| { |
| "field_id": "F041", |
| "field": "computational linguistics", |
| "directions": [ |
| { |
| "direction_id": "G0", |
| "name": "Machine translation", |
| "scope": "Produce better translations between natural languages or better translation systems as data availability, domain, and language coverage vary." |
| }, |
| { |
| "direction_id": "G1", |
| "name": "Conditioned text production", |
| "scope": "Generate text that meets task goals based on supplied material, events, plans, or content and form requirements. This includes summaries that compress source information and other controlled or grounded prose generation." |
| }, |
| { |
| "direction_id": "G2", |
| "name": "Correction and restoration of existing text", |
| "scope": "Detect and repair grammatical, spelling, or transcription-form errors in existing text to produce more appropriate written text." |
| }, |
| { |
| "direction_id": "G3", |
| "name": "Evaluation of generated language", |
| "scope": "Develop or validate judgments of the quality, correctness, faithfulness, and usefulness of already generated text." |
| }, |
| { |
| "direction_id": "G4", |
| "name": "Evidence-based answering and multi-step inference", |
| "scope": "Use supplied evidence or organize intermediate reasoning to derive answers to questions and other language-based inferences. Evidence retrieval and reasoning-step design are parallel branches, not requirements that every study must meet together." |
| }, |
| { |
| "direction_id": "G5", |
| "name": "Structured meaning from language", |
| "scope": "Produce verifiable structured representations from language input, either by mapping requests to formal queries and programs or by recording entities, relations, and events expressed in text." |
| }, |
| { |
| "direction_id": "G6", |
| "name": "Affective and harmful-language interpretation", |
| "scope": "Identify emotions, opinions, or harmful social meanings directed at people that are expressed in language. Affective judgments and harm judgments retain distinct targets; neither is a prerequisite for the other." |
| }, |
| { |
| "direction_id": "G7", |
| "name": "Social bias and fairness in language systems", |
| "scope": "Measure, explain, or reduce social-group-related stereotypical associations and unequal outcomes in language models and NLP systems." |
| }, |
| { |
| "direction_id": "G8", |
| "name": "Reusable language models under capability and resource constraints", |
| "scope": "Build language-model capabilities reusable across multiple tasks and study how model size, inference cost, or processable context length limits those capabilities. Specialist-domain adaptation and resource optimization are parallel branches." |
| }, |
| { |
| "direction_id": "G9", |
| "name": "Task learning with limited supervision and changing distributions", |
| "scope": "When direct labels for a target task are insufficient or languages, categories, or deployment distributions change, use cross-lingual knowledge, create training examples, or apply transferable classification mechanisms to maintain downstream prediction capability. The three approaches need not occur together." |
| }, |
| { |
| "direction_id": "G10", |
| "name": "Probing linguistic and reasoning capabilities", |
| "scope": "Use controlled comparisons, probes, and challenge tasks to diagnose what linguistic structures, meanings, and inference capabilities existing models or representations capture." |
| }, |
| { |
| "direction_id": "G11", |
| "name": "Conversational understanding and response", |
| "scope": "Use multi-turn context to sustain useful interaction by understanding utterances and persistent state, or by generating or planning an appropriate next response from that context. Each study need not contribute to both." |
| } |
| ] |
| } |
|
|
