Detecting relevant segment of text in legal domain

研究的目标是调查、设计和实现算法，以检测(或识别)和提取相关文本片段，预测和识别法律实体和上下文，并最终生成适当的元数据存储在结构化数据库中。从用户查询到法律从业人员的法律研究，该数据库可用于多个场景。“相关语段”的概念被定义为与感兴趣的问题(或简称查询)有关的一段连续的文本。关联性可以用不同的方法来衡量，这取决于对关联性的解释。如果我们在法律文件中寻找法官的名字，我们可以使用广泛的信息提取(IE)工具。IE利用了广泛的技术，从图像分割到条件随机场(CRF)和马尔可夫模型，再到机器学习和分类，从图像分割(当文档的图像可用并且在特定区域中高度期望相关片段)。虽然可以使用IE技术提取实体等结构化信息片段，但对于法律文件中更深层次、含糊和概念性的部分，如损害类型或法官的决定和案件结果，我们需要开发一个超越IE技术的有监督的机器学习算法。这个问题既不是传统的IE问题，也不是文本分类问题。为了解决这个问题，法律文档被划分为概念上相关的部分，如标题、案例、引用、损害赔偿、判决等。这一步骤称为分区，可以使用监督或非监督学习方法来执行。一些区域，如页眉，预计会出现在文档的第一部分，因此它们可以通过无监督技术进行检测。另一方面，还有其他组成部分，如“损害”，可能出现在文档的任何部分，需要使用基于词典或手动标记的宏大真理或两者兼而有之的监督模型。

The goal of the research is to investigate, design, and implement algorithms to detect (or recognize) and extract the relevant segment of text, predict and recognize legal entities and context, and finally generate an appropriate metadata to be stored in a structured database. The database can be utilized in several scenarios from the user query to legal research by law practitioners. The notion of "relevant segment" is defined as a contiguous piece of a text which is relevant to the question of interest (or simply query). Relevance can be measured by different methods depending how relevance is being interpreted. If we are looking for the name of a judge in a legal document, we can use a wide range of information extraction (IE) tools. IE takes advantage of a broad spectrum of techniques from image segmentation, when the image of the document is available and a relevant segment is highly expected in a specific zone, to Conditional Random Fields (CRF) and Markov Models to Machine Learning and classification. While the structured pieces of information such as entities can be extracted using IE techniques, for deeper, ambiguous, and conceptual components of a legal document such as the type of damage or the judge's decision and case outcome, we need to develop a supervised machine learning algorithms beyond IE techniques. This problem is neither a traditional IE problem nor a text classification. To solve this problem, a legal document is partitioned into conceptually-related segments such as header, case, citations, damages, decision, and so on. This step is called zoning and can be performed using supervised or unsupervised learning methods. Some zones such as headers are expected to appear in the very first section of the document and so they can be detected by unsupervised techniques. On the other hand, there are other components such as "damages" which may appear in any part of the documents and needs a supervised model using either lexicon-based or manually-labeled grand truth or both.