Detecting Chemical Reactions in Patents

Detecting Chemical Reactions in Patents Hiyori Yoshikawa1,3, Dat Quoc Nguyen1, Zenan Zhai1, Christian Druckenbrodt2, Camilo Thorne2, Saber A. Akhondi2, Timothy Baldwin1, Karin Verspoor1 1The University of Melbourne, Australia; 2Elsevier; 3Fujitsu Laboratories Ltd. 1{hiyori.yoshikawa,dqnguyen,zenan.zhai,tbaldwin,karin.verspoor}@unimelb.edu.au 2{c.druckenbrodt,c.thorne.1,s.akhondi}@elsevier.com Abstract Extracting chemical reactions from patents is a crucial task for chemists working on chem- ical exploration. In this paper we introduce the novel task of detecting the textual spans that describe or refer to chemical reactions within patents. We formulate this task as a paragraph-level sequence tagging problem, where the system is required to return a se- quence of paragraphs that contain a descrip- tion of a reaction. To address this new task, we construct an annotated dataset from an ex- isting proprietary database of chemical reac- tions manually extracted from patents. We in- troduce several baseline methods for the task and evaluate them over our dataset. Through error analysis, we discuss what makes the task complex and challenging, and suggest possible directions for future research. 1 Introduction Chemical patents are a crucial resource for chem- ical research and development activities. In fact, many compounds are reported first in patents and only a small fraction of them appears in the chem- ical literature after 1 to 3 years (Senger et al., 2015), meaning that chemists habitually search over both academic papers and patent databases. Moreover, as the number of chemical patents awarded each year is ever-increasing, there is an increasing urgency to perform patent searches to establish the novelty of chemical compounds (Akhondi et al., 2014). Text mining methods are a vital tool in this process, enabling a significant reduction in associated time and effort. Most previous research in text mining of chem- ical information has focused on named entity recognition (NER) of chemical concepts, and sev- eral publicly-available NER corpora have been de- rived from both scientific literature (Kim et al., 2003; Corbett et al., 2007; Krallinger et al., 2015) and chemical patents (Akhondi et al., 2014). Some studies have also addressed relation extraction be- tween chemical entities and other concepts such as protein and diseases (Wei et al., 2015; Krallinger et al., 2017). However, there has been limited work on au- tomatically extracting chemical reactions from patents. A chemical patent usually contains a de- scription of chemical reactions that are relevant to its claims. Figure 1 shows an example of a chemical reaction description. Generally speak- ing, a chemical reaction is a process where a set of chemical compounds is transformed into another set of chemical compounds. A reaction descrip- tion may include the source chemical compounds, solvents and reagents involved in the reaction, re- action conditions, and materials obtained as a re- sult of the reaction. Despite the fact that such in- formation is crucial for a comprehensive under- standing of chemical patents, there are — to the best of our knowledge — few methods or anno- tated resources that can be used for this purpose. As a first step to extracting chemical reactions, a filtering step must take place to determine …

Content truncated to 500 words