Where should a post about a 2006 speech-recognition system for an under-resourced language go? Speech Recognition, Low-Resource Languages, and Research History all describe it, but giving each label equal weight makes the archive harder to browse. The first question is simpler: where would a reader look for more posts like it?
For a growing historical-technical blog, categories are the broad shelves of the archive. They should let readers browse a subject without knowing a particular paper, laboratory, or algorithm. A useful category still makes sense as the archive grows; it does not depend on one conference season or a few articles sharing an incidental keyword.
Start with the posts that actually exist
Inventory published articles and likely near-term assignments before naming categories. For each post, note its main research problem, the evidence it examines, and the neighboring posts a reader might want next. A spreadsheet with columns for working title, primary subject, possible secondary subject, and proposed category will do. This is an editorial exercise, not a tally of every term in the text.
Say the archive covers language identification, speech synthesis, speaker recognition, document layout analysis, object tracking, stereo reconstruction, and medical-image segmentation. Those posts point toward durable areas such as Speech and Language, Computer Vision, and Document Analysis. Medical Imaging might earn its own category as coverage grows. A category called 2005 Research, by contrast, puts unrelated problems together because they share a date. Dates belong in titles, timelines, or archive filters; they rarely make a useful primary shelf for readers following a technical subject.
Use a small test set
Assign ten to twenty representative posts before changing the live site. Include straightforward cases, borderline cases, and at least one planned article. If most have an obvious home, the scheme is promising. If a document OCR post could land in four categories with overlapping definitions, sharpen those definitions. If every planned post needs a new category, the categories are probably too narrow.
Do not turn a minimum post count into a rule. A young archive may have only two posts in a clear, recurring area. Ten posts, meanwhile, do not justify a category if all they share is a conference name. What matters is whether the category page offers a coherent reading path the publication intends to maintain.
Choose one principle for the top level
Top-level categories become confusing when they mix different organizing principles. Speech Recognition names a research task; 2000s names a period; Academic Conferences names a publication setting; Applications describes a vague purpose. A post might fit all four, yet still be hard to find. For this archive, research domains or recurring technical problems are usually the clearest basis for top-level categories.
Other dimensions still matter. An article might discuss hidden Markov models, an evaluation dataset, and work presented at a mid-2000s meeting. Those details can live in its title, prose, or controlled tags rather than compete with its main subject category. A reader opening Computer Vision should not have to guess which vision posts were filed instead by year or conference.
Write a one-sentence scope note for each proposed category. For example: “Speech and Language covers recognition, synthesis, language identification, and related evaluation of spoken or written language technologies.” Scope notes reveal gaps and collisions that names alone can hide. If Document Analysis covers OCR and page-layout methods while Computer Vision covers image understanding broadly, decide whether document posts primarily belong in Document Analysis, with Computer Vision used as an occasional secondary category.
Distinguish a category from a tag
A category is a dependable route through a substantial part of the blog. A tag, if the site uses them, marks a narrower connection across categories: hidden Markov models, low-resource languages, evaluation datasets, or a research method. Tags need limits too. Separate tags for “low resource,” “low-resource,” and “under-resourced” would create three nearly empty routes to similar material. Keep an approved vocabulary, and add a tag only when it connects posts readers would expect to find together.
A category can be narrower than a discipline once the archive has the depth to support it. Medical Imaging may become useful if posts regularly cover segmentation, feature extraction, registration, and evaluation in clinical images. One segmentation post does not call for a separate shelf. The existing post on medical image analysis before deep learning, for instance, could sit in Computer Vision while medical coverage develops.
Make borderline assignments deliberately
Historical research rarely respects tidy boundaries. Speaker recognition draws on speech processing and biometrics. OCR uses image processing but serves document analysis. Stereo reconstruction can overlap with tracking when a study follows objects in three-dimensional scenes. Categories should reflect those connections without putting every article everywhere.
Choose the primary category by the question the article asks, not every technique it mentions. A post about identifying speakers from short recordings fits Speech and Language when the audio problem is its focus. A broader comparison of identity-verification methods might fit Biometrics, if that category exists. For an OCR paper concerned with recovering page structure, Document Analysis is the stronger primary home even when image features play a major role.
Add a secondary category only when it offers another meaningful browsing route. If a stereo-reconstruction post substantially compares geometric vision and object-tracking methods, two categories may help. If tracking appears only in the introduction, an Object Tracking assignment misrepresents the post. A workable house rule is one primary category and a second only for substantial overlap; an editor should be able to explain that second assignment in one sentence.
| Post emphasis | Likely primary category | Possible secondary category |
|---|---|---|
| Recognizing an under-resourced spoken language | Speech and Language | Low-Resource Technologies, if established |
| Recovering text blocks from scanned pages | Document Analysis | Computer Vision, if methods are central |
| Segmenting anatomical structures in scans | Medical Imaging, if established | Computer Vision |
| Estimating depth from paired camera views | Computer Vision | None unless another domain is central |
These are editorial judgments, not fixed definitions of the fields. The same method may deserve a different home when an article asks a different historical question.
Name categories for readers, not filing systems
A category label should make sense without an explanatory paragraph. “Speech and Language” is easier to scan than an internal abbreviation or a laboratory’s program name. Keep historically precise terms in the articles, but use category names readers will recognize across decades of work. “Computer Vision” can hold early object-tracking and stereo-reconstruction posts without implying those methods began or ended in the period the blog covers.
Avoid near-duplicates such as Speech Technology and Speech Processing unless readers can tell how their scopes differ. “Research” and “History” are usually too broad for a blog where nearly every post concerns research history. The labels need to distinguish posts from one another, not describe the publication as a whole.
Look at each category page as a reader would. Does its title describe the posts listed? Do the first few entries show the range the label promises? A Biometrics category with one substantial article and several posts that mention face recognition in passing is not yet a useful browsing route. It can remain a candidate while those posts stay in their main subject areas.
Know when to split, merge, or leave a category alone
More posts do not automatically mean more categories. A large Speech and Language category can still work if its posts belong together and the page is easy to scan. Consider a split when recognition, synthesis, and language identification each have enough coverage for their own reading paths, new posts arrive regularly, and the broader page has become difficult to browse. The point is to improve a reader’s choices, not balance the category counts.
Merge categories when editors cannot explain the distinction or repeatedly disagree about where posts belong. If Document Processing and Document Analysis have become interchangeable, choose one label and reassign the affected posts. Before renaming a published category, check whether its page address will change. If so, redirect the old address and update any navigation pointing to it; otherwise, bookmarks and existing links may lead to a missing page.
Review categories after a substantial batch of new posts rather than after every publication. Look for empty categories, near-identical labels, pages filled with weakly related posts, and articles editors repeatedly struggle to place. Keep a brief record of scope decisions. A future editor should be able to see why an OCR article went into Document Analysis while a post on general image-feature extraction went into Computer Vision.
Give each category page a real purpose
A category page is part of the publication, not just a list generated by software. A short description can set its scope and help readers decide whether to browse it. Speech and Language, for example, might mention recognition, synthesis, and language identification in historical academic research. Do not promise complete coverage of a field when the archive offers selected case studies.
Order matters too. A newest-first list may bury a foundational explanation beneath later case studies. If the platform allows an introduction or featured entries, use them to orient readers. Otherwise, descriptive titles and consistent primary assignments carry more of the load. Category pages do not need lengthy repeats of article summaries; they need to help someone choose what to read next.
Before adding Low-Resource Technologies, list the posts it would contain and write its scope note. If that list holds just one speech-recognition article and an unrelated mention in a document-processing post, keep the term as a controlled tag for now. Revisit the category when several articles make limited data or resources central to their research questions.