While named entity recognition is a much addressed research topic, recognizing companies in text is of particular difficulty. Company names are extremely heterogeneous in structure, a given company can be referenced in many different ways, their names include person names, locations, acronyms, numbers, and other unusual tokens. Further, instead of using the official company name, quite different colloquial names are frequently used by the general public. We present a machine learning (CRF) system that reliably recognizes organizations in German texts. In particular, we construct and employ various dictionaries, regular expressions, text context, and other techniques to improve the results. In our experiments we achieved a precision of 91.11% and a recall of 78.82%, showing significant improvement over related work. Using our system we were able to extract 263,846 company mentions from a corpus of 141,970 newspaper articles. 1. FINDING COMPANIES IN TEXT Named entity recognition (NER) defines the task of not only recognizing named entities in unstructured texts but also classifying them according to a predefined set of entity types. The NER task was first defined during the MUC6 conference , where the objective was to discover general entity types, such as persons, locations, and organizations as well as time, currency, and percentage expressions in unstructured texts. Subsequent tasks, such as entity disambiguation, question answering, or relationship extraction (RE), rely heavily on the performance of NER systems, which perform as a preprocessing step. This section highlights the particular difficulties of finding company entities in (German) texts and introduces our industrial use-case, namely risk management based on companyrelationship graphs. 1.1 Recognizing company entities Although there is a large body of work on recognizing c ©2017, Copyright is with the authors. Published in Proc. 20th International Conference on Extending Database Technology (EDBT), March 21-24, 2017 Venice, Italy: ISBN 978-3-89318-073-8, on OpenProceedings.org. Distribution of this paper is permitted under the terms of the Creative Commons license CC-by-nc-nd 4.0 entities starting from persons and organizations, to entities like gene mentions or chemical compounds, the current research often neglects the detection of more fine-grained subcategories, such as person roles or commercial companies. In many cases, the “standard” entity classes turn out to be too coarse-grained to be useful in subsequent tasks, such as automatic enterprise valuation, identifying the sentiment towards a particular company, or discovering political and company networks from textual data. What makes recognizing company names particularly difficult is that in contrast to person names they are immensely heterogeneous in their structure. As such, they can be referenced in a multitude of ways and are often composed of many constituent parts, including person names, locations, and country names, industry sectors, acronyms, numbers, and other tokens, which makes them especially hard to recognize. This heterogeneity is expected to be true particularly for the range of medium-sized to small companies. Regarding examples like “Simon Kucher & Partner Strategy & Marketing Consultants GmbH”, “Loni GmbH”, or “Klaus Traeger”, which all are official names of German companies, one can easily see that they vary not only in length and types of their constituent parts but also in the position where specific name components appear. In the example “Clean-Star GmbH & Co Autowaschanlage Leipzig KG” the legal form “GmbH & Co KG” is interleaved with information about the type of the company (carwash) and location information (Leipzig, a city in Germany). What is more, company names are not required to contain specific constituent parts: the example “Klaus Traeger” from above is simply the name of a person. It does not provide any additional information apart from the name itself, which leads to ambiguous names that are difficult to identify in practice. Additionally, and in contrast to recognizing named entities from English texts, detecting them in German texts presents itself as an even greater challenge. As pointed out by Faruqui and Padó, this difficulty is due to the high morphological complexity of the German language, making tasks such as lemmatization much harder to solve . Hence, features that are highly effective for English often lose their predictive power for German. Capitalization is a prime example of such a feature. Compared to English, where capitalization of common nouns serves as a useful indicator for named entities, in German all nouns are capitalized, which drastically lowers the predictive power of the feature. We propose and evaluate a named entity recognizer for German company names by training a conditional random field (CRF) classifier . Besides using different features, Industrial and Applications Paper Series ISSN: 2367-2005 610 10.5441/002/edbt.2017.82 Figure 1: An example of a company graph. the fundamental idea is to include domain knowledge into the training phase of the CRF by using different real-world company dictionaries. Transforming the dictionaries into token tries enables us to determine efficiently whether the analyzed text contains companies that are included in the dictionary. During a preprocessing step, we use the token trie to mark all companies in the analyzed text that occur in the used trie. In addition, we automatically extend the dictionaries with carefully crafted variants of company names, as we expect them to occur in written text. 1.2 Use case: Risk management using company graphs Among the many possible applications for a companyfocused NER system, we focus on modern risk management in financial institutions as one that would benefit from such a system. Named entity recognition and subsequent relationship extraction from text for the purpose of risk management in financial institutions is particularly important in the context of illiquid risk . Illiquid financial risks basically represent contracts between two individuals, e.g., a bank granting a credit over 1 Mio USD (creditor) to a private company (obligor). Because the risk that the credit-taking company will not honor its repayment obligations cannot be easily transferred to other market participants, assessing the creditworthiness of an obligor is of major importance to the relatively small number of its creditors and other business partners. Also, insights gained by one bank on the obligor’s ability to pay back are usually not shared. Hence, obtaining adequate and timely information about non-exchange-listed obligors becomes a difficult task for creditors. To circumvent this difficulty, financial institutions rely on the so-called “insurance principle”: pooling a huge number of independent gains or losses ultimately results in the diversification of risk, which in turn eliminates almost all of it. Unfortunately, risk mitigation based on the insurance principle relies on the independence assumption between individual gains or losses. At the latest with the financial crisis of 2008/2009, this low dependency assumption has turned out to be devastatingly wrong. Information on the economic dependency structure between contracting parties and assets can be seen as the holy grail of financial risk management. Traditionally, the internal and external data sources used to assess credit risk focus on individual customers, not on the relationships between them. Dependency information is inferred from exposure to common risk factors and thus is inherently symmetric. Direct non-symmetric dependencies, such as supply chains, are not captured. Fortunately, with the growing amount of openly available data sources, there is justified hope that dependency modeling becomes significantly easier by leveraging this vast amount of data. Sadly, most of those data sources are textbased and require considerable effort to extract the contained knowledge about relationships and dependencies between the entities of interest. The desired outcome of such an extraction effort can be organized in a graph as shown in Figure 1. The figure shows an example of an actual company graph. To be able to automatically extract such graphs from large amounts of unstructured data, a reliable NER system constitutes the first decisive prerequisite for a following relation extraction step. As pointed out at the beginning, the described use case is merely one of many possible use cases, others might include semantic role labeling, machine translation, and question answering systems. 1.3 Contributions and structure We address the problem of recognizing company names from textual data by incorporating dictionary matches into the training process using a feature that represents whether a token is part of a known company name. Our evaluation focusses on analyzing the impact of using a perfect dictionary and different real-world dictionaries, as well as the effects of different ways to integrate the knowledge contained in the dictionaries on the performance of the NER system. In particular, we make the following contributions: • Creation of a NER system capable of successfully recognizing companies in German texts with a precision of 91.11% and a recall of 78.82%. • Analysis of the impact of various dictionary-based feature strategies on the performance of the NER. The remainder of this paper is organized as follows: Section 2 discusses related work, while Section 3 presents the baseline configuration for the CRF. In Section 4 we give an overview of the text corpus and the dictionaries we used. We describe the key data structures and technical aspects of the approach in Section 5. Finally, Section 6 presents our experimental results and Section 7 concludes the paper.