A Note on Entity-Linking

PrivateGPT has the capability to track different mentions of the same entities over multiple messages. Importantly, however, this is only possible if exactly the same names are used in each reference to the entity. If you refer to an entity with multiple different names, PrivateGPT may not be able to correctly link the entity in its response. This includes variations like different spellings, different cases, and typos.

In the following example, the user has asked PrivateGPT to draft a letter. Note that in the redacted text sent to ChatGPT, both mentions of Sun Care Health are redacted as [ORGANIZATION_1], demonstrating that PrivateGPT is linking both mentions to a single entity.

In the following prompt, the user asks PrivateGPT to add a contact number for Sun Care Health to the letter. Because they have used the exact organization name that was used in the first prompt, PrivateGPT understands that the user is again referring to the same entity, and once again it is redacted as [ORGANIZATION_1].

Because the entity is properly linked to the previous mentions, the phone number is added in the subsequent response, as requested by the user.

In the third prompt, however, the user refers to the organization with an acronym, SCH, rather than its full name. PrivateGPT detects this as a new entity, and it is redacted as [NAME_2]. ChatGPT then incorrectly associates the acronym (and following address) with the addressee of the letter. The resulting response contains inaccurate contact information for both parties.

Similar confusions can result from referring to an individual by different names (Susan Cheng / Susan / Ms. Cheng), with case mismatches (Vikram / vikram), or typos (Ben Jameson / Ben Jamesson), etc.

To avoid these outcomes, take care to always refer to each entity consistently.