As an operations leader, optimizing knowledge processes is an important part of your strategy. Reducing effort, Improving accuracy and Optimizing costs are recurrent goals for the Operations team of any enterprise. Named Entity Recognition is a vital part of the knowledge process team for any enterprise, be it a commercial real estate services firm, a law firm or an enterprise software sales firm. All medium to large enterprises consume and process large amounts of legal documentation and spend huge costs on processing these documents. This blog aims to recommend the latest approaches to implementing LLM-based solutions for NER tasks to improve knowledge process strategies for enterprises.

The Current State of Knowledge Processes

The most common practice is to outsource knowledge processes to third-party firms that employ SMEs or automation tools to implement the aforementioned processes. This practice worked well due to the absence of models and tools that can work with unstructured and rapidly changing documentation. Known limitations of this practice include vendor lock-in, administrative overhead and black box tooling.

The Future of Knowledge Processes with LLMs

LLMs have transformed the NER frameworks from bulky rule-based regular expression search engines to token classification models. The advantages of this advancement include significantly less preprocessing of input text, more structured output for consumption, generalization like multi-lingual and multi-domain implementations and MLOPs readiness to name a few. Transfer learning is particularly useful to organizations to avoid training models from scratch and realize value faster. Tools like ChatGPT allow organizations to circumvent the need for training models completely and directly jump into inference for operationalization.

Let us look at an example of NER on Court case documents to illustrate the advantages LLMs bring to the process. We will use the legal court cases dataset from the OpenNyAI team. This team took a base roBERTa model and trained it on the Indian court cases corpus and thereby using transfer learning to leverage the pre-trained LLM to quickly develop a domain-specialized model.

Here’s the output of the custom-trained LLM:

To train a base LLM model for a specialized task, developers need training data. This might not be feasible for all organisations. In such a scenario, Generative AI models provide an easier way to implement NER. Below is what OpenAI’s ChatGPT-3.5 can extract using a little prompt engineering.

Based on the amount of prompt customization, ChatGPT can match the output of a specialized LLM in this particular case. Other major advantages of using ChatGPT are the overall reduction in TCO, technical debt and time to production.

Conclusion

This blog post has shown how Operations teams can improve their knowledge processes using the advancements in Natural Language Understanding like LLMs and LLM-based Generative AI models. Depending on the requirements of the organisation, they can choose to build custom models or leverage out-of-the-box intelligence to achieve tasks related to knowledge processes. If you are interested in learning about how LLMs and Vector Databases can optimise your entity resolution (ER) workflows, then please read our previous blog here.