This module enables analyzing the text content of webpages, including with natural language processing methods. The module uses Mozilla Readability in a content script to parse document title and content when possible.
Training, Testing, and Deploying Natural Language Processing Models
A motivating use case for this module is applying natural language processing methods to webpage text. The module provides infrastructure for NLP models, but leaves implementation and evaluation of models to study authors. We recommend using existing toolkits for NLP feature generation (e.g., Natural or NLP.js) and for working with models (e.g., TensorFlow.js, ONNX.js, WebDNN, or sklearn-porter). We also recommend using the same codebase for collecting data (e.g., with web crawls), constructing models, evaluating models, and deploying models in browser-based studies. When maintaining multiple NLP codebases for a browser-based study, subtle inconsistencies are easy to introduce and can call into question NLP model performance.
Web Crawls to Collect Natural Language Processing Training Data
Because WebScience integrates with ordinary browser extensions, you can use this module in a web crawl to collect page text content as NLP training data. All the major browser automation toolkits (e.g., Selenium, Puppeteer, Playwright, and WebdriverIO) support running web crawls with browser extensions installed. We recommend running an online crawl to collect NLP data, using this module to extract webpage text, then training and testing models offline. If you use web crawl data to construct an NLP model for a browser-based study, be sure to carefully consider how the distribution of pages in the crawl compares to the distribution of pages that a user in the study might visit. If a crawl is not representative of user browsing, NLP model performance on crawl data might significantly differ from performance when deployed in a browser-based study.
Implementing Natural Language Processing in Web Workers
Because natural language processing methods can be computationally
expensive, it is very important to offload NLP tasks from an extension's
main thread. We recommend pairing this module with the workers module to
implement NLP tasks inside of Web Workers, which run in separate threads
and will not block the extension's main thread. Some NLP toolkits support
additional optimizations, such as WebAssembly or WebGL, and we recommend
enabling all available optimizations to minimize the possibility of impact
on the user's browsing experience.