Week 8 (August 15th - August 19th)

August 16th:

Previosly my mentor and I discussed changing the word US Representative to US House Representative hoping the specificity would help with the skewed results. Thus, I ran the hard tests on each category again. I ran it over on the hard test by type of politician as well as by gender and political affiliation. From the results it showed that using US House Representative helped tremendously in how CLIP classified certain images. Along with this sice the initial sentiment tests where so close in percentage spanning anywhere in the 40 percents it was challenging to analyze the results. Therefore, I created a table that showed the different results based on each specific prompt. Here were the results:

August 18th:

Today I kept working on creating the different scripts for the PKS tests, refining it to what Vicente and I talked about. After a long day I was able to create a draft website of the PKS Scores. Moreover, I read the recommended article Debiasing word embeddings using PCA. This article talked about how machine learning models have some textual biases relating to professions. Based on the image given to vision and language models they can confuse men for a male-dominated profession and confuse women for a female-dominated profession althought that is not what the image is of. This related back to my work as in my Basic Testing the Medium Test consisted of several different professions that could be confused for women and men politicians.

Upcoming

• Finish Calculating PKS Scores
• Create HTML Page for PKS Scores
• Brainstorm how to better analyze sentiment tests