Microsoft recently introduced Communication Compliance to M365 with Supervision Policies offering control over lexicon, including OOT Classifiers such as Profanity, Targeted Harassment, Threat, Adult Images, Gory Images, and Racy Images.
However, with the quick deprecation of OOTB Offensive Language Classifiers and the added overhead of Communication Compliance moderation and administration, creating a custom solution became essential.
This custom Communication Compliance solution was specifically tailored to address profanity and offensive language challenges.
Building Profanity Dictionaries
Building a comprehensive dictionary for detecting offensive language is no small feat. Cultural nuances play a significant role, making it challenging to account for every context and interpretation.
For example, in my culture of origin, being called the name of a specific innocuous animal is deeply offensive. Translated into English, the word may lose its impact, but it can still be perceived as offensive. Similar nuances exist across different cultures, with seemingly innocuous words carrying significant weight in certain contexts.
Attempting to build a dictionary that covers every cultural nuance while offending nobody is an impossible task. Striving for such perfection would only result in an abundance of false positives, rendering the system ineffective.
Instead of aiming for an unattainable level of perfection, my focus was on identifying well-known profanity and offensive words that are common to both English and French languages and cultures. By concentrating on these widely recognized terms, I aimed to create a practical solution that balances cultural sensitivity with effective content moderation.
In my role as an architect of bilingual systems, designing solutions for clients with bilingual requirements, it was imperative to address profanity detection in both English and French languages. To meet this need, I sought out banned word lists online for both languages.
Creating Sensitive Information Type (SIT) and DLP Policies
To make sure my files fit within the size limits for DLP and transport rules in M365, I used PowerShell to split them. Each file had a list of offensive words, but they were too big. So, I ended up with 15 smaller files instead. (I am not posting resulting files here, so that I do not offend anybody.)
After splitting the files, I proceeded to create custom Sensitive Information Types (SITs) by selecting the "Keyword Dictionary" type. These SITs were then utilized in crafting Data Loss Prevention (DLP) policies for SharePoint Online and MS Teams Workloads.
The Outcome
As a result, Microsoft Chat posts containing profanity or offensive language were automatically blocked. Instead of the post, users received a notification explaining why it was blocked.
Similarly, attempts to upload documents containing offensive language to SharePoint Online or Microsoft Teams were met with automatic blocking and an explanatory message.
Furthermore, if a user attempted to modify a document to include offensive language, the update would not be accepted, and the user would receive a notification.
Administrators were promptly notified of any incidents, receiving detailed information including the user's name, the location of the incident, and the exact word match found in the dictionary.
This comprehensive approach ensured effective content moderation and enforcement of compliance standards across the M365 environment.
Mitigating Religious Context Challenges
When proposing this solution to Correctional Services, they quickly identified a significant issue that needed to be addressed. Canadian French profanity includes sacred religious words that are commonly used in Church settings but become offensive when used outside of their religious context.
Given that Correctional Services had a Chapel with documents stored in a designated SharePoint site, the initial implementation of the solution inadvertently rendered the Chapel SharePoint site and its associated Microsoft Teams channels dedicated to religious services effectively unusable.
This presented a unique challenge where the list of offensive words for the specific location needed to be compiled with the exclusion of religious terms for the French dictionaries. To address this, the Chapel SharePoint site had to be excluded from the DLP Policy Locations for the original dictionary.
Subsequently, a new DLP Policy was created utilizing an amended dictionary that excluded religious terms. This new policy was then applied solely to the Chapel Site, ensuring that the content moderation measures were appropriately tailored to the unique needs of that specific location.
Comments