The Bloomberg Law newsroom is working on a new AI-powered tool to help edit articles for style and cut down on the time editors spend copy editing articles. The tool, called the Stylebook Assistant, will use a large language model to copy edit articles so they align with our newsroom’s specific style standards.
There are text editors for professional publishers that already exist, like Tansa or Grammarly, that use AI or other technology to edit articles for grammar, sentence structure and tone. But the Stylebook Assistant provides a different take as an editing tool, moving its capabilities from centering around grammar editing to focusing on style editing.
Most newsrooms follow standards set in the AP Stylebook, applying grammar, punctuation and usage rules to their sentences to maintain a standard format for all articles. Some newsrooms have their own entries that dictate style rules around particulars in the industries or localities they cover. But only a special handful of newsrooms, like The New York Times and Bloomberg Law, have their very own stylebook. And honestly, it’s very difficult to remember the nuances of rules when reporters and editors write articles for different newsrooms throughout their careers.
Enter the Stylebook Assistant
The Stylebook Assistant is still in development, but our approach so far is straightforward with a heap of engineering help. Newsroom editors will send an unpublished article to a large language model that we host in our AWS Bedrock account, ensuring that our data is safe and not being sent into the large language model for training. The large language model sifts through the article, sentence by sentence, to find if a style rule is applicable to an individual sentence. For example, the sentence below would be flagged for misusing our rule on “Big Four” firms:
The “Big Four firms” include Deloitte, Ernst and Young, KPMG, and PwC.
Bloomberg Law style states that a journalist in our newsroom should use “the Big Four firms,” without the quotation marks, when referring to the accounting firms Deloitte LLP, Ernst and Young LLP, KPMG LLP, and PwC. The large language model will identify the sentence as breaking the style rule for Big Four firms. Then, the LLM will refer to the Big Four rule written as a prompt in order to suggest an edit so that the sentence instead says:
The Big Four firms include Deloitte, Ernst and Young, KPMG, and PwC.
The Bloomberg Law newsroom has over 250 rules, meaning we have to create and test 250 individual prompts that will run over each unpublished article. We have a long road ahead of us in this project, but we are excited to watch its progress and potential.
Using AI to Copy Edit in Your Newsroom
If I haven’t lost you yet, I have a few suggestions on how smaller newsrooms with less engineering power can approach this application of large language models on a task that can be deduced to rote simply because of the rules we follow as journalists.
A newsroom that uses the AP Stylebook in its purest form to copy edit articles can try creating a prompt in ChatGPT that asks it to edit an article for AP Style. A prompt that completes this task might look something like this:
“Edit the article in <document> tags below using Associated Press style, ensuring consistency in formatting, punctuation and language usage. Make all necessary edits to the article while preserving its original meaning and intent. After editing, provide a brief summary of the major changes you made, highlighting any significant style corrections or adjustments related to the specific rules mentioned. <document> [insert article here] </document>”
A newsroom that uses a combination of the AP Stylebook rules along with more nuanced newsroom-specific rules can try creating a prompt in ChatGPT that asks it to edit an article for AP Style, but also includes a list of those specific rules in instructional format. A prompt that completes this task might look something like this:
“Edit the article in <document> tags below using Associated Press style. Pay special attention to the following specific rules: 1) Do not use the word “between” for a range that includes the numbers on each end. Instead, use “from [start year] to [end year]” format. Make all necessary edits to the article while preserving its original meaning and intent. After editing, provide a brief summary of the major changes you made, highlighting any significant style corrections or adjustments related to the specific rules mentioned. <document> [insert article here] </document>”
These prompts are merely suggestions and haven’t been tested at scale. But each prompt should give you an idea on how to approach creating an effective prompt that could help alleviate the amount of time spent copy editing articles.
As you work to find a prompt, remember to:
- Test your prompt on a variety of articles to ensure the style rules can be applied to variety of writing styles.
- Verify all edits to ensure they do not change the meaning or substance of the article.
- Try out different models.
- Be wary of the ever-present possibility for inaccuracies introduced by a large language model.
Happy prompting!