Research has found that the concept of “persona” is predominantly deployed in Human-Computer Interaction (HCI) studies, but there is a stream of Natural Language Processing (NLP) studies that also refer to the concept of persona. This post aims at making a conceptual distinction in these two different types of referring to personas.
First, HCI studies tend to focus on user personas or design personas. These are personas created by humans, algorithms, or collaboration of humans and algorithms, to represent current or potential users of software applications, services, or other products or outputs an organization creates. The users of user or design personas are designers, software developers, marketers, or other stakeholders engaged in user-centric decision making — in a word, these are humans interpreting the personas’ information for a given use case.
In contrast, the existing NLP studies seem to focus on “linguistic personas”, not on user personas or design personas. These linguistic personas can be seen as pseudo-personas based on linguistic patterns. A pseudo-persona means that that these NLP personas are not develop to the extent that they present rounded persona profiles with human-interpretable information about users, such as goals, interests, demographics, etc. An NLP persona does not have a name, gender, or interests. Rather, it is a vector of numbers (embedding), potentially useful for machine-learning tasks but possibly useless for a human designer.
Technically, there may be similarities in how (data-driven) user personas and NLP-based personas are created. For example, both may be based on data dimensionality reduction. In data-driven user personas, however, the critical step is what happens after dimensionality reduction — namely, personification by adding name, picture, and other information to create complete (also called rounded) persona profiles. In NLP personas, no such step is ever conducted (at least in any research paper that I have personally seen).
This makes me think NLP personas are not meant for human decision makers, but instead their purpose is to create user models that other algorithms could leverage in downstream tasks, e.g., to model or predict user behavior. (Also see our discussion about the role of personas in automated decision making.)
The above points raise a couple of important questions for using the concept of personas in NLP studies:
(1) First, it is worthwhile to ask, why not just refer to linguistic patterns instead of using the concept of personas? I.e., what separates personas from linguistic patterns? Is “persona” just misleading vocabulary when deployed to refer to linguistic patterns?
(2) Personas in HCI are used for understanding people, i.e., to form empathetic understanding of user behavior towards creating user-centered design outputs. What is the equivalent purpose for NLP personas? Again, is it necessary to refer to “personas” or should one use terms, such as “user model” or “user embedding”?
Conclusion
Using personas for NLP tasks is an interesting path but fundamentally different from the HCI paradigm of personas, where personas are created for human designers and not for algorithms. The use of “persona” concept in NLP research can be confusing for an HCI researcher, prompting NLP researchers to consider using other terms such as user model and user embeddings. Particularly the concept of user model is a broadly established term when discussing implicitly modeling users for downstream tasks such as personalization and recommendation.