News MRN - Entity Linking to listed companies in South Korea

Why Aicel MRN?Enable users to exploit signals in news for quantitative strategies and systematic tradingAnalyze across thousands of companies in real timeMitigate risk with an early warning signalData mapped to Bloomberg stock symbols (tickers)Automated entity linking with accuracy: Aicel proprietary NLP Technology OverviewNamed Entity RecognitionOne of the key requirements for any NLP technology is identifying what the different parts of any given texts represent. NER plays an important role in determining the parts of text to focus on within a news article. Aicel’s NER model classifies texts into ten types of entities (as opposed to the usual four – person, location, organization, other) and is trained from the ground-up using text data tailored specifically from the financial domain. Knowledge GraphsKnowledge Graph augments value extracted from unstructured data by storing and modeling the relationships between different entities. Capturing these relationships allows our system to infer various information from texts even if they are not directly mentioned in the data. Aicel’s knowledge graph has been built using our own proprietary data in the financial domain and has been expanded via various linked open data such as Wikipedia. Named Entity LinkingIdentifying individual entities from documents is necessary to extract value and capture signals from unstructured texts for humans and machines alike. Just as humans identify what a word represents using both its textual form and any contextual information surrounding it, Aicel combines both semantic meaning extracted from deep learning model and contextual meaning from Knowledge Graph when identifying entities for any given mentions in texts. News ClusteringCertain stories may appear in multiple different articles as we source news from multiple sources with coverage over a broad domain. Our clustering algorithm allows for a filtering of these redundancies to quickly identify unique events, as well as utilizing the redundancies when identifying significant events to serve high-quality signals across unstructured news data. FAQHow do you handle corporate events?IPOs: News data for the new ticker is delivered from its first trading day.Company spin-offs: If company B spins-off from company A and starts trading in the stock exchange, news data for company B will become available from the first trading day of company B.M&A (or acquisition of a brand): If a listed company A acquires a private company B, news data for company B will start to be mapped to the ticker of company A. If company A acquires brand B, news for brand B will be mapped to company A. The ticker mapping process is automated by Aicel’s knowledge graph, and Aicel’s research team also manually manages and regularly update mapping information for [Brand] : [Ticker] and [Subsidiary Company] : [Parent Company]. How often do you train your Knowledge Graph?Changes regarding listed companies are tracked and updated daily along with the stock market. Non-listed companies and other entities (such as brands, products, etc) do not have a fixed update cycle and are updated whenever a change is planned or detected. Models that require knowledge graph for training (such as for knowledge graph embeddings used in NEL) are trained once a week at the latest. What is the methodology of news clustering?There are often many different news articles written on the same topic due to the number of sources we collect the articles from. We provide a way to identify and group these articles via clustering. Aicel’s news clustering methodology incorporates many features in addition to the content itself; such as vectorized words to handle synonyms, and mapped tickers to discern articles with similar content on different companies. These features are used to calculate unique vectors for each article, which are then projected to a vector space where a variable number of high-density clusters can be identified. How is fact relevance (saliency) measured?Fact relevancy measures how strongly related the given news article is to a mapped ticker (and its entities). The relevancy score starts off with the positioning of the mentioned entity as the baseline. A high relevancy score will be assigned to an entity in the headline and lower scores to entities mentioned further down the article. The relevancy score may then be adjusted if the entity is identified to play a significantly higher or lower role than the score calculated by its positioning.

为何选择 Aicel MRN？

启用用户利用新闻中的信号进行量化策略和系统化交易。
实时分析数千家公司。
通过早期预警信号降低风险。
数据映射至彭博股票符号（交易代码）。
自动化实体链接，精确度高达 Aicel 独家自然语言处理技术。

技术概览

命名实体识别
任何自然语言处理技术的关键要求之一是识别给定文本中不同部分所代表的内容。命名实体识别在确定新闻文章中应关注的文本部分方面发挥着重要作用。Aicel 的命名实体识别模型将文本分为十种实体类型（相较于常见的四种——人物、地点、组织、其他），并且从零开始，使用专门针对金融领域的定制文本数据进行训练。

知识图谱
知识图谱通过存储和建模不同实体之间的关系，增强了从非结构化数据中提取的价值。捕捉这些关系允许我们的系统从文本中推断各种信息，即使这些信息没有直接在数据中提及。Aicel 的知识图谱是使用我们自己的金融领域专有数据构建的，并通过诸如维基百科等各种链接的开放数据进行了扩展。

命名实体链接
从文档中识别个体实体对于从非结构化文本中提取价值并捕捉信号，无论是对于人类还是机器都是必要的。正如人类通过单词的文本形式及其周围的任何上下文信息来识别一个词所代表的内容一样，Aicel 在识别文本中的任何提及的实体时，结合了从深度学习模型中提取的语义意义和知识图谱中的上下文意义。

新闻聚类
某些故事可能出现在多个不同的文章中，因为我们从多个来源收集文章，涵盖广泛的领域。我们的聚类算法允许过滤这些冗余，快速识别独特事件，同时在识别重大事件时利用冗余，为非结构化新闻数据提供高质量的信号。

常见问题解答

如何处理公司事件？
IPO：新交易代码的新闻数据从其首个交易日提供。
公司分拆：如果公司 B 从公司 A 分拆并开始在证券交易所交易，公司 B 的新闻数据将从公司 B 的首个交易日提供。
合并与收购（或品牌收购）：如果上市公司 A 收购私人公司 B，公司 B 的新闻数据将开始映射到公司 A 的交易代码。如果公司 A 收购品牌 B，品牌 B 的新闻将映射到公司 A。Aicel 的知识图谱自动化了交易代码映射过程，并且 Aicel 的研究团队也手动管理和定期更新 [品牌] : [交易代码] 和 [子公司] : [母公司] 的映射信息。
知识图谱多久训练一次？
关于上市公司的变化每天都会跟踪和更新，与股市同步。非上市公司和其他实体（如品牌、产品等）没有固定的更新周期，在计划或检测到变化时进行更新。需要知识图谱进行训练的模型（如用于命名实体链接中的知识图谱嵌入）每周至少训练一次。
新闻聚类的原理是什么？
由于我们从收集文章的多个来源中经常看到关于同一主题的许多不同的新闻文章，我们提供了一种通过聚类识别和分组这些文章的方法。Aicel 的新闻聚类方法除了内容本身外，还结合了许多其他功能；例如，使用向量化单词处理同义词，以及映射交易代码以区分关于不同公司的内容相似的文章。这些功能用于为每篇文章计算独特的向量，然后将其投影到向量空间中，在那里可以识别出可变数量的高密度聚类。
如何衡量事实的相关性（显著性）？
事实相关性衡量给定新闻文章与映射交易代码（及其实体）的相关强度。相关性分数以提及实体的定位为基准。对于标题中提到的实体将分配高相关性分数，而对于文章中提到的较后的实体分配较低分数。如果发现实体扮演比计算出的分数显著更高或更低的角色，相关性分数可能会进行调整。