Skip to content

Gender bias and prompt tuning

Systematic Review

Research Questions

  1. How is the gender classification bias problem addressed?
  2. How did they measured bias?
  3. What is the main contribution, did they prompt engineered?
  4. What are their baseline works?
  5. Database: name, is it public, private, balanced, bias focused?
  6. Which VLM did they use, which backbone and data source?
  7. Did they tried to replicate CLIP results?

Research String

keywords: Gender bias, VLMs, CLIP, prompt tuning

String 1: ("prompt tuning" OR "prompt engineering") AND ("gender bias classification") AND ("VLMs" OR "CLIP")

Inclusion / Exclusion Criteria

#### Inclusion Criteria

  1. Must address societal bias
  2. Must employ VLMs
  3. Must be a full paper
  4. Must be submitted to conference or journal

Exclusion Criteria

  1. Paper not in English
  2. Paper is older than 5 years
  3. Paper is poorly written
  4. Paper is a thesis or workshop

Research Protocol

For a systematic review, a research protocol is in order. Here we will define the steps required to efficiently progress through the research.

  1. Explore the digital library search engine
  2. Adapt the search string and tune in search parameters
  3. Evaluate initial results
  4. Collect and export results

After that, we should have a tabular data structure with all the retrieved papers. All that is left is filter them for the truly relevant ones, and here are the steps.

  1. Filter by title
  2. Filter by abstract
  3. Filter by abstract, introduction, and conclusion (first pass)
  4. Filter by Inclusion / exclusion criteria (second pass)
  5. Read full paper

Science Databases

We will start from Google Scholar and then extend it if necessary.

  1. Google Scholar
  2. CORE
  3. Science.gov
  4. Scopus
  5. ACM
  6. IEEExplore

Google Scholar

String: (CLIP OR VLMs) AND (prompt engineering OR prompt tuning) AND gender classification bias -> 801 results

(A lot of the results are about text-to-image models, probably due to "prompt engineering / tuning" in the search string)

Title filter: 101

Abstract Filter: 69

Filter by Criteria: 12