{ "title": "Building Human-Language Technology", "type": "SIDTODAY", "description": "Human Language Technology relies on corpora--collections of linguistic data. For the NSA's linguistic tools to analyze SIGINT accurately, their datasets need to include analysis of specific terms from classified materials, meaning that the corpora themselves must be classified and the analysts who work on them must have security clearance.", "long_description": "This SIDtoday newsletter article, part of the Human Language Technology (HLT) series, describes the Corpora Activity within NSA's HLT Program Management Office. It explains that linguistic corpora—annotated collections of language data—are the foundation for developing and evaluating HLT tools used by analysts. The article contrasts open-source, unclassified corpora used in the commercial and academic worlds with the classified corpora NSA must build from SIGINT intercept, noting that such classified data requires cleared, qualified linguist annotators and adherence to security and policy restrictions. It also promotes diversity tours and details available to language analysts to support this work.", "tags": [ "Human Language Technology", "corpora", "linguistics", "language analysis", "media mining", "annotation", "tool development" ], "document": { "creator": "NSA", "date_created": "2006-09-07", "classification": [ "SECRET", "SI" ], "dissemination": [], "banner": "DYNAMIC PAGE -- HIGHEST POSSIBLE CLASSIFICATION IS TOP SECRET // SI / TK // REL TO USA AUS CAN GBR NZL", "genre": "newsletter", "series": { "name": "(U) HLT", "part": 7 }, "programs": [ "HLT" ], "countries": [], "people": [] }, "publication": [ { "publisher": "The Intercept", "date_published": "2019-05-29", "link": "https://theintercept.com/snowden-sidtoday/5987334-building-human-language-technology/", "doc_source": "https://theintercept.com/snowden-sidtoday/5987334-building-human-language-technology/", "doc_url": "https://s3.documentcloud.org/documents/5987334/Building-Human-Language-Technology.pdf" } ], "pdf_sha256": "2fc625536d6bc9f1f88b1b66ccdddc8d41a72241c60f150cd31012620a934bd6", "pages": 2 }