Very good article, I enjoyed it and learnt something new. But what I will say coming from an accounting background is that understanding the relationships between the various line items is the bare minimum knowledge requirement for financial analysis, eg an analyst’s job is not to classify inventory as a current asset, this is assumed knowledge. Much of the complexity from financial analysis and normalisation of statements comes from non-standardised reporting and subjective interpretation of the accounting standards rather than difficulties with understanding the relationship between items.

For example to understand whether a certain lease is to be expensed or capitalised an analyst will have to read the accompanying notes. And quite often items will be unique to that reporting period and/or company. In these instances, the analyst will still have to manually specify the entity to entity relationship on a case by case basis thus I don’t think the knowledge graph explicitly addresses normalisation.

In addition, accounting is rife with semantics, despite similarities of the word embeddings in vector space, acquisition and gains are very different concepts. Getting NLP to work with accounting terms will most likely require training on a very specific corpus of domain vocabulary or a much more nuance word embeddings model.

Regardless, good stuff, I will look into this further after your post.

Researcher | Investor | Data Scientist | Curious Observer. Thoughts and insights from the confluence of investing and machine learning.