g-simmons / persona-research-internship

1 stars 0 forks source link

Proj: Generate example movie reviews #51

Open AbhiramBorra opened 1 year ago

AbhiramBorra commented 1 year ago

Generate example movie reviews. Some examples in each direction (Direction, Acting, Cinematography)

AbhiramBorra commented 1 year ago

Example

Direction: Abhiram Borra's direction in The Dreams is nothing short of exceptional. He masterfully guides the narrative, maintaining a perfect balance between suspense and emotional depth. His ability to create a dreamlike atmosphere throughout the film is truly remarkable. Borra's storytelling unfolds like a carefully orchestrated symphony, with every note striking a chord in the audience's hearts.

Acting: The ensemble cast's performances in The Dreams are a testament to the power of exceptional acting. Each actor brings their character to life with such authenticity that it's impossible not to become fully immersed in their world. The chemistry between the lead actors is palpable, delivering an emotional resonance that resonates long after the credits roll. Standout performances from the entire cast elevate the film to a level of cinematic brilliance rarely seen.

Cinematography: The cinematography in The Dreams is nothing short of breathtaking. The film's visual compositions are a visual feast, capturing the audience's attention from the very first frame. The use of colors, lighting, and camera angles is sheer artistry, creating a dreamy, ethereal quality that complements the story beautifully. Every scene feels like a meticulously crafted painting, and the cinematography is a vital element in making The Dreams a feast for the eyes.

g-simmons commented 1 year ago

@AbhiramBorra Could you add a brief writeup of instructions for how to write the edits?

Basic idea was something like edit only one region to achieve low sentiment score on that region, leave the other regions unchanged?

AbhiramBorra commented 1 year ago

The basic idea is to edit only one aspect of the review such as the Direction, Acting, or Cinematography while leaving the other two unchanged. Additionally, every edit we make should change the text and the next edit made should add on to the edited text.

AbhiramBorra commented 1 year ago

Example

  1. Abhiram Borra's direction in The Dreams is very good(Edit 1). He masterfully guides the narrative, maintaining a perfect balance between suspense and emotional depth. His ability to create a dreamlike atmosphere throughout the film is truly remarkable. Borra's storytelling unfolds like a carefully orchestrated symphony, with every note striking a chord in the audience's hearts.
  2. Abhiram Borra's direction in The Dreams is very good(Edit 1). He smoothly(Edit 2) guides the narrative, maintaining a perfect balance between suspense and emotional depth. His ability to create a dreamlike atmosphere throughout the film is truly remarkable. Borra's storytelling unfolds like a carefully orchestrated symphony, with every note striking a chord in the audience's hearts.
  3. ......

The general idea is to gradually edit the text and change the sentiment from a very positive review to a neutral review and finally to a very negative review. (Repeat for each section)

AbhiramBorra commented 1 year ago

Link to document: https://docs.google.com/document/d/10kqgfjSLBSMfI49zFGEb2w6yQgJi8mAPflKYAmuuAao/edit?usp=sharing

g-simmons commented 1 year ago

@AbhiramBorra Looks like good progress 😄

What do you think about moving to google sheets? I am suggesting this because it will be difficult to interact with Google Docs programatically. At some point we will need to download this data as e.g. a CSV file to use it in other parts of the project.

g-simmons commented 1 year ago

I would also suggest keeping the different components of the prompt in separate columns and concatenating them programmatically when needed.

g-simmons commented 1 year ago

@AbhiramBorra Here is a google sheet with what I'm thinking

https://docs.google.com/spreadsheets/d/1HC6ZK2Lo40qR3V1ttZdpKhi-uQZ2BEb5AekPkOxzWUc/edit#gid=0

Explanation:

original_<component> stores the original sentences for each component (direction, acting, cinematography) edited_<component> stores the potentially modified sentences for each component is_<component>_edited are boolean columns indicating which components have been edited sentiment_scorer is the name of the sentiment model used to calculate sentiment scores original/edited_<component>_sentiment are the sentiment scores for each component

Why are we doing it this way?:

You may be thinking "aren't some of these redundant?", and indeed they are. The idea is to prioritize machine operability over storage space. For example, having the is_<component>_edited columns allows us to quickly sort the google doc to see all the rows with edited cinematography.

"Isn't sentiment_scorer going to be the same for the whole spreadsheet?" Also yes :). Again, we are prioritizing machine operability and, in this case, thorough experimental records over storage space. Having the scorer listed on every row makes it (painfully) obvious how we obtained the scores.

Other notes:

"Keep things machine-readable" is generally a good rule to follow. We can move faster and be more thorough when it's very easy to do calculation and analysis, run experiments, etc. Having everything in standardized machine-readable formats is one of the most important practices to enable this.

If you've taken a class in database systems you might recognize that we're storing the data in denormalized form. Denormalization is a "move" that database admins can do to increase read performance at the cost of write performance. Same tradeoff here :)