Messy Web Text Project: Green Claims and Climate Communication

• Project name: messy_text
• Research question (example): Do company sustainability pages differ linguistically from public-interest climate information pages?
• Programming language: python (suggested) or R (allowed)
• Expert contact: TBD, Anastasia?

Canonical course conventions live in project_guidelines.md. That file is the source of truth for the four required workflow files (week1_explore.qmd, week2_operationalize_clean.qmd, week3_model.qmd, week4_storytelling.qmd), the data/model_data.rds -> data/model_results.rds pipeline, the raw-data policy, quality-check requirements, decision logs, and contribution tracking. Read it before starting and treat anything below as project-specific guidance on top of those conventions.

Tutorial framing

Web text is complex because the data arrive wrapped in markup, navigation, scripts, boilerplate, duplicated page elements, and inconsistent page structure rather than as analysis-ready documents.

Students should learn three main things about these data:

1. How web text is produced and represented through HTML, DOM trees, URLs, HTTP requests and responses, CSS, JavaScript, metadata, and page templates.
2. How to turn raw pages into a clean corpus or analysis table by choosing a unit of analysis, extracting meaningful text, removing boilerplate, preserving source metadata, and documenting text-cleaning choices.
3. How extraction, tokenization, repeated page elements, and publisher purpose affect linguistic features, models, visualizations, and the claims that can be made from a small web corpus.

Peer-teaching checklist

Dimension	This project teaches
Data structure	HTML documents, DOM trees, text corpus, nested page metadata, and document-feature or document-term representations (which are sparse matrices — the same data structure used for network adjacency).
Storage system	Raw downloaded HTML files, with NoSQL/document-store storage such as MongoDB discussed as an optional comparison rather than a required implementation.
File formats	HTML, JSON metadata or exports, TXT, and CSV/RDS-style clean analysis outputs.
Encoding	UTF-8 text, HTML markup, and JSON serialization for metadata or document-style records.
Model	Group comparison, logistic or linear regression, clustering, or another small interpretable model using transparent text features.
Key aspects to explain	DOM structure, HTTP status codes, robots.txt, scraping legality and ethics, boilerplate removal, tokenization, document-term or TF-IDF features, and sensitivity to extraction choices.

Resources

Data sources

• Raw HTML pages from corporate sustainability pages and public-interest climate information pages.
• Possible corporate sources: Shell, ExxonMobil, TotalEnergies, or other firms identified through Orbis or a similar source. [TODO: to be downloaded before the course]
• Possible public-interest sources: UN climate pages, National Geographic, government climate pages, or climate-focused NGOs. [TODO]

Knowledge sources

• Basic HTML and DOM tutorials.
• Python packages such as requests, webSweep, beautifulsoup4, TBD
• R packages such as rvest, TBD

Week-by-week

Week 1

Inspect raw HTML, explain who published it and why, and identify the DOM structure and markup noise that matter for extraction.

• What is HTML? How does it relate to CSS, JavaScript, and server-side systems such as PHP?
• How do users and scripts interact with websites through HTTP or HTTPS requests?
• What is the unit of raw data in this project: a page, a paragraph, a text block, a sentence, or something else?

Prepare for roundtable in week 2:

• What is the advantage of markup languages such as HTML? Where else are they used?
• Explain why raw HTML is not the same thing as clean text, and introduce the DOM tree and how to use it to extract information.
• Explain the practical and ethical constraints of web data collection, including error codes, scraping legality, and the role of robots.txt, even when the project uses already-downloaded pages.
• Explain how publisher type and page purpose shape the data-generating process, and how the unit of analysis was created from messy pages, for example by extracting comparable paragraphs or text blocks.

Week 2

Operationalize the question by turning raw pages into one analysis table with transparent text-cleaning choices.

• What, exactly, counts as corporate climate communication or public-interest climate communication?
• Which parts of each page should be kept or removed: headers, menus, cookie banners, captions, footers, links, boilerplate, and repeated slogans?
• Which text representation fits the research question? Start with a transparent count-based or TF-IDF representation. Embeddings are an optional extension only if time allows and only after a count/TF-IDF baseline has been built and interpreted.
• The document-term matrix you build is typically extremely sparse (most documents do not contain most terms). This is the same sparse-matrix concept that the Networks group teaches with adjacency matrices — note this connection so the two groups can teach it jointly.
• Optional cross-modality reflection: how is turning text into a model table similar to turning images, audio, or video into model inputs (pixels, spectrograms, frames, embeddings, labels, or extracted features)? Skip if time is tight.
• What source metadata should stay attached to each unit, such as publisher type, URL, date collected, page title, or page section?

Prepare for roundtable in week 3:

• Explain the basic NLP ideas used in the project, such as tokenization, hand-built lexical features, or document-term style representations, without overselling their sophistication.
• Explain why the document-term matrix is sparse and how that connects to the sparse-matrix idea also seen in network adjacency matrices.
• Optional: compare text with images, audio, and video — what is the raw file format, what counts as the unit of analysis, and what choices are needed before the data become a table or model input?
• Explain why a simple text model can still be useful if the feature engineering is transparent and tied to a substantive framing question.
• Explain how extraction choices, token choices, and repeated text from the same page can change the model result.

Week 3

Fit a small interpretable text model using hand-built features or simple document representations, and show sensitivity to extraction or tokenization decisions.

• Is the goal to describe linguistic differences, classify source type, or estimate an association between publisher type and framing?
• Which features answer the research question: risk terms, urgency terms, technology terms, efficiency terms, sentiment, modal verbs, or topic-like clusters?
• Build the baseline model on count or TF-IDF features first. Only add an embedding-based representation if the baseline is fully working and there is real time left.
• Which model or comparison is small enough to explain clearly: group differences, logistic regression, linear regression, clustering, or a simple classifier?
• How do results change if boilerplate removal, tokenization, stopword choices, or repeated pages are handled differently?

Prepare for roundtable in week 4:

• Explain what the model is actually learning from the text and what it is not learning.
• Explain which visual summary best teaches the contrast across sources: keyword frequencies, coefficient plots, or source-level framing differences.
• Explain the main limitation of drawing broad conclusions from a small purposive web corpus.

Week 4

Visualize and tell a story about the contrast between sources, while making the limits of the corpus and preprocessing choices explicit.

• What is the context? What is the main result? Why is it important?
• Which visualizations support the finding without pretending the model understands the full meaning of the pages?
• Which extraction, labeling, or source-selection choices could change the story?
• What are the assumptions and limitations of your design?