Strategies and methods for clustering insurance documents using natural language processing

Translator|Li Rui

Reviewer|Sun Shujuan

Natural language processing (NLP) in the insurance industry can benefit from hybrid machine learning/symbolic approaches to improve scalability , while leveraging advanced symbolic reasoning.

Insurance Documents and Policies: Complex Use Cases

It is well known that up to 87% of data science projects fail to move from proof of concept to production; natural in insurance Language processing (NLP) projects are no exception. They must overcome some of the difficulties inevitably associated with this space and its intricacies.

The main difficulties come from:

• The complex layout of insurance-related documents.
• Lack of large corpus with relevant annotations.

The complexity of layout is so great that the same language concept can drastically change its meaning and value depending on where it is placed in the document.

Look at a simple example: If you try to build an engine to identify whether "terrorism" coverage exists in a policy, you will have to assign a different value regardless of where it is placed:

(1) Sub-limit part of the declaration page.

(2) "Exclusion" section of the policy.

(3) Add one or more insurance endorsements.

(4) Add specific endorsements to the coverage.

The lack of high-quality, appropriately sized annotated corpora of insurance documents is directly related to the inherent difficulty of annotating such complex documents and the amount of effort required to annotate tens of thousands of policies.

And this is just the tip of the iceberg. In addition to this, the need to normalize the concept of insurance must also be considered.

Language standardization: an invisible but powerful force in insurance language

When dealing with databases, the standardization of concepts is a well-understood process. Because it is key to applying reasoning and increasing the speed of the annotation process, it is also crucial for NLP in the insurance field.

The concept of normalization means grouping elements under the same tag language, which may look very different. While there are many examples, the most important ones come from insurance policies that cover natural disasters.

In this case, different sub-limits will apply to different flood zones. Areas with the highest risk of flooding are often referred to as "high risk flood zones". This concept can be expressed as:

(1) Level 1 Flood Area

(2) Flood Risk Area (SFHA)

(3) Flood Area A

etc

In fact, any insurance coverage can have many terms that can be grouped together. Depending on the specific geographical area and its inherent risks, the most important natural disaster coverage even has two Difference between tiers or tiers (I, II and III).

Multiply this by all possible elements that can be found, and the number of variants can quickly become very large. This causes both machine learning annotators and natural language processing (NLP) engines to get bogged down when trying to retrieve, infer, or even label the correct information.

New Language Clustering: Hybrid Approaches

A better way to solve complex natural language processing (NLP) tasks is based on hybrid (machine learning/symbolic) techniques that use machine learning-based The clustering of micro-languages ​​improves the outcomes and lifecycle of insurance workflows, which are then inherited by the symbolic engine.

While traditional text clustering is used in unsupervised learning methods to infer semantic patterns and group together documents with similar topics, sentences with similar meanings, etc., hybrid methods are significantly different . Microlinguistic clusters are created at a granular level using machine learning algorithms trained on labeled data using predefined normalized values. Once a microlinguistic cluster is inferred, it can be used in further machine learning activities or in hybrid pipelines driven inference logic based on symbolic layers.

This is in line with the traditional golden rule of programming: "break the problem down." The first step in solving a complex use case (like most use cases in the insurance space) is to break it down into smaller, more palatable chunks.

What tasks can mixed language clustering accomplish, and how is its scalability?

Symbolic engines are often labeled as extremely accurate but not scalable because they do not have the flexibility of machine learning when dealing with situations not seen during training.

However, this type of language clustering solves this problem by leveraging machine learning to identify concepts that are then passed to the complex and precise logic of the symbolic engine next in the pipeline.

The possibilities are endless: for example, the symbolic step can change the intrinsic value of machine learning recognition based on the document segment to which the concept belongs.

Here is an example of using the notation process of "segmentation" (splitting text into relevant regions) to see how to use the labels passed by the machine learning module.

Imagine that the model needs to understand whether certain coverages are excluded from a 100-page policy.

The machine learning engine will first cluster together all possible variations of the "Arts" coverage:

• "Fine Arts"
• " "Work of Arts"
• "Artistic Items"
• "Jewelry"
• and so on.

Following this, the symbols portion of the pipeline will check to see if the "Exclusions" section mentions the "Arts" tag to understand if the coverage is excluded from the policy, or if it is covered ( as part of the sub-limit list).

Thanks to this, machine learning annotators don’t have to worry about assigning different labels to all Arts variants based on their position in the policy: they just need to annotate their variants Normalized value for "Arts", which will act as a micro-language cluster.

Another useful example of complex tasks is data aggregation. If the hybrid engine is designed to extract sub-restrictions of a specific coverage, as well as coverage normalization issues, there is an additional layer of complexity to deal with: the ordering of language items used for aggregation.

Consider that the task at hand is to extract not only the sub-limits of a specific coverage, but also its qualifiers (per event, aggregation, etc.). The three items can be arranged in several different orders:

• Fine Arts $100,000 Per Item
• Fine Arts Per Item $100,000
• Per Item $100,000 Fine Arts
• $100,000 Fine Arts
• Fine Arts $100,000

Exploiting all of these permutations while aggregating data can significantly increase the complexity of a machine learning model. A hybrid approach, on the other hand, would let the machine learning model identify the normalized labels and then let symbolic reasoning identify the correct order based on the input data from the machine learning part.

These are just two examples that demonstrate that an unlimited amount of complex symbolic logic and reasoning can be applied on top of scalable machine learning algorithms to identify canonical concepts.

Scalable workflows that are easier to build and maintain

In addition to scalability, symbolic reasoning brings other benefits to the overall project workflow:

• Instead of implementing different machine learning workflows for complex tasks, different tags need to be implemented and maintained. Additionally, retraining a single machine learning model is faster and consumes less resources than retraining multiple models.
• Since complex parts of business logic are handled symbolically, it is much easier for data annotators to add human annotations to machine learning pipelines.
• For the same reasons mentioned above, it is also easier for testers to provide feedback directly to the machine learning standardization process. Additionally, since the machine learning part of the workflow normalizes language elements, users will have a smaller list of tags to label documents with.
• Symbol rules do not need to be updated frequently: what is updated frequently is the machine learning part, which also benefits from user feedback.

Conclusion

• Machine learning in complex projects in the insurance field may suffer because the inference logic is difficult to compress into simple tags; this also makes the annotator's life more difficult.
• Text location and inferences can dramatically change the actual meaning of concepts with the same linguistic form.
• In a pure machine learning workflow, the more complex the logic, the more training documents typically required to achieve production-level accuracy.
• For this reason, machine learning requires thousands (or even tens of thousands) of pre-labeled documents to build effective models.
• Complexity is reduced by taking a hybrid approach: machine learning and user annotations create language clusters/tags, and these are then used as starting points or building blocks for the symbolic engine to achieve its goals.
• User feedback, once validated, can be used to retrain the model without changing the most granular parts (which can be handled by the symbolic part of the workflow).

Original title: ​​Insurance Policies: Document Clustering Through Hybrid NLP​​, author: Stefano Reitano

The above is the detailed content of Strategies and methods for clustering insurance documents using natural language processing. For more information, please follow other related articles on the PHP Chinese website!