This week, I finished the conversion of the prompts from singular to plural. This took me a while but it finally worked. Now that I have all the code and all the peices, I was able to obtain the data and create the graphs for the final report. I started writing the final report this week. I realized that it was a good idea to keep all of your sources in one place before writing the paper as it was a pain to recollect all of the sources and cite them while writing the paper.

This week, I finished the scores to calculate the scores of all the output tokens. I am almost done with the conversion of of the singular prompts to plural prompts. I realized that the spacy library has a function that can be used to stem a word and this could be used to convert a supporting word from plural to singular. I would probably need to go through all of the sentences once again once they have beeen converted but the code kmakes my job easier.

This week I experimented a little more with the SpaCy library. I tried different types of sentences and they had different types of relations. One of the most common changes that I need to make in addition to changing the pronoun is changing the auxialiary verb in the sentence. However, a simple if statement will not suffice. I tried it and it ended up changing the whole meaning of the sentence. So, I am going to have to experiment a little more and find a way to make this conversion a little better.

This week, I realized that a bulk of my experiment came from turning the sentences with singular pronouns into plural. This was a problem because I would have to manually change the grammatical structure of all the sentences. This means that there were 800 sentences that I would have to go through. Unfortunately, I could not find the code to change from singular to plural online because of the inconsistent nature of natural languages. However, I learned about a library called spacy; spacy helps you find the grammatical and syntactical relationships between words. I think that I would be able to take advantage of this to convert the sentences from singular to plural.

This week, the professor asked me to pass all the prompts in the Winobias dataset into various large language models. I did not know how to automate this as I have never used Python before. So, I started passing all of the prompts individually. I thought that it would help me notice a pattern better if I looked at the individual outputs from the LLM. Clearly, this was a bad idea. There were around 1600 prompts total and I also had to experiment with all 1600 of these in their third person plural version as well. I was able to do around 30 by the end of it.

Upon further consideration, I realised that I am more passionate about this other porject idea that I had. It was the bias that non-binary people face because of the training of ML models. I discussed the idea with Dr.Ordonez and he said that it was a good idea. He pointed me to a test set of prompts that his lab previously worked on. This was called the Winobias dataset. I am still going through that paper and deciding what to do with it. The professor thinks that it would be a good idea for me to run the prompts by the LLM’S and acl;culate the accuracy of the coreference resolution.I am currently working on that

This week, I met with the professor and we spoke about the kind of projects that I would be interested in. I was speaking to Veronica about the same and she told me about how the project that she was working on had to do with language visual models and gender bias. I thought the idea of ML and gender bias was really interesting to me. We have spoken about a concept called social computing at UCSD; the idea of ML models being biased because of an inherent bias in training data seemed really interesting to me. I conveyed the same to the professor.

This week I met with Dr. Ordeonez. We talked about the kind of papers I would be interested in writing. I went through some of the projects that the lab was currently working on. They all seemed vey interesting. I was going through the descriptions for all of them and Veronica, the other intern under Dr. Ordonez, toldd me about what she worked on last summer and what the other intern last suummer worked on.

After a week of hard work, I finally finished the paper. I also had a lot of ideas on how to better the Winobias dataset and spoke to the professor about it. He said that I could continue working on it under his guidance in the same lab. I was beyond grateful for the opportunioty. The idea that I had was to elongate the prompts so that there would be a greater distance between the pronoun and the subject. This would really test the co-reference capabilities of the model.

During the first week, we spent out time going though some literature and reading papers. I read some papers about something known as the CLIP model. This model was a model that essentially a model that takes in an input prompt and outputs an image that matches the input prompt. I learned a lot about zero-shot processing and other machine learning techniques that I think will be useful for my project later on. I also read some papers about neural networks and the likes of them.