Words move people.

Think about the last time you read a job description. This is the first front between you and the rest of the company. The analagous for researchers is the paper abstract. Like a company, researchers want to connect to people through their work. Unfortunately, most people don't get past the paper abstract. So the question is: how do we write abstracts that are not only concise and compelling, but also connects us to the right person? This is what I built Velma for.

Velma is a pipeline for 45 million research papers that provides examples of high quality abstracts from top papers in Computer Science, Neuroscience, and Biomedical.

Say I'm a computer scientist writing a paper on reinforcement learning. I want to pull up a few examples of abstracts so I can have a sense of what works well. Say I know the abstract from "Mastering the game of Go without human knowledge" has a great abstract. I'll provide the title "Mastering the game of Go without human knowledge" to Velma, and Velma will provide me not only with the abstract from "Mastering the game of Go without human knowledge", it will also provide me with 5 additional abstracts that are similar to the abstract from "Mastering the game of Go without human knowledge". Now, I have not one but six examples of good abstracts I can use as templates to write my own abstract. If you've convinced your reader to continue past the abstract, our work is done and yours is just starting.

Click here to see a demo: Velma

We use a lot of papers. Papers are collected from the Open Research Corpus (~45 million) from three main categories -- Computer Science, Neuroscience, and Biomedical -- and are stored as raw text files in AWS S3. Then, the papers are preprocessed into a large dataframe, where each row is a paper and each column is a feature of that paper like id, title, abstract, number of citations, and tags. Spark is used because I want to distribute the work to multiple workers, saving time and solving out of memory issues. After data driven filtering by number of citations and relevant tags, Velma compares the simililarity of abstracts using the Jaccard Index, a measure of word choice similarity. For each abstract, I find the top 5 most similar abstracts and store these as key value pairs. Redis is perfect for this because of its key-value based store and constant time access. To display Velma's abstracts, I use Flask because of its template based development, allowing for a wide range of customization as well as "just working".

Velma will provide abstracts using the common language of your field, so you can be confident that you're using the right words.

7 orders of magnitude. Comparing the jaccard index of each paper to every other paper is O(n^2). With 45,000,000 research papers, this task is impossible -- on the order of 2 quadrillion computations. (That's a 2 followed by 15 zeros!) So what next? Taking a closer look at the data, I determined three filters that not only reduced the number of computations but also increased the quality of abstracts.

• The data isn't very clean. Roughly 35% of papers from the data did not contain abstracts. So I drop them.

• Only about 13% of papers have more than 10 citations. Since we're in the business of top papers and its much more likely that a user will give Velma a high quality paper, I decide to only store these top papers in our database.

• You don't want to waste your time comparing papers about gene splicing to papers about database architecture. For this reason, I only compare papers that share at least one tag. This way, only papers that are relevant to each other will be compared to each other.

With these three filters, I am able to reduce our computation by 7 orders of magnitude from 2 quadrillion to 900 million computations, distributed across 8 Spark workers. This is like reducing the size of the Earth down to the size of the mystery machine. I was not only able to reduce our computation time by a significant amount, I was also able to provide higher quality abstracts to our user. Everyone is happy!

Input the title of a paper and Velma will provide the abstract from that paper and 5 of the most similar abstracts to it. For example, the user enters the paper title "Mastering the game of Go without human knowledge". Velma will then search through the database for that paper. If the paper exists in the database, Velma will display the 5 most similar abstracts to the abstract from "Mastering the game of Go without human knowledge".

The user will enter a keyword(s) that they want in their abstract. For example, the user inputs = "deep learning". Velma will scan through the database, pull every abstract with the keyword "deep learning", sort them by highest citation, take the top 20% of these papers, compute the jaccard index, and display the 5 papers with the highest jaccard index.

A user will enter a tag(s) and Velma will provide the top 5 abstracts with those tags. For example, the user inputs the tag "biomedical". Velma will look through only the abstracts with the tag "biomedical", compute the jaccard index for the top 20% of papers and display the top 5.

Allows the user to filter by year. Ex. only look at the papers from 2015-2019.

Different journals, even on the same subject, look for different kinds of abstracts. Velma will allow you to filter by journal so you can write your abstract in a way that targets whichever journal you are submitting to.

There are potentially millions of other papers from other domains on sites like arXiv, Academia, and PubMed (> 200GB)

Expanding Velma’s service to other summary type writings like job descriptions, executive summaries, and intros to articles.

There are many other features to a paper that indicate quality papers and possibly good abstracts. Velma would like to be able to weigh your abstracts suggestions by author, out citations, and title.

Provide examples and learn how to recommend good words and style for people. Velma could recommend ways to write your abstract, intro paragraph, or executive summary while keeping your distinct voice.

A big thank you to Insight and the DE SEA 19A Fellows for your help and constant feedback, without which this project would not have been possible. Special thanks to Sriram and Curtis and the rest of Insight Data Science staff.