Cutting-Edge AI for SEC EDGAR: Introducing MemWalker

[Elijas]

Hello everyone! I'm thrilled to share a new approach to long text reading, inspired by a series of tweets from Howard Chen (@__howardchen) and a research paper titled "Walking Down the Memory Maze: Beyond Context Limit through Interactive Reading". This approach, called MemWalker, introduces a fundamentally different method to handle long text reading, which could be particularly beneficial for processing large documents like SEC EDGAR filings.

Why MemWalker?

Through interactive reading on this model-constructed memory tree, MemWalker outperforms other long context baselines as well as retrieval and recurrent variants, especially with the longer examples.

MemWalker: A Two-Stage Process

MemWalker is a two-stage process:

1. Constructing the Memory Tree: The long text is first segmented and summarized into summary nodes. These summary nodes are further summarized into higher-level nodes until it reaches the root.

2. Navigation: Upon seeing a query, the Large Language Model (LLM) navigates the tree to look for relevant information and crafts an appropriate response. The LLM achieves this by reasoning – whether to commit to an answer, choose to go further down the path, or retract if it finds itself astray.

The navigation can be achieved by zero-shot prompting and is readily applicable to any LLM of your choice.

Two critical components enable the effectiveness of MemWalker:

1. Working Memory: When the LLM is allowed to carry information along the path it traverses, the LLM has a better global context.

2. The Reasoning Ability of the LLM: MemWalker is effective when the LLM passes a reasoning threshold. Error cascades too fast during navigation when the reasoning capability is below the threshold. Therefore, using a better-performing (as of October 2023) model like GPT-4 would be preferred.

MemWalker and sec-parser

Our sec-parser project simplifies extracting meaningful information from SEC EDGAR HTML documents by organizing them into semantic elements and a tree structure. This semantic tree can easily be converted into a memory tree, which puts you in a perfect position to apply the MemWalker technique in AI applications or AI agents. This approach could potentially enhance the efficiency and effectiveness of information extraction and navigation in large documents like SEC EDGAR filings.

For more details about MemWalker, you can check out the research paper and the Twitter thread by Howard Chen.

We're excited about the potential of this approach and look forward to seeing its applications with sec-parser. As always, we welcome your thoughts and feedback in the discussions!