Tag: nlp

The trigger vocabulary that finds surrender statements was grown and tested: 373 mined candidates, 45 tested against ten labeled sources spanning four tiers of label quality, 9 promoted on human gold. Three measured regularities came out: precision rises with phrase length (0.603 at two words to 0.771 at five, over 134,600 decided spans); first-person forms separate self-acknowledged limitations from reviewer criticism (0.955 versus 0.000 for the same phrase family); and mining scores do not predict trigger quality. Every number is a measurement with its label tier stated.

Three models one size class up — Qwen3-14B, Phi-4, and Mistral-Small-24B at Q4_K_M quantization — measured on the same 1,505-sentence limitation gold set, same prompt, same harness as the earlier six-model comparison. Qwen3-14B scores F1 0.8272 as a single model, above the previous two-model union (0.8143) and the published rule-system reference (0.800). VRAM at load: 10.6 GB of 16.3 GB. Every number is a full-split measurement.

Scientific papers state their own defeats in words: a sample that could not be enrolled, a computation that could not be afforded, data that could not be obtained. A phrase list finds those statements with precision 0.23. Judging each match in context — with rules, two trained classifiers, and a pair of 8-billion-parameter models on two 16GB GPUs — raises the measured operating point to a 0.945-precision auto-admit gate and a two-judge F1 of 0.814, beating the published rule-based reference on the same gold set. Every number is a full-split measurement.

Terms of service are revised for two decades, and clauses come and go. Scoring every archived revision of YouTube, Reddit, and Twitter against the same 34 consumer issues — and measuring how much each revision changed — surfaces what a single snapshot hides: rewrites that leave the length untouched, and clauses removed for years that quietly return. YouTube's terms drop machine-learning language in the 2010s and bring it back, spelled out, in July 2022 — a revision that replaced 92% of the document's wording with a net length change of 21 words.

A word count tells you a terms-of-service document got longer. It does not tell you what the document covers. This is a different measurement: a fixed list of 34 consumer issues, checked against real terms-of-service and privacy documents, with every clause scored three ways — how closely it matches the issue, how much text it spends on it, and how harshly it reads. It starts with three documents read closely and ends with an interactive leaderboard of sixty-four, ranked by what they actually address.

In 2016 the Defense Technical Information Center required 1,400 new thesaurus terms to be categorized against its 272 Field & Group categories. A WordNet-based run performed 380,800 semantic comparisons in about 4.3 hours of machine time, replacing an estimated one to two man-years of manual review with a ranked shortlist for human confirmation.

Daydreamer generates associations without being prompted. Given a lexical model of English — WordNet, VerbNet, SentiWordNet, and the NRC Emotion Lexicon — it draws links between ideas, uses the oppositions among them as a signal, and uncovers the concepts that unify them: concepts its input never named.

The complete DOJ Epstein release is 1,378,652 documents. Instead of reading headlines about them, we indexed, embedded, and machine-read the entire corpus — full-text search, semantic vectors, and computer vision — measuring 46 topics, 48 named individuals, and every image, to establish what is actually in the files and what is not.