In this work a natural language processing workflow to extract sequential activities from large collections of medical text documents is developed. The approach utilizes graph structures to process, link and assess activities found in the documents. * Contact: aniekler@informatik.uni-leipzig.de relation extraction, natural language processing, graph processing, process models
Medical publications, surgical procedure reports or medical records typically contain procedural descriptions. For example, all activities included in a medical study must be documented for reproducibility purposes, in surgical reports a stepwise description of included procedures is documented and in medical records a history of medical treatment is listed. Additionally, related studies or reports describe alike activities with some alterations or rely on preceding activities that may be described in other documents. Consider for example the preparation steps before DNA could be sequenced which are described in scientific papers. They are often the same but need to be documented for each study. Such redundant activity descriptions can be found amongst many documents describing research within the same domain or field of research. Nevertheless, differences amongst the activities in related documents also exist. A complete overview of activities from a defined document collection provides an easy insight to workflows and paradigms within a domain or field of study.Thus, finding this link between the documents and aligning the activities w.r.t. redundant activities helps to structure and analyze procedural knowledge from topical- or domain-related medical texts. For example, early stages or parts of a larger process might be documented separately to other parts or later stages.
In this work a general natural language processing workflow to extract sequential activities from large collections of medical text documents is developed. A graph-based data structure is introduced to merge extracted sequences which contain similar activities in order to build a global graph on procedures which are described in documents on similar topics or tasks. In this section we describe a methodology which extracts and links activities from medical text documents. The described system follows a sequence of procedures in order to create an activity graph as a result. First, the text sources have to be processed in order to access the entity items in the text. Different entities in a sentence are related and form an expressed activity. Therefore, the extraction of valid relations that form activities is introduced to the text processing step. The second step in our proposed methodology is the creation of a directed graph structure which can be further used for the representation of the activities contained within a text collection. 2.1
The text sources must be separated into sentences and tokens first by using state of the art tools. Additionally, POS-Tagging must be applied to the text sources. To extract the procedural knowledge from the texts, named entity recognition (NER) is required as a pre-processing step. In the separated and preprocessed sentences multiple entities may form an activity. Consider the sentence “Real-time JJ PCR NNP was VBD done VBN using VBG the DT fluorescent-labelled JJ oligonucleotide NN probes NNS”. “Realtime PCR” and “fluorescent-labelled oligonucleotide probes” are the identified entities which form the activity “done”. This activity can be part of a chain of activities throughout multiple documents.
The characteristics of activities or relations between entities change within different domains or described procedures. Thus, the process for identifying and connecting entities to activities within the sentences should not be fixed or static. To answer this fact the identification of relations or activities is defined as classification task using a Support Vector Machine (SVM) along with word- and POSTag-level features GuoDong et al. [2005]. If a sentence contains an entity E1 and E2 the two words before E1, the two words after E2 and all words between E1 and E2 are extracted as features. Furthermore, the POS-tags of the extracted words are used as features for the SVM.
Before the training process is applied the user must define the type and the form of the desired relation. On the basis of this definition training examples are collected from the data. For this purpose an active learning procedure is introduced where the user iteratively collects training data with the support of an automatic classification. An initial search for sentences that include a minimum of entities and verbs that indicate an activity is conducted. The set of matching sentences which contain this custom pattern is presented to the user. Correct entities are selected from the proposed sentences along with the definition whether there is a relation between them or not. The features are extracted automatically and the set of positive and negative examples is used to train an initial SVM model. The trained model is used to identify additional examples in the data. The user judges on those examples and with every batch of new examples the classifier can be refined. If the training quality of the SVM does not change with new examples a final model is trained and applied to all documents. The result is a set of sentences from a document collection where each sentence contains an activity or valid relation between entities.
can be said, A0 is an unconnected graph where a set N of connected components can be identified. This set represents different graphs where the interaction and coherence of related processes, described in different documents, is encoded. In order to quantitatively judge on the quality of the extraction process an evaluation dataset and evaluation strategy needs to be developed as prequisite for future work. More research on suitable similarity functions for relations which can also handle semantic similarities will optimize the quality of the graph merging process. Future work will also include the adoption of domain knowledge from knowledge graphs. Those has been described as very helpful resources in order to adopt to a domain in Roberts and Harabagiu [2011]. The links and dependencies between entities and their possible representations in the data can be encoded in those data structures by domain experts. This will add supervision and control to the graph creation process and thus allows for a higher precision of the graph. Additionally, anaphora resolution can be modeled with knowledge bases to connect graph structures where the relations represent processes which produce other entities as results. Such edges can’t be established with character or semantic comparison of the relations. In the moment a connection can only be established if the producing process is encoded within a single document. Furthermore, the introduction of manual corrections steps to refine the graph and the optimization of the relation extraction classification may also optimize the quality.
ExB Labs GmbH kindly helped to compile and preprocess the corpus for this work.
Funding: The project is funded by European Regional Development Fund (ERDF/EFRE) and the European Social Fund (ESF).