My last blog triggered a flurry of responses, especially on the question of whether ChatGPT would be a game-changer for IDP. As I reported, the IDP experts at Super.ai looked at ChatGPT in its current iteration and concluded it was too early to tell. They wrote, “these models are not directly applicable to document processing and the unstructured data processing market. We haven’t seen any research on closing this (invoice processing) gap with LLMs.” My cheeky summary was: are we there yet? No, we are still parked in our driveway waiting for a roadmap. 

If this was a proper enquiry in a court of law, our opinions and comments would be dismissed as speculative or hearsay evidence. Because I watched every episode of Law and Order, I felt qualified to put the robot itself on the stand to testify in its own words if or how it can help IDP. Followed by an expert rebuttal witness. Here is that testimony, verbatim. (Please note that ChatGPT completely ignored the advice of its attorney to give the briefest possible answers and insisted on over-explaining.) 

Mr. Lucarini: Can ChatGPT be used for Intelligent Document Processing?

ChatGPT: Intelligent Document Processing (IDP) refers to the use of AI and machine learning technologies to automatically extract information from unstructured or semi-structured documents, such as invoices, receipts, contracts, and emails. ChatGPT can be used for IDP in a number of ways:

1. Document Classification: ChatGPT can be used to classify documents based on their content or structure. For example, it can be trained to distinguish between different types of invoices, or to recognize whether a document is a contract or a memo.

2. Named Entity Recognition (NER): ChatGPT can be used to identify and extract named entities such as people, organizations, and locations from documents. This can be useful for tasks such as extracting contact information from business cards or identifying key stakeholders in a contract.

3. Text Extraction: ChatGPT can be used to automatically extract text from images or scanned documents using Optical Character Recognition (OCR) technology. This can be useful for tasks such as digitizing paper documents or extracting information from forms.

4. Text Summarization: ChatGPT can be used to summarize the content of long documents or emails, making it easier for humans to quickly review and understand the key points.

5. Sentiment Analysis: ChatGPT can be used to analyze the sentiment of documents, such as customer reviews or social media posts, to identify positive or negative feedback about a product or service.

Overall, ChatGPT’s ability to process and analyze natural language can be leveraged in a wide range of IDP tasks, helping to automate document processing and improve efficiency in a variety of industries such as finance, healthcare, and legal.

Mr. Lucarini: What else can ChatGPT do for IDP?

ChatGPT: I can be used to create synthetic data to train IDP AI models. Synthetic data is artificially generated data that is designed to mimic real-world data, and can be used to supplement or replace actual training data.

To generate synthetic data, ChatGPT can be trained on a set of existing documents, and then used to generate new, synthetic documents that are similar in structure and content to the original documents. These synthetic documents can be used to train IDP AI models in place of, or in addition to, real-world data.

One advantage of using synthetic data is that it can be generated quickly and at scale, which can be particularly useful in situations where real-world data is scarce or difficult to obtain. Additionally, synthetic data can be used to test and refine AI models in a controlled environment, without the risks associated with using sensitive or confidential data.

However, it’s important to note that the quality of the synthetic data generated by ChatGPT, or any other AI model, is dependent on the quality of the original training data. If the original training data is biased or incomplete, the synthetic data generated by the model may also exhibit similar biases or limitations. Therefore, it’s important to carefully evaluate the quality and accuracy of any synthetic data generated by ChatGPT before using it to train IDP AI models.

Mr. Lucarini: Can I start training ChatGPT today?

ChatGPT: As an end-user, you cannot directly train ChatGPT. If you’re interested in training your own language model, OpenAI offers a cloud-based platform called GPT-3 API, which allows developers to create and fine-tune their own language models using a large dataset of text. However, access to the GPT-3 API is currently limited and requires an application and approval process.

Alternatively, there are other machine learning platforms and frameworks, such as TensorFlow or PyTorch, that you can use to train your own language models from scratch, although this requires a significant amount of technical expertise and computational resources.

Mr. Lucarini: So your testimony is that, with training you can help IDP; but to train you, I need to use the GPT-3 API?

ChatGPT: Yes, that’s correct. To train a language model like ChatGPT, you would need access to a large dataset of text and a machine learning framework like TensorFlow or PyTorch. The GPT-3 API is one such framework that allows developers to create and fine-tune their own language models using a large dataset of text.

Once you have trained a language model, you can then use it to perform a variety of natural language processing tasks, including Intelligent Document Processing (IDP), as described earlier. However, it’s important to note that training a high-quality language model requires significant technical expertise and computational resources, and may not be practical for all users.

If you don’t have the technical expertise or resources to train your own language model, you can still leverage the capabilities of existing language models like ChatGPT through APIs or pre-trained models provided by companies like OpenAI. These APIs and pre-trained models can be used to perform a variety of natural language processing tasks, including IDP, without requiring users to train their own models from scratch.

Mr. Lucarini: I have no further questions for the witness at this time. The prosecution now calls industry expert Alan Pelz-Sharpe to the stand as a rebuttal witness. Mr. Pelz-Sharpe, in your expert opinion, does IDP really need ChatGPT?

Mr. Pelz-Sharpe: ChatGPT is  just another Large Language Model (LLM); or more precisely, it’s an MLM (M for Massive). I am all in for using LLMs for document-centric work, be it enterprise search or document analysis or IDP. LLMs are already being used for search and certainly could be used for IDP, at the least to generate training data for the AI. But I’m at a loss why one would use the ChatGPT version. My point is, why use a generic MLM like ChatGPT, when a custom LLM built specifically for, let’s say, the commercial boat insurance industry would be smaller, cheaper, faster, and more accurate?  Choosing ChatGPT is like using an articulated truck to drive from your house in Cornwall to the local shop via the backroads. Yes it will get you there, but at what cost and how long will it take? On your single-lane roads, a Ford Fiesta would be the better choice. 

Mr. Lucarini: Thank you, Mr. Pelz-Sharpe. No further questions at this time.