Exploiting machine learning to improve the proposal submission process

为期3个月的第一阶段提案描述了创建一个原型系统，以评估提交给Innovate UK的提案。该项目有三个主要目标：将建议书与评估员配对，查明重复和重新提交的建议书，然后查明欺诈行为。该系统将使用自然语言处理（NLP），机器学习和数据分析的组合来理解提案的内容，并将其分配给相关的评估人员。该系统使用主题建模来识别提案的关键概念，而NLP则提取人、组织和关键字等实体，以构建提案内容的丰富图片。然后将提取的数据存储到图形数据库和JSON-LD中。然后采取类似的方法从评估员的个人资料、简历、撰写的文章、博客文章、推特或期刊论文中提取内容，以创建评估员知识和技能的丰富图片。从提案和评估员材料中提取的内容，主题，关键字和实体，然后以结构化的方式存储在Elasticsearch和图形数据库（Neo4j）中的JSON-LD中。搜索查询，社交网络分析和文本向量机学习聚类技术，然后使用链接的基础上，他们的技能和提案主题的相似性的提案评估。该提案的第二个目的是利用称为文件指纹和筛选的技术，查明重新提交的提案、重复的内容和改写的提案。这些技术将文档分解成块，然后可以在提交的提案的语料库中识别这些块。然后使用神经网络机器学习方法对这些结果进行聚类。最终目的是通过评估参与提案的人员和组织，针对公开收集的信息，确定可能的欺诈行为。我们建议使用NLP来提取人员，组织和网站实体，然后将其与来自Companies House和Gateway to Research API的数据进行交叉引用。同样，我们将使用具有社交网络分析和机器学习技术的图形数据库来识别欺诈性提案，例如以公司子公司名称提交的多个提案。

The 3-month Phase 1 proposal describes the creation of a prototype system to assess proposals that are submitted to Innovate UK. The project has three main goals: matching proposals to assessors, identifying duplicates and resubmitted proposals and then identifying fraudulent behaviour. The system will use a combination of Natural Language Processing (NLP), machine learning and data analytics to understand the content of a proposal and the assign this to a relevant assessor. The system uses topic modelling to identify the key concepts of a proposal, while NLP extracts entities such as people, organisations and keywords to build a rich picture of the proposal content. The extracted data is then stored to a graph database and in JSON-LD. A similar approach is then taken to extract content from assessor’s profiles, resumes, authored articles, blog posts, tweets or journal papers to create a rich picture of an assessor’s knowledge and skills. The extracted content, topics, keywords and entities from the proposals and assessor’s material, is then stored in a structured manner in JSON-LD in Elasticsearch and a graph database (Neo4j). Search queries, social network analysis and text vector machine learning clustering techniques are then used to link assessors to proposals based upon the similarities of their skills and the proposal topics. The second aim of the proposal is to identify resubmissions, duplicate content and reworded proposals using techniques called document fingerprinting and Winnowing. These techniques break a document into chunks that can then be identified across a corpus of submitted proposals. These results are then clustered using a Neural Network machine learning approach. The final aim is to identify possible fraudulent behaviour, by assessing the people and organisations involved in the proposal against publicly gathered information gathered information. We propose to use NLP to extract the people, organisation and websites entities and then cross-reference this against data from Companies House and Gateway to Research APIs. Again, we will use a graph database with social network analysis and machine learning techniques to identify fraudulent proposals, for example multiple proposals that have been submitted under company subsidiary names.