SBIR Phase I: Building a Flexible, Technology Adaptive Architecture to Support Processing of Content by Knowledge Workers

这个小企业创新研究（SBIR）第一阶段项目解决了当今自然语言处理系统的能力与公共关系、外交事务和危机管理等语言敏感领域的知识工作者（分析师和研究人员）的准确性要求之间的差距。许多组织中的知识工作者监控和分析印刷品和网络报道，以获取感兴趣的内容。当搜索结果量很大时，一些人使用自然语言处理（NLP）系统使用复杂的词库、模式和特定于上下文的算法对结果进行过滤、分类和评分。然而，用户抱怨这些系统没有达到预期的准确性，缺少反讽、讽刺、隐喻、双关语和对参考文献的不正确解释等修辞手段。因此，对准确性要求较高的用户转而使用手动流程来补充或替代技术。该项目将测试一个原型架构，允许快速插入、测试和文本分析算法的适应，以及有效集成人类审查判断的工作流过程。一旦商业化，该系统将使知识工作者更快地采用技术。在精度要求较高的领域，对人工判断的需求限制了技术在搜索等离散领域的使用，而在随后的处理步骤中，分析人员必须手动捕获、分类、评分、分析和报告输出。

This Small Business Innovation Research (SBIR) Phase I Project addresses the gap between the capabilities of today's natural language processing systems and the accuracy requirements of knowledge workers (analysts and researchers) in language-sensitive fields such as public relations, foreign affairs, and crisis management. Knowledge workers in many organizations monitor and analyze print and web coverage for content of interest. When the volume of search results is large, some filter, classify and score the results with Natural Language Processing (NLP) systems using complex libraries of words, patterns, and context-specific algorithms. However, users complain that these systems fall short of desired accuracy, missing rhetorical devices such as irony, sarcasm, metaphors, double entendre, and improperly interpreting references. Users with high thresholds for accuracy thus turn to manual processes to either supplement or substitute for technology. This project will test a prototype architecture allowing rapid insertion, testing, and adaptation of text analysis algorithms and a workflow process efficiently integrating human review judgment. Once commercialized, the system will enable more rapid adoption of technology by knowledge workers. In fields with high accuracy requirements, the need for human judgment has constrained technology use to discrete areas like search, while in subsequent processing steps, analysts must manually capture, classify, score, analyze, and report on the output.