 103, e.g. information on the insured object, a desired coverage, information on existing insurance policies on the insured object, or a desired premium. The broker 103 may adapt the provided underwriting information 102 to generate a request 104 for underwriting an insurance policy. The request 104 may be an unstructured electronic data massage, e.g. an e-mail, a text message, or an instant message, that is communicated to an insurer 105. The insurer 105 may be any party, e.g. an insurance company, that undertakes to pay compensation to the policyholder 101 in the event of a loss incurred on the insured object in exchange for a predetermined fee, i.e. a premium.

[44] Typically, an expert underwriter first analyses 106 the received request 104 for underwriting an insurance policy on behalf of the insurer 105. This analysis 106 includes extracting underwriting information from the request 104 such as, for example, information on the prospective policyholder 101, information on the broker 103, information on the insured object, and information on the desired coverage.

Based on the extracted underwriting information, the underwriter can then search, collect 107, and select 108 data or information that is relevant for assessing the risk of underwriting the insurance policy. In a following step 109, the underwriter may use his/her expertise to determine the risk of underwriting the insurance policy based on the selected data.

[45] In a following step 110, the underwriter can determine whether the risk is acceptable based on the relevant data obtained during steps 107 and 108, and the risk analysis performed during step 109. If the risk is determined to be unacceptable, underwriting of the insurance policy is refused. If the risk is determined to be acceptable, a policy proposal may be determined 112. Determining a policy proposal 112 may for example include determining a premium, terms, and conditions for the requested insurance policy. Alternatively or complementary, determining a policy proposal 112 may include granting the policy proposal as requested 104 by the broker 103. In a final step, the proposal or refusal can be communicated 111, 113 to the prospective policyholder 101 via the broker 103.

[46] Fig. 2 shows steps 200 according to a computer implemented method for underwriting an insurance policy in an automated way according to an embodiment.

The method illustrated in Fig. 2 can for example automate, or partially automate, the underwriting process illustrated in Fig. 1. In a first step 201, an unstructured electronic data message 211 is received. Such an unstructured electronic data message 211 can for example be, amongst others, an e-mail, a text message, or an instant message. The unstructured electronic data message 211 comprises a request for underwriting an insurance policy, e.g. for an insured object of a small and medium-sized enterprise, SME. In particular, the requested insurance policy may be a customized insurance policy that may be shared between a plurality of insurers. In other words, the insured object may partially be insured by a plurality of insurers. The request can include underwriting information on the requested insurance policy such as, for example, an identification of the policyholder, information on the insured object, and the desired coverage.

[47] In a following step 202, this underwriting information is extracted as structured underwriting data 212 by means of natural language processing. The structured underwriting data 212 can for example be keywords, tokens, values, or subsections in the unstructured electronic data message 211. Natural language processing may for example include named entity recognition, also named entity identification, entity chunking, or entity extraction, wherein underwriting data 212 is located in an unstructured text and classified into pre-defined categories, ie. structured. The structured underwriting data 212 comprises at least an identification of a policyholder, an insured object, and a desired coverage. Identification of a policyholder may for example be a name, an address, or a VAT identification number of the policyholder.

Identification of an insured object may for example be an identification of a building, a machine, a ship, a vehicle, computer equipment, or any other asset that is covered by the insurance policy. A desired coverage may describe the risk, liability, or potential loss that is covered by the insurance policy.

[48] As such, interpreting a substantially unique request for underwriting an insurance policy can be automated. It is an advantage that the request for underwriting an insurance policy can be substantially automated without using standardized forms, as the appropriate underwriting information in each request can vary substantially between customized insurance policies.

[49] Ina next step 203, at least one query 213 is generated based on the extracted underwriting data 212 to query a knowledge graph 204. A knowledge graph 204, also semantic network, represents a network of real-world entities 221, 222, 223 and the relationships 231, 232 between them. An entity 221, 222, 223 can for example be a policyholder, a company, a name, a number, an address, an asset, financial information, a vessel, a broker, a coverage, a policy, an insured object, a contract, a claim, a location, or any other real-world object or event. An entity 221, 222, 223 is further characterized by one or more labels, e.g. a category, an entity type, or a risk level. Relationships 231, 232 between entities can for example be, amongst others, "located in’, 'owned by’, 'insured by', ‘operates in’, or 'negotiating with’. The knowledge graph 204 can be represented as a network of nodes 221, 222, 223 and edges 231, 232, respectively representing the entities and the directional relationships between these nodes. The knowledge graph 204 can for example be stored in a graph database or NoSQL database.

[60] The knowledge graph 204 allows to integrate high-dimensional information extracted from multiple structured and/or unstructured data sources and allows to encode the meaning of this data. Querying the knowledge graph 204 allows to discover how entities 221, 222, 223 from different parts of the information domain relate to each other. Querying the knowledge graph 204 further allows to uncover hidden or previously unknown knowledge in the knowledge graph 204. Querying the knowledge graph 204 can for example be performed by a latent feature model, a graph feature model, and/or by graph-computing techniques such as shortest path computations and network analysis. As such, the at least one query 213 can be used to search and discover information encoded in the knowledge graph 204 that relates to the underwriting data 212, i.e. the requested insurance policy.

[61] Querying the knowledge graph results in a collection of entities 214 that are related to the structured underwriting data 212, i.e. the requested insurance policy. In a following step 205, a selection of entities 215 can be obtained by selecting the entities from the collection of entities 214 that comprise one or more predetermined labels. In other words, the entities in the collection of entities 214 that are not labelled with the one or more predetermined labels are omitted from the selection 215.

Alternatively, this selecting can be performed while querying the knowledge graph 204, e.g. by adding the predetermined labels as a condition to the query. The predetermined labels can be indications that an entity 221, 222, 223 is relevant to the risk of an insurance policy. For example, entities 221, 222, 223 labelled with the entity type bankruptcy, default, or labelled with a high risk level. In doing so, entities 221, 222, 223 and relationships 231, 232 that are relevant to the risk of an insurance policy are obtained, i.e. the selection of entities 215. It is an advantage that unknown relevant entities can be discovered by exploring the relationships between entities in the knowledge graph 204, which can result in a better assessment of the underwriting risk and can thus enhance the underwriting process.

[62] In a next step 207, one or more peers 216 of the insurance policy are identified from a portfolio of insurance policies 206 based on the selection of entities 215. The portfolio of insurance policies 206 is a collection of previous and current, i.e. active, insurance policies underwritten by the insurer. A peer of the insurance policy is an insurance policy that is characterized by one or more similar entities, e.g. an insurance policy with similar coverage, a similar policyholder, a similar geographical location, or a similar line of business. In other words, by the obtained selection of entities 215 one or more insurance policies that are substantially similar to the insurance policy to be underwritten can be identified from the portfolio 206. The portfolio 206 data can be stored in a database or a knowledge base.

[63] A peer 216 can be identified by matching one or more entities from the selection 215 to data of the insurance policies in the portfolio 206. Alternatively, this portfolio data can be encoded in knowledge graph 204. By identifying peers 216 based on the similarity between one or more entities, peers of an insurance policy can be identified even tough the insurance policies are substantially unique such as, for example, in customized policies for SMEs. In other words, this allows to identify peers 216 even if the insurance policy to be underwritten is substantially unprecedented.

[54] In a final step 208, a loss ratio for underwriting the insurance policy can be determined based on the loss ratios of the one or more peers 217. The loss ratio is a cost that can be expressed as the ratio of paid claims to the collected premiums of an insurance policy over a certain period, e.g. a month, a year, or a decade. The loss ratios of the peers 217 may be obtained from the database or the knowledge base wherein the portfolio 206 is stored, e.g. by analysing the claim history of the peers 216.

[65] Determining 208 the loss ratio of the insurance policy can further be based on a probability distribution of the loss ratios of the peers 217. By the probability distribution, the most probable loss ratio for the insurance policy can be determined or predicted. For example, the loss ratio 218 can be determined 208 to be the average loss ratio of its peers 216. It is thus an advantage that the loss ratio 218 for underwriting an insurance policy can be determined even if the prospective policyholder has not previously taken out any insurance policies with the insurer. The determined loss ratio 218 further allows to determine whether the risk is acceptable.

This can be performed manually, e.g. by providing the determined loss ratio to an underwriter, or can be performed automatically, e.g. by a rule-based algorithm.

[66] Additionally, a premium for the insurance policy can be determined based on the determined loss ratio 218. The premium can directly be determined from the loss ratio 218, e.g. by dividing the average claim amount of the peers 216 by the average loss ratio 218 of the peers 216. The determined premium can further be expanded with a correction factor, i.e. a safety margin to correct for prediction errors in the determined loss ratio 218. As such, the determined loss ratio 218 allows to determine a suitable premium for the insurance policy and whether the risk is acceptable, even when an insurance policy relates to a substantially unique case.

[57] Thus, the above described method substantially differs from a mere automatization of the manual underwriting process as performed by an underwriter, in that it provides a machine architecture that allows the underwriting process to be performed in a different and improved way. In other words, the method allows to underwrite an insurance policy in a substantially different way, i.e. not just by executing the performed steps of an underwriter on a computer.

[58] Fig. 3 shows steps 300 to encode unstructured data 311 and structured data 312 extracted from at least one data source 310 into the knowledge graph 204.

Knowledge graph 204 can thus be obtained by parsing structured data 312 and/or scraping unstructured data 311 associated with the insurance policy from at least one data source 310, preferably a plurality of data sources. The data sources 310 can be internal data sources such as data stored on private servers, or the data sources 310 can be external data sources such as data publicly available on the Internet. A web crawler, also spider or spiderbot, can be implemented to systematically browse the

World Wide Web to collect structured 312 and/or unstructured data 311.

[59] The unstructured data 311 can for example be, amongst others, text, audio, pictures, and video. Natural language processing can be implemented for extracting and structuring 320 data from unstructured text and audio. For example, natural language understanding can be used to extract and structure text data, while speech recognition can be used to extract and structure audio data. Computer vision algorithms can be implemented for extracting and structuring 320 data from unstructured pictures and video.

[60] The structured data 312 and its meaning can then be encoded 330 in the knowledge graph 204 by assigning an entity identifier 331 and one or more relationships 332 to an entity, i.e. structured data 312. As such, the structured data 312 and its meaning can be mapped on nodes 221, 222, 223 and edges 231, 232 in the knowledge graph 204. In other words, the structured data 412 can be used to construct interconnected facts comprising a subject, predicate, and object that can be stored in a knowledge database, graph database, or NoSQL database. An entity 221, 222, 223 can further be characterized by one or more labels 333, e.g. a category, an entity type, a risk level, or a relevance to underwriting an insurance policy.

[61] These labels 333 can be inferred from the extracted data 312. Alternatively or complementary, the labels 333 can be determined based on an observed action history of underwriters. An action history of an underwriter refers to the actions an underwriter 330 performs when collecting 351 and/or selecting 352 data associated with the risk of the insurance policy. These actions can be observed by monitoring and logging 353 the actions of underwriters in a computer application 350. The computer application 350 can be a desktop application or web application that provides a user interface wherein an underwriter can search, select, and collect relevant underwriting data. The computer application 350 can further be configured to receive the request for underwriting an insurance policy, i.e. the unstructured electronic data message. This allows to label 354 entities 221, 222, 223 in the knowledge graph 204 based on the logged actions 353, e.g. by adding labels relating to the risk of an entity 221, 222, 223 as perceived by an underwriter 340. As such, querying the knowledge graph 204 can result in a selection of entities that substantially matches the selection made by underwriters 340. In other words, this allows to encode the expertise of underwriters in the knowledge graph 204. It is thus an advantage that the experience and expertise of underwriters can be embedded in the knowledge graph 204.

[62] Additionally, the selection of entities obtained by querying the knowledge graph 204 can be exposed 356 to an underwriter 340 via computer application 350.

This can allow the underwriter 340 to enhance and/or redact the selection of entities.

The actions performed by the underwriter 340 can be monitored and logged 353.

Such actions can for example include omitting entities from the selection, adding entities to the selection, or re-querying the knowledge graph 204 for additional information. The labels 333 of the entities 221, 222, 223 in the knowledge graph 204 can then be adjusted 356 according to the logged actions 353 of the underwriter, i.e. the underwriter feedback. It is thus an advantage that the knowledge graph 204 can be continuously improved by embedding the underwriter feedback.

[63] The knowledge graph 204 can further be enriched over time by adding additional entities 221, 222, 223 and relationships 231, 232. Entities and relationships can also be extended and adjusted as additional entities are considered for addition to the knowledge graph 204. Further periodic curating of the knowledge graph 204 may include link prediction to predict missing edges 231, 232, entity resolution to merge redundant nodes 221, 222, 223, and link-based clustering to group entities 221, 222, 223 based on a similarity of their edges 231, 232.

[64] Fig. 4 shows a supervised training process 400 of a classifier 330 to extract structured underwriting data 440 from an unstructured electronic data message 410 comprising a request 411 for underwriting an insurance policy. The training process 400 can start by gathering a plurality of unstructured electronic data messages 410 for underwriting an insurance policy, i.e. training data 401. In a next step, the unstructured electronic data messages 410 can be annotated 422, 423, 424 with information on the underwriting data 412, 413, 414 that is to be extracted by the classifier 430. This underwriting data 412, 413, 414 can for example be, amongst others, an identification of a policyholder 412, an insured object 413, and a desired coverage 414. Annotating 402 an unstructured electronic data message 410 may include locating and labelling desired keywords, tokens, values, or subsections in the unstructured electronic data message 410, e.g. by an expert underwriter. In doing so, a dataset of annotated electronic data messages is constructed that comprises, amongst others, keywords, tokens, values, and positions of the underwriting data 412, 413, 414 in the message.

[65] Alternatively or complementary, the classifier 430 can be pre-trained on a benchmark dataset, e.g. the Enron dataset, the Google Blogger Corpus dataset, or the 20 Newsgroups dataset. The classifier 430 can further be trained by unsupervised learning or by reinforcement learning. Training the classifier 430 by machine learning can thus result in a machine learning model 403 that can locate and extract structured underwriting data 440 from an unstructured electronic data message 410. The trained classifier 430 can for example be, amongst others, a neural network, a support vector machine, a Bayesian network, a conditional random field, a maximum entropy model, or any other classifier known to the person skilled in the art.

[66] The received unstructured electronic data message 410 comprising a request for underwriting the insurance policy can include errors, e.g. in the spelling of a company name or in the VAT identification number of the policyholder. An error can also be induced in the underwriting data 440 by the extracting thereof from the unstructured message 410. To this end, a correction may advantageously be performed wherein the extracted underwriting data 440 is checked for correctness.

Extracting the structured underwriting data 440 can thus further comprise correcting errors in the extracted underwriting data 440. This correcting can be performed substantially after the extracting of structured underwriting data 330, or during the extracting of structured underwriting data 330. For example, correcting of the underwriting data 440 can be performed after the trained classifier 430 extracts the data 440.

[67] An error in structured underwriting data can be corrected by matching the erroneous underwriting data to correct underwriting data describing the same characteristic feature of an entity, e.g. the name of a policyholder. For example, an error in underwriting data that describes the name of a policyholder can be corrected by matching it to other extracted underwriting data that correctly describes the name of that policyholder. Alternatively, an error in structured underwriting data can be corrected by comparing the meaning of the erroneous underwriting data to the meaning of correct underwriting data describing the same entity, e.g. by comparing the name and the national registration number of a policyholder. For example, an error in extracted underwriting data that describes the name of a policyholder can also be corrected by comparing it to correct underwriting data that describes the national registration number of said policyholder as both keywords describe the same entity, i.e. the policyholder. Alternatively or complementary, keywords can be checked for correctness by searching databases, e.g. a public register. This can allow to collect additional data to check the correctness of underwriting data.

Approximate string matching can further be used when erroneous underwriting data cannot be used to search a database for additional data.

[68] In a substantially similar manner, a natural language processing algorithm can also be implemented to extract structured cost data from a history of claims, ie. previous unstructured claims. A claim can be an unstructured text document comprising a request for compensation of a loss covered by an insurance policy. The previous unstructured claims can be stored in a database or a knowledge base. By the natural language processing algorithm, the previous unstructured claims can be interpreted to extract structured cost data. Structured cost data can include elements of a claim such as, for example, one or more keywords, subsequences of a sentence, or one or more sentences included in the unstructured text of the claim.

The structured cost data can further include the cost associated to the respective claim elements. In other words, the structured cost data allows to classify claim elements in previously filed claims according to their respective cost.

[69] In a substantially similar manner, the terms and conditions of previous insurance policies can also be analysed by means of natural language processing.

The terms and conditions are a part of an insurance policy that describe how the policy will operate and specify what is covered by the insurance policy. The terms and conditions can be an unstructured text document. By processing the terms and conditions, structured elements can be obtained such as, for example, one or more keywords, subsequences of a sentence, or one or more sentences included in the unstructured text of the terms and conditions. These structured elements of the terms and conditions can then be compared with and associated to the structured cost data of the respective claim. This can allow to assign a cost to elements in the terms and conditions. In other words, the structured elements of the terms and conditions can be classified according to their respective cost in the claims. This classification allows to adjust the terms and conditions to optimize the loss ratio of an insurance policy, ie. reduce the loss ratio. It is an advantage that the exposure of an insurer can be controlled by excluding high-cost elements in the terms and conditions of an insurance policy.

[70] The computer implemented method can further comprise optimizing the portfolio of insurance policies. Portfolio optimization relates to the process of maximizing long-term performance metrics of the portfolio by carefully selecting the individual insurance policies that are underwritten. To this end, a future state of the portfolio of insurance policies can be predicted. A portfolio state can be characterized by one or more performance metrics, such as, for example, expense ratio, retention rate, average claim cost, claim frequency, global loss ratio of the portfolio, a gross sum of insured damages, market share, diversification, and profit margins.

[71] Predicting the state of the portfolio can be based on the loss ratio for underwriting the insurance policy, previous states of the portfolio, and previous underwritten insurance policies of the portfolio. A previous period of insurance can be divided into a discrete sequence of states and actions. An action can for example be underwriting a new policy, renewing an existing policy, or adjusting an existing policy.

Each action can further be characterized by the associated individual policy, e.g. characterized by the premium, terms and conditions, and coverage of the individual policy. A state-action model can then be used to model a reward for every previous action based on the portfolio state resulting from that previous action. Such a state- action model can then allow to predict a future state of the portfolio when underwriting an insurance policy based on the determined loss ratio for that insurance policy. The state-action model can be based on a reinforcement learning method such as, for example, the state-action-reward-state-action method, SARSA, the Q-learning method, or a temporal difference learning method.

[72] In other words, the effect of underwriting an individual insurance policy on the global insurance portfolio of an insurer can be predicted. This can allow to optimize the portfolio by adjusting the insurance policy according to the predicted state of the portfolio, i.e. by adjusting the policy proposal. It is a further advantage that the global portfolio of insurance policies is considered in the underwriting process of an individual insurance policy.

[73] Fig. 5 shows a suitable computing system 500 enabling to implement embodiments of the above described method according to the invention. Computing system 500 may in general be formed as a suitable general-purpose computer and comprise a bus 510, a processor 502, a local memory 504, one or more optional input interfaces 514, one or more optional output interfaces 516, a communication interface 512, a storage element interface 506, and one or more storage elements 508. Bus 510 may comprise one or more conductors that permit communication among the components of the computing system 500. Processor 502 may include any type of conventional processor or microprocessor that interprets and executes programming instructions. Local memory 504 may include a random-access memory (RAM) or another type of dynamic storage device that stores information and instructions for execution by processor 502 and/or a read only memory (ROM) or another type of static storage device that stores static information and instructions for use by processor 502. Input interface 514 may comprise one or more conventional mechanisms that permit an operator or user to input information to the computing device 500, such as a keyboard 520, a mouse 530, a pen, voice recognition and/or biometric mechanisms, a camera, etc. Output interface 516 may comprise one or more conventional mechanisms that output information to the operator or user, such as a display 540, etc. Communication interface 512 may comprise any transceiver- like mechanism such as for example one or more Ethernet interfaces that enables computing system 500 to communicate with other devices and/or systems such as for example, amongst others, a remote server 550. The communication interface 512 of computing system 500 may be connected to such another computing system by means of a local area network (LAN) or a wide area network (WAN) such as for example the internet. Storage element interface 506 may comprise a storage interface such as for example a Serial Advanced Technology Attachment (SATA) interface or a Small Computer System Interface (SCSI) for connecting bus 510 to one or more storage elements 508, such as one or more local disks, for example

SATA disk drives, and control the reading and writing of data to and/or from these storage elements 508. Although the storage element(s) 508 above is/are described as a local disk, in general any other suitable computer-readable media such as a removable magnetic disk, optical storage media such as a CD or DVD, -ROM disk, solid state drives, flash memory cards, etc. could be used.

[74] Although the present invention has been illustrated by reference to specific embodiments, it will be apparent to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied with various changes and modifications without departing from the scope thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. In other words, it is contemplated to cover any and all modifications, variations or equivalents that fall within the scope of the basic underlying principles and whose essential attributes are claimed in this patent application. It will furthermore be understood by the reader of this patent application that the words "comprising" or "comprise" do not exclude other elements or steps, that the words "a" or "an" do not exclude a plurality, and that a single element, such as a computer system, a processor, or another integrated unit may fulfil the functions of several means recited in the claims. Any reference signs in the claims shall not be construed as limiting the respective claims concerned. The terms "first", "second", third", "a", "b", "¢", and the like, when used in the description or in the claims are introduced to distinguish between similar elements or steps and are not necessarily describing a sequential or chronological order. Similarly, the terms "top", "bottom", "over", "under", and the like are introduced for descriptive purposes and not necessarily to denote relative positions. It is to be understood that the terms so used are interchangeable under appropriate circumstances and embodiments of the invention are capable of operating according to the present invention in other sequences, or in orientations different from the one(s) described or illustrated above.

Claims

00. CONCLUSIONS

A computer-implemented method (200) for granting an insurance policy, comprising: - receiving (201) an unstructured electronic data message (211) containing a request to grant the insurance policy; - extracting (202) structured underwriting data (212) from the unstructured electronic data message by means of natural language processing including at least an identification of a policyholder, an insured item and a desired coverage; generating (203) at least one query (213) based on the structured underwriting data for querying a knowledge graph (204), the knowledge graph being a network of entities (221, 222, 223) containing one or more labels and their respective relationships (231, 232); obtaining (205) a selection of entities (215) from the knowledge graph by querying the knowledge graph with the at least one query, where an entity of the selection has a relationship with the structured underwriting data and one or more predetermined labels includes; — identifying (207) one or more peers (216) of the insurance policy from a portfolio of insurance policies (206) based on the selection of entities, where a peer is characterized by a set of entities that are substantially similar to the selection of entities; and - determining (208) a loss ratio (218) for granting the insurance policy based on the loss ratios of the peer(s).

A computer-implemented method according to claim 1, wherein extracting the structured underwriting data is performed by a classifier (430) trained to extract structured underwriting data (440) from an unstructured electronic data message (410).

A computer-implemented method according to any one of the preceding claims, wherein extracting (202) the structured insurance technical data further comprises correcting errors in the extracted structured insurance technical data (440) by comparing erroneous structured insurance technical data with other structured insurance technical data.

A computer-implemented method according to any one of the preceding claims, wherein one or more labels (333) of the respective entities (221, 222, 223) in the knowledge graph (204) are determined based on an observed merchant action history .

A computer-implemented method according to any one of the preceding claims, further comprising obtaining (355) merchant feedback on one or more entities in the selection of entities, and modifying (356) the labels of the respective entities (221, 222, 223) in the knowledge graph (204) according to the feedback.

A computer-implemented method according to any one of the preceding claims, wherein determining (208) the loss ratio for granting the insurance policy is based on a probability distribution of the loss ratios of the one or more peers (216).

A computer-implemented method according to any one of the preceding claims, further comprising determining a premium for the insurance policy based on the determined loss ratio (218) for granting the insurance policy.

A computer-implemented method according to any one of the preceding claims, further comprising predicting a condition of the portfolio of insurance policies (206) at the issuance of the insurance policy based on the loss ratio for the issuance of the insurance policy, previous conditions of the portfolio, and previously issued insurance policies of the portfolio.

A computer-implemented method according to any one of the preceding claims, further comprising extracting structured cost data from previous unstructured claims by natural language processing, and associating the structured cost data with the terms of the insurance policy for which an unstructured claim has been made.

A computer-implemented method according to claim 9, further comprising determining the terms of the insurance policy based on the structured cost data and the determined loss ratio for granting the insurance policy.

A computer-implemented method according to any one of the preceding claims, further comprising parsing structured data (312) and/or scraping unstructured data (311) associated with the insurance policy from at least one data source (310) to to encode (330) into the knowledge graph (204).

A computer-implemented method according to claim 10, wherein the unstructured data is structured (320) by natural language processing and/or computer vision.

A data processing system arranged to perform the computer-implemented method of any one of claims 1 to 12.

A computer program comprising instructions which, when executed by a computer, cause the computer to perform the computer-implemented method of any one of claims 1 to 12

A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to perform the computer-implemented method of any one of claims 1 to 12.