Quantifying Writing Quality With Nlp And Sociolinguistics

---

I am a graduate researcher at MIT CSAIL studying human perception and assessment of writing quality by leveraging data from a short story competition (user submitted written prose and their associated rankings). Currently, there exists no objective linguistic metrics that capture what it means for something to be well written, and even state of the art NLP models lack quality prediction capabilities. However, humans seem to have an innate understanding of what writing quality signifies. We aim to quantify what it means for something to be "well written" by analyzing both qualitative and quantitative data from the competition, as well as through the development of state of the art NLP models. Indeed, we can learn much by analyzing both how humans score various short stories and comparing it to how NLP classification models perform on such task. Lastly, studying how a plethora of psycholinguistic features and differences affect scoring might give us interesting insights as to how language affects writing perception.

• Date: May 2023
• Service: Masters Research

Close

A Noisy Channel Framework For Sentence Comprehension In Spanish And French

---

Linguistic communication is no easy task, and is very often subject to noise. Speakers can make errors or introduce ambiguity, and our environment can introduce other sounds which might make it difficult to comprehend one another. Nonetheless, people can generally understand each other easily, therefore suggesting that the way humans process and understand language accounts for the omnipresent noise in our lives. The noisy channel framework is a model that explains how humans are able to make inferences about speakers' intended meaning when utterances have been subjected to potential noise. Given prior information and a noise model, that is, what they know about the world, people can indeed infer more than what they simply perceive.

My project endeavored to apply and evaluate the noisy channel framework to the French language, and replicatd recent results where the model was tested on the Spanish language.

See my full paper here!

• Date: May 2023
• Service: Computational Psycholinguistics

Close

Light Echo: Exploring The Relationship Between Art And Neuroscience

---

Perception is malleable. By using different speeds, motions, and light frequencies, my piece Light Echo illustrates the illusory nature of object perception. Objects exist within a rich and complex environment, and they may look different based on the perceptual parameters applied to the object. In this piece, the invisible relationship between motion and light is brought to the realm of the physical by creating shape deformations, what I called the “light echoes” of the rotating cube, where the eponymous name of the piece comes from.

The goal of my perceptual experience is to observe the cube projection, and watch how light helps construct and deconstruct the objects that we perceive. On the other side of the projection, a wireframe cube is attached to a rotating motor turning at around ~50 rpm. Moreover, a red laser line created from a laser pointer and cylindrical optic lens is reflected by a mirror octagonal prism. This laser line reflection sweeps the cube as it is itself turning onto itself. The result? Projections of deformed and contorted shapes. Can viewers figure out what the original shape was?

This piece is on display in the MIT Museum studio as part of the Garden of the Lights exhibition.

Watch an excerpt of the experience here!

• Date: December 2022
• Service: Art & Neuroscience

Close

Revealing Surprisal Neural Representation In Language Processing

---

I investigated the neural basis of surprisal by analyzing relationship between cognitive activity and surprisal scores, given 55h+ of intracranial recordings and annotated treebank. In this project, we computed word probabilities for our surprisal scores using LSTM, n-gram, and GPT-2 models. Above are raster plots and graphs of the neural voltage activity in the superior temporal gyrus region of the brain for the movie Cars 2. Surprisal was computed as the negative log of word probability, using word probabilities from the GPT-2 language model.

See my full paper here!

• Date: October 2022

Close

Game Theoretic Modeling Of Patrolling Stackelberg Games For Mas

---

For my Cognitive Robotics final project, I worked on creating simulations of multi-agent systems whose aim was to patrol a house. We use game theory, and specifically Stackerlberg games to frame our project, and solved for the optimal policies using a Decomposed Optimized Bayesian Stackelberg Solver (DOBSS).

With Bayesian Stackelberg games we can frame a patrolling scenario as a solvable problem. Normally, those who are patrolling are at a direct disadvantage compared to those who are trying to attack, since they have to make their move first and they can be observed. However, this exact scenario can be framed as a Stackelberg game where the leader is the security doing patrol and the follower is an attacker. While accounting for possible attacker types, the optimal policy for the leader is solved with DOBSS. For games where number of valuables, length of patrol, or number of attacker types remains under 5, the game is solved on average under .10 seconds.

See the full report here!

• Client: Cognitive Robotics
• Date: May 2022
• Service: Robotics simulation

Close

Statera: Researching The Hidden Link Between Art And Mathematics

---

In late high school, I spent some time studying the dichotomy between mathematics and art. The math and art couple always seemed to be incredibly intersting to me, and art and mathematics are not often thought of together. Art is often seen as something beautiful, from a purely aesthetic POV. We often forget its codified and strict aspects. It is a subject that is arguably close to mathematics, which on the other hand, is a discipline meant to demonstrate facts about the world meaningfully and logically. However, math can also be seen as beautiful. As a result of these thoughts, I began my research project. I worked with Stanford lecturer and mathemartist John Edmark, wrote a 30-page dissertation, and created a 3D printed sculpture on SketchUp, aptly named Statera, the balance.

• Date: March 2018
• Service: Art & Mathematics

Close