Identifying the Client Associated with a Legal Document

The main objective was to identify the client(s) associated with each document through one of the following identifiers:

The goal was to extract client names from legal documents using Named Entity Recognition (NER). Here's how I approached the task:

Data: I had a collection of legal documents in PDF format. The task was to identify the clients mentioned in each document using one of the following identifiers:

Approximate client name (e.g., "John Doe")

Precise client name (e.e., "Doe, John A.")

Approximate firm name (e.g., "Doe Law Firm")

Precise firm name (e.g., "Doe, John A. Law Firm")

About 5% of the documents didn't include any identifying entities.

Dataset: For developing the model, I used 710 "true" PDF documents, which were split into three sets: 600 for training, 55 for validation, and 55 for testing.

Labels: I was given an Excel file with entities extracted as plain text, which needed to be manually labeled in the document text. Using the BIO tagging format, I performed the following steps:

Mark the beginning of an entity with "B-".

Continue marking subsequent tokens within the same entity with "I-".

If a token doesn't belong to any entity, mark it as "O".

Alternative Approach: Models like LayoutLM, which also consider bounding boxes for input tokens, could potentially enhance the performance of the NER task. However, I opted not to use this approach because, as is often the case, I had already spent the majority of the project time on preparing the data (e.g., reformatting Excel files, correcting data errors, labeling). To integrate bounding box-based models, I would have needed to allocate even more time.

While regex and heuristics could theoretically be applied to identify these simple entities, I anticipated that this approach would be impractical, as it would necessitate overly complex rules to precisely identify the correct entities among other potential candidates (e.g., lawyer name, case number, other participants in the proceedings). In contrast, the model is capable of learning to distinguish the relevant entities, rendering the use of heuristics superfluous.