SBIR Phase I: Representation and Deep Learning for Free Text Applications

这个小型企业创新研究(SBIR)第一阶段项目寻求验证一种处理文本材料的新方法，以便计算机能够更好地学习与自然语言相关的应用程序。一个目标是使计算机学习方法能够更好地考虑句子中的句法结构，这在目前是困难的。将建造两个实验原型。其中一个重点是自动确定关于公司、品牌或机构的在线帖子是正面的还是负面的(情绪分析)。这很困难，因为句子结构很重要，还因为在线帖子可能是非正式的，包含俚语等。第二个原型在搜索过程中找到最重要的段落，也考虑了句子结构。这种段落级别的信息检索有助于从运行的文本或网页中提取事实信息。新方法使用单个高维向量(例如，1,000个数字的列表)同时表示具有解析结构的单词。一个成功的SBIR项目将最终改进机器学习应用程序，用于广泛的任务，包括文档检索、摘要和自动翻译。此外，同样的技术可以应用于表示机器学习之前的任何结构化集合，包括图像和基因组信息。该项目更广泛的影响/商业潜力是加强自由文本和其他结构化数据的自动处理能力。这种方法的动机来自神经网络，反过来，它也应用于神经建模和我们对大脑如何处理信息的理解。在软件行业，商业创新继续围绕网页的自动化处理，这在创建许多新公司方面发挥着关键作用。因此，处理自由文本的能力变得越来越重要。一种更好地表示用于机器学习的文本的方法将在必须考虑句子结构的任何地方打开新的功能。这将导致新的创业，并为消费者提供新的产品和服务。一个成功的项目将验证新技术，这些技术可以为利用新技术的公司带来巨大的竞争优势，为消费者提供新的更好的能力，并促进那些经济受益于新技术创新的国家，如我们自己的国家。

This Small Business Innovation Research (SBIR) Phase I project seeks to validate a new way to process textual material, so that computers can better learn applications related to natural language. A goal is to enable computer learning methods to take better account of parse structure in sentences, which is currently difficult. Two experimental prototypes will be constructed. One will focus upon automatically determining whether online posts about a company, brand or institution are positive or negative (sentiment analysis). This is difficult because sentence structure is important, and also because online posts can be informal, contain slang, etc. The second prototype finds the most important paragraph during a search, also taking sentence structure into account. Such paragraph-level information retrieval helps when extracting factual information from running text or web pages. The new method represents words simultaneously with parse structure using a single high-dimensional vector (for example, a list of 1,000 numbers). A successful SBIR project will ultimately improve machine learning applications for a wide range of tasks, including document retrieval, summarization, and automated translation. Moreover, the same techniques can be applied to represent any structured collection prior to machine learning, including images and genomic information. The broader impact/commercial potential of this project is to enhance capabilities for automated processing of free text and other structured data. Motivation for this approach comes from neural networks and, in turn, it has applications to neural modeling and our understanding of how the brain processes information. In the software industry, commercial innovation continues to revolve around automated processing of web pages, which plays a key role in creating many new companies. Therefore, the ability to handle free text is increasing in importance. A better way to represent text for use with machine learning will open new capabilities wherever the structure of sentences must be taken into account. This will lead to new startups, and provide consumers with new products and services. A successful project will validate new technology that can give a huge competitive edge to companies that take advantage of it, provide new and better capabilities for consumers, and advance those countries, such as our own, whose economies benefit from new technological innovations.