In the realm of data privacy, the complexity of regulations such as the European Union’s General Data Protection Regulation (GDPR) can pose a significant challenge for organizations. Compliance with GDPR requirements, which include obtaining user consent, providing access to personal data, ensuring its portability, and facilitating data erasure upon request, is difficult to enforce at a technical level.
Additionally, making informative yet anonymous data available for engineers and data teams to develop products and extract insights is a daunting task. To achieve this, complex workflows must be developed to handle various specific use cases, such as differences in language, variations in personally identifiable information (PII) structure, and the use of text dictionaries for identifying names and locations. These intricate workflows require a considerable amount of effort and can result in a maintenance-heavy codebase if not efficiently managed.
Consequently, organizations often find themselves in a labyrinthine web of regex coding, with a multitude of cascaded and nested IF-ELSE statements. The outcome is an engineering nightmare that requires significant effort to maintain efficiency and ensure up-to-date compliance. Thus, a more efficient solution is warranted.
The Solution:
Large language models (LLMs) !
Consider leveraging the expertise of a LLM to handle complex text-based tasks. These machin learning models have been trained on extensive data and excel at executing clearly defined, intricate assignments.
To illustrate the potential benefits of using a LLM, let’s consider an example involving email text anonymization, though this could be applied to any type of text input. I input the following prompt into the model:
Read the following email text, that can be in any language. Remove any personally identifiable information (PII),
and replace it with the appropriate placeholder. For example, replace the name "John Doe"
with "[NAME], »:
Hi Christine,
I'm writing to you because I noticed that you recently bought a new house. I'm a salesperson at a local construction business (Mavericks Constructions Ltd.), and I wanted to let you know that we have great deals on home renovations.
If you're interested, please let me know.
Thanks,
Jean Delasalle
Phone: +33712345678
Email: j.delasalle@mavericks.com
Adresse: 3, Rushmore Street, 44300 Nantesand got this as an output:
Hi [NAME],
I'm writing to you because I noticed that you recently bought a new house. I'm a salesperson at a local construction business, and I wanted to let you know that we have great deals on home renovations.
If you're interested, please let me know.
Thanks,
[SENDER]
Phone: [PHONE]
Email: [EMAIL]
Address: [ADDRESS]Amazing ! right ? Here’s the same thing with French text:
Input:
Read the following email text, that can be in any language. Remove any personally identifiable information (PII),
and replace it with the appropriate placeholder. For example, replace the name "John Doe"
with "[NAME], »:
Bonjour Madame Nguyen,
Je vous écris parce que j'ai remarqué que vous aviez récemment fait l’acquisition d’une nouvelle maison.
Je représente une entreprise de construction locale (Mavericks Constructions SAS) et je voulais vous faire savoir que nous avons d'excellents prix sur les travaux de rénovation.
Si vous êtes intéressée, merci de me le faire savoir.
Merci,
Jean Delasalle
Téléphone : +33712345678
Courriel : j.delasalle@mavericks.com
Adresse : 3, rue Rushmore, 44300 NantesOutput:
Bonjour [NAME],
Je vous écris parce que j'ai remarqué que vous aviez récemment fait l’acquisition d’une nouvelle maison.
Je représente une entreprise de construction locale et je voulais vous faire savoir que nous avons d'excellents prix sur les travaux de rénovation.
Si vous êtes intéressé(e), merci de me le faire savoir.
Merci,
[SENDER]
Téléphone : [PHONE]
Courriel : [EMAIL]
Adresse : [ADDRESS]As you can see here, utilizing a pre-trained model publicly available as an API can yield excellent results, even for lengthy inputs up to 4000 words.
This approach offers several benefits, such as working across all languages and eliminating the need for complex custom-made workflows, long chains of heavy data processing tasks, and extensive text dictionaries.
Moreover, this method can be applied to all types of text input and eliminates the requirement to define and maintain a comprehensive list of personal information.
To experience the advantages of this approach firsthand, you can try it out for yourself.
OpenAI’s text-davinci-003 model was used in this example. But, it is possible to run the same task on your cloud or even locally on your servers, provided that the appropriate model is selected and properly called.
This example is straightforward, but there are numerous additional, more complex and valuable applications that can be envisioned.