Knowledge-enhanced information extraction across languages for pharmacovigilance

如今，科学知识可以在许多不同的形式和来源中以数字方式发布，例如百科全书，科学论文，以及结构化的知识来源，如本体或知识库。除此之外，新闻文章，博客文章或社交媒体也包含相关信息。所有这些都是以大量不同的语言发布的。例如，在MEDLINE中，每年都有近一百万篇新文章被收录。本项目旨在设计人工智能（AI）方法，自动消化这些不同类型的文本源，并联合提取这些知识和观察结果，以填充现有的知识库。我们的项目展示了药物警戒领域的这些方法，其致力于保持最新的药物不良反应（ADR）知识，以利于公众健康。在这一领域，权威来源包括科学期刊和药品标签，而基本观察结果则来自患者记录和社交媒体。当前主流的信息抽取方法都是从大型文本语料库中自监督抽取词汇表示，往往忽略目标领域的已有知识。相反，本项目旨在将现有知识整合到单词表示获取和信息提取过程中，以提高新信息和知识的提取。此外，它将利用以多种语文出版的类似信息，汇集各国的知识。需要跨越这些障碍阐述相关的调查结果，这需要时间和精力来收集和翻译成相应的语言。在不久的将来，工具将帮助研究人员和其他公民查找和链接分布在不同来源和语言的信息。在本项目中，我们将帮助改进此类技术，并将其用于药物警戒。拟议的三边合作明显有利于这种跨语文的工作。为了加强我们的合作和相互了解，我们计划在合作伙伴的共同监督下，为其他两个合作伙伴团队的早期职业研究人员提供实习机会，以及全体联合授课的培训行动，为他们提供共同的国际视野。该联盟由三个国际公认的自然语言处理专业团队组成。NAIST（JP）为日语创建了事实上的NLP工具。DFKI（DE）在语料库生成、通用信息抽取和生物医学文本处理方面有着很强的背景。LIMSI（FR）在语料库注释、混合信息提取和问答方面拥有丰富的经验，在生物医学语言处理方面拥有强大的背景，包括患者论坛的药物警戒。

Nowadays scientific knowledge can be published digitally within many different forms and sources, such as encyclopedias, scientific papers, but also structured knowledge sources like ontologies or knowledge bases. Beside that also news articles, blog posts or social media contain relevant information. All this is published everyday in a large number of different languages.In MEDLINE for instance every year close to one million new articles are included.The present project aims to design Artificial Intelligence (AI) methods that automatically digest these different types of text sources and jointly extract such knowledge and observations in order to populate existing knowledge bases.Our project showcases these methods in the domain of pharmacovigilance, which endeavors to maintain up-to-date knowledge on adverse drug reactions (ADR) for the benefit of public health. In this domain, authoritative sources include scientific journals and drug labels while elementary observations are reported in patient records and social media.Current mainstream information extraction methods use self-supervised extraction of word representations from large text corpora and tend to neglect existing knowledge on the target domain. In contrast, the present project aims to integrate existing knowledge into the word representation acquisition and information extraction processes to improve the extraction of new information and knowledge. Additionally, it will take advantage of the existence of similar information published in multiple languages to pool knowledge across countries.Language barriers hamper the free flow of knowledge and thought across languages. Relevant findings need to be articulated across these barriers, which requires time and effort to collect and translate into the respective languages. In the not too distant future, tools will assist researchers and other citizens in finding and linking information distributed across sources and languages. In this project, we will help to improve such technologies and will demonstrate them for pharmacovigilance. This cross-language dimension obtains a clear benefit from the proposed trilateral collaboration. To strengthen our collaboration and mutual knowledge, we have planned internships for early career researchers at each of the other two partner teams under the joint supervision of the partners, as well as plenary, jointly taught training actions, to provide them with a shared international exposure.The consortium is composed of three internationally recognized teams specialized in natural language processing. NAIST (JP) has created the de-facto NLP tools for Japanese. DFKI (DE) has a strong background in corpus generation, general information extraction and biomedical text processing. LIMSI (FR) has a long experience in corpus annotation, hybrid information extraction and question-answering, and a strong background in biomedical language processing, including pharmacovigilance from patient forums.