Experiment 3: Markov Chains

I thought it would be interesting to use a markov chain to generate some predictive text. For this week's experiment, I decided to gather my personal messages and see what predicative text would be generated based on those personal messages.

In order to do this, I downloaded all my Facebook Messages and cleaned my conversations up. I went through all the json files and copied all the messages that I sent to a csv file. Now that I had all my messages, I implemented a markov model using Markovify (https://github.com/jsvine/markovify). Next, I used Flask to create a web server and hosted it on Amazon Web Services (specifically, Amazon Lambda - to make a serverless web application). Then, I generated a phone number using Twilio; my goal was to allow someone to text the number and get a reply back with a 200 character response of something I'd say. I used the API endpoint generated by Lambda as a webhook for the Twilio SMS API. If you text 631-318-6563, you'll get a text back with something I'd say that was generated using a markov chain. I am using Twilio via a trial account, so only approved/verified numbers receive replies, unfortunately. The reply messages are very interesting, though! For some examples of replies, visit: https://jb2du5s042.execute-api.us-west-2.amazonaws.com/dev.

Comments

Post a Comment

Popular posts from this blog

Assignment 9

This week's reading draws an interesting parallel between machine translation and natural language. At one point, the author argues that there is an important difference between programming languages and natural language, arguing that we can never truly have perfect machine translation. I thought that this was an interesting parallel because language is grounded in rules and programming languages consist of a series of rules. Thus, we ought to be able to reconstruct something that makes up a series of rules using other rules. The author treats machine translation as a black-box, whereas we know exactly how these models are being generated. The program is being fed a corpus that contains a body of text and the "right" translation. Hence, we have complete control over how texts are being translated; the only necessary component to change is the corpus to obtain a better translation. Knowing this, we are able to produce very literal translations. The author uses the concept
Read more

Final Project Update

I made a good amount of progress on my final project. I scraped all the articles from the Brown Daily Herald and used a sentiment analysis classification tool to assign each of the articles a score (https://github.com/aanshshah/dla_final_project). Next, I started the scene that I am going to be using for the project in a-frame. I created a 'wormhole' and the article text is read to the user via a text to speech API while the user is brought further into the 'wormhole'. The user is stuck in the wormhole during the entire experience as any interaction with the webpage repositions them within the wormhole. The next tasks I need to complete are adding articles from 2009-present to the experience, changing the background color, changing the wormhole color, and adding more interactions to the web-experience. The current version iteration of my final project can be viewed here:  http://aanshshah.com/dla_final_project/
Read more

Experiment 6

Many polls have shown that liberals view Starbucks more positively and that conservatives champion Walmarts. I decided to explore how the ratio of Starbucks to Walmarts per state aligned with political ideology to see if any interesting trends emerged. View Here
Read more