杰瑞·韦的斯坦福录取申请案例 | 深度解析

George Orwell didn’t just write dystopian novels, he also inspired me to apply my love for computer science in the most unexpected of places: studying linguistics. Let me elaborate—during my freshman year, I read 1984 and discovered the language of Newspeak, in which words such as “outstanding” and “wonderful” were replaced with unambiguous words such as “doubleplusgood.” By enforcing a limited vocabulary, Newspeak prevented the working class from thinking anything that diverged from the ruling party’s beliefs, essentially acting as a vessel to promote an implicit agenda. Through my morning routine of keeping up with the news, I noticed a trend in today’s news that was eerily similar to Orwell’s warning of linguistic totalitarianism. News sources often report on the same topic but present information in a way that only supports their own viewpoint, implicitly pushing their political agenda on readers. Soon I began to wonder: is there a way to identify how systemic and implicit biases manifest in today’s news? Strikingly, the answer came from another passion of mine—computer science. In my sophomore year, I took an online Coursera course on machine learning in which I learned about the recurrent neural network, a statistical model that could automatically discover linguistic patterns in large amounts of text. I thought the model could potentially help answer my question, so I set out to gather a large dataset that I could use to train it. Over the course of two weeks, I scoured the internet and compiled a dataset of over 20,000 articles about Donald Trump, one of the most polarizing topics in today’s news. Each article came from one of four major news sources that each represent different points on the political spectrum according to Media Bias/Fact Check, a recognized website that categorizes the biases of news organizations. Now that I had my dataset, I used differentiable programming tools to create a custom model in Python for my project, debugging over a dozen failed configurations in the process. In the end, my model successfully predicted a given news article’s source with 85% accuracy by detecting the article’s hidden political biases. Now that my model could accurately identify an article’s source, I coded a way to extract how much bias it associated with certain phrases of interest so that I could interpret the patterns that it had learned. I found that the phrases it perceived as biased matched with what humans would expect. For instance, the model recognized that, when referring to Trump, liberal sources tended to use phrases with negative connotations such as “Trump lost because he” and “Trump has a history of,” whereas conservative sources tended to use titles of respect such as “President” and “Commander-in-Chief.” In addition to the newest statistical models, I also analyzed my dataset using two methods from classical computational linguistics. One popular concept, the age of acquisition (AoA), measures the age at which words are typically learned (for example, the AoA of “run” is 4.5 years old while the AoA of “abscond” is 13.4 years old). I found that in my dataset, articles from conservative sources, which typically target older audiences, had a higher average AoA than articles from liberal sources. In terms of lexical diversity, the proportion of unique words in a text, I discovered that sources that were categorized as more biased used a smaller proportion of unique words in their articles relative to more-neutral sources. This may suggest that more-biased sources tend to narrowly select words that support their agenda. By using both the latest statistical models and more-traditional computational linguistics methods, I discovered new insights on how bias manifests in today’s news. Although George Orwell might not understand the mathematics behind my model, he certainly shares my sentiment about the importance of language. As technology continues to improve, I hope to continue searching for new ways to use computational methods for analyzing different types of biases in language.

2019 was the most exciting year in computational linguistics since the beginning of the field in the 1950s due to the development of an algorithm called BERT. By learning from the entire Wikipedia database of over 2.5 billion words, BERT has acquired a wealth of general information about language and therefore can be used to tackle new linguistic tasks with only a limited amount of data. Even I experienced the effects of BERT. Before BERT came out, I had conducted research on using machine learning to detect political bias in news articles, a project for which I had to spend four weeks collecting and processing twenty thousand articles in order to sufficiently train a machine learning model. In my next project, which I worked on after BERT came out, I analyzed questions that people asked online about COVID-19, but was only able to collect a few hundred questions for my dataset. Whereas previous machine learning algorithms could not be sufficiently trained with such a small dataset, using BERT, I was able to perform complex linguistic analyses with surprising accuracy, despite the limited amount of data. After witnessing BERT’s effects firsthand, I was both astounded and inspired by how a single idea can change the way we do research in computational linguistics. In the future, I will continue studying computational linguistics and hope to one day develop my own algorithm that can allow others to utilize the power of computational methods to more accurately and accessibly study linguistic phenomena.

Hi there, I’m stoked to room with you! My brother left for college when I was thirteen, so I’ve had my own room since then. It’s nice to get to have a roommate again. Here are some fun “me” facts… I adore computers and anything to do with them. You’ll probably see me arranging my custom computer setup on my desk and plugging in a bunch of wires everywhere (I’ll do my best to keep the wires from turning into a big heap of wire monster, but no promises). I love linguistics, and I’m interested in how our use of language affects how we perceive the world. I just finished reading Metaphors We Live By (which is about how we use metaphors to understand abstract concepts), and I am now reading Syntactic Structures, a classic linguistics book by Noam Chomsky, the father of linguistics. Hope you’re okay with seeing boxes of protein bars—I’ve been powerlifting (a type of weightlifting consisting of the bench press, squat, and deadlift) for about four years now. Feel free to lift with me sometime! Personality-wise, I’m an architect, like Michelle Obama and Elon Musk. Architects are “imaginative yet decisive” and “curious about everything but remain focused.” Spot-on—one time my cat was scratching my door while I was writing an essay, and I finished writing the entire essay before opening the door to see what he wanted. Turns out, he just wanted to hop onto my window sill to feel the cool breeze of fresh air.

One alarming discovery that I made while studying computational linguistics is that artificial intelligence algorithms can be implicitly biased as a result of hidden biases in the data that they are trained on. As a classic example, computational linguistics algorithms often associate the word “doctor” with “man” and “nurse” with “woman,” even though gender is not referenced in the definition of “doctor” or “nurse.” When algorithms with implicit biases make it into the real world, there are severe consequences. For example, Amazon developed an algorithm in 2014 that analyzed resumes to identify which job candidates to interview. A year after deploying it, however, Amazon found that the algorithm tended to select men because it had an implicit bias that women were less qualified than men for positions in the technology industry solely because of their gender. These biased algorithms troubled me not only because they had invaded my home territory of computational linguistics, but also because they challenged an ideal that I was raised to believe in—equal opportunity. As computational linguistics algorithms improve and take on larger roles in decision-making, I have a duty as an aspiring computational linguist to work on ensuring that our algorithms are fair and give everyone an equal opportunity. As I continue studying computational linguistics, I will also continue to analyze bias in algorithms. My current research has already shown that algorithms can discover implicit political biases in language, and moving forward I plan to find new methods to actively reduce bias in computational linguistics algorithms.