A list of novel but doable projects in natural language processing
Inspired by YCombinator’s Requests for Startups, here is a list of novel but doable projects in natural language processing.
The dataset for a project may not exist yet but we know how to generate it at scale with a bit of scripting.
Given a URL, predict whether or not it is an image.
Given an HTML element, predict whether or not it is an advertisement that should be blocked. (See rules files for adblockers)
Language identification for written variants of a language like British, American, Canadian and Indian English
Given an HTML element, predict whether or not it is main content (like Chrome, Safari and Firefox for mobile, see Readability and AMP)
Identify propaganda, hate speech and threats of war to monitor regimes and predict conflict
Given a string, fix the encoding to make it more human-readable if necessary
Given HTML, generate Markdown (.md) Given text (without formatting, eg from copy-pasting a website), generate Markdown (.md)
Given a GitHub repo, generate a README.md
Given a README.md, generate a GitHub repo
Given a GitHub Pages-enabled repo, generate HTML
Given a screenshot of a website, generate HTML (What to do about images?)
Given a research paper, generate a blog post
Given a blog post, generate a research paper
Given a question on StackExchange, generate an answer or comment
Translation between variants like British and American English, or Eastern and Western Armenian
Generate and/or disambiguate between variants and translations of proper names - person names, placenames and organisation names
Transliteration for multiple languages and scripts in one model
Translation for multiple language pairs in one model
Transcribe and index the audio of a large set of YouTube videos so that they are searchable
Augment text data with realistic noisification
Build text data and pre-trained models from arXiv papers, to find quality authors, topic trends, previous work… and slice by time, region, university…
Build text data and pre-trained models from GitHub repo READMEs
Sentence representations
Combine or diff fastText models trained on different datasets or with different parameters
Visualise/interpet a text classification model’s predictions 1) on a row/sentence 2) in aggregate for a model and dataset 3) across models differing by epochs or other hyperparameters
Visualise/interpet a seq2seq model’s output 1) on a row/sentence 2) in aggregate for a model and dataset 3) across models differing by epochs or other hyperparameters
Modify fastText test to list out rows from the test set that were most strongly assigned each label, the rows most strongly assigned the wrong label, and the rows barely assigned a label
Diff fastText models trained on different datasets or with different parameters
Accent reduction - for foreign or regional accents