Coreference annotation and resolution in the Colorado Richly Annotated Full Text (CRAFT) corpus of biomedical journal articles.

BACKGROUND: Coreference resolution is the task of finding strings in text that have the same referent as other strings. Failures of coreference resolution are a common cause of false negatives in information extraction from the scientific literature. In order to better understand the nature of the phenomenon of coreference in biomedical publications and to increase performance on the task, we annotated the Colorado Richly Annotated Full Text (CRAFT) corpus with coreference relations. RESULTS: The corpus was manually annotated with coreference relations, including identity and appositives for all coreferring base noun phrases. The OntoNotes annotation guidelines, with minor adaptations, were used. Interannotator agreement ranges from 0.480 (entity-based CEAF) to 0.858 (Class-B3), depending on the metric that is used to assess it. The resulting corpus adds nearly 30,000 annotations to the previous release of the CRAFT corpus. Differences from related projects include a much broader definition of markables, connection to extensive annotation of several domain-relevant semantic classes, and connection to complete syntactic annotation. Tool performance was benchmarked on the data. A publicly available out-of-the-box, general-domain coreference resolution system achieved an F-measure of 0.14 (B3), while a simple domain-adapted rule-based system achieved an F-measure of 0.42. An ensemble of the two reached F of 0.46. Following the IDENTITY chains in the data would add 106,263 additional named entities in the full 97-paper corpus, for an increase of 76% percent in the semantic classes of the eight ontologies that have been annotated in earlier versions of the CRAFT corpus. CONCLUSIONS: The project produced a large data set for further investigation of coreference and coreference resolution in the scientific literature. The work raised issues in the phenomenon of reference in this domain and genre, and the paper proposes that many mentions that would be considered generic in the general domain are not generic in the biomedical domain due to their referents to specific classes in domain-specific ontologies. The comparison of the performance of a publicly available and well-understood coreference resolution system with a domain-adapted system produced results that are consistent with the notion that the requirements for successful coreference resolution in this genre are quite different from those of the general domain, and also suggest that the baseline performance difference is quite large.

A corpus of full-text journal articles is a robust evaluation tool for revealing differences in performance of biomedical natural language processing tools.

BACKGROUND: We introduce the linguistic annotation of a corpus of 97 full-text biomedical publications, known as the Colorado Richly Annotated Full Text (CRAFT) corpus. We further assess the performance of existing tools for performing sentence splitting, tokenization, syntactic parsing, and named entity recognition on this corpus. RESULTS: Many biomedical natural language processing systems demonstrated large differences between their previously published results and their performance on the CRAFT corpus when tested with the publicly available models or rule sets. Trainable systems differed widely with respect to their ability to build high-performing models based on this data. CONCLUSIONS: The finding that some systems were able to train high-performing models based on this corpus is additional evidence, beyond high inter-annotator agreement, that the quality of the CRAFT corpus is high. The overall poor performance of various systems indicates that considerable work needs to be done to enable natural language processing systems to work well when the input is full-text journal articles. The CRAFT corpus provides a valuable resource to the biomedical natural language processing community for evaluation and training of new models for biomedical full text publications.

Componential Analysis of English Verbs.

Computational lexical resources such as WordNet, PropBank, VerbNet, and FrameNet are in regular use in various NLP applications, assisting in the never-ending quest for richer, more precise semantic representations. Coherent class-based organization of lexical units in VerbNet and FrameNet can improve the efficiency of processing by clustering similar items together and sharing descriptions. However, class members are sometimes quite different, and the clustering in both can gloss over useful fine-grained semantic distinctions. FrameNet officially eschews syntactic considerations and focuses primarily on semantic coherence, associating nouns, verbs and adjectives with the same semantic frame, while VerbNet considers both syntactic and semantic factors in defining a class of verbs, relying heavily on meaning-preserving diathesis alternations. Many VerbNet classes significantly overlap in membership with similar FrameNet Frames, e.g., VerbNet Cooking-45.3 and FrameNet Apply_heat, but some VerbNet classes are so heterogeneous as to be difficult to characterize semantically, e.g., Other_cos-45.4. We discuss a recent addition to the VerbNet class semantics, verb-specific semantic features, that provides significant enrichment to the information associated with verbs in each VerbNet class. They also implicitly group together verbs sharing semantic features within a class, forming more semantically coherent subclasses. These efforts began with introspection and dictionary lookup, and progressed to automatic techniques, such as using NLTK sentiment analysis on verb members of VerbNet classes with an Experiencer argument role, to assign positive, negative or neutral labels to them. More recently we found the Brandeis Semantic Ontology (BSO) to be an invaluable source of rich semantic information and were able to use a VerbNet-BSO mapping to find fine-grained distinctions in the semantic features of verb members of 25 VerbNet classes. This not only confirmed the assignments previously made to classes such as Admire-31.2, but also gave a more fine-grained semantic decomposition for the members. Also, for the Judgment-31.1 class, the new method revealed new, more fine-grained existing semantic features for the verbs. Overall, the BSO mapping produced promising results, and as a manually curated resource, we have confidence the results are reliable and need little (if any) further hand-correction. We discuss our various techniques, illustrating the results with specific classes.

Semantic Representations for NLP Using VerbNet and the Generative Lexicon.

The need for deeper semantic processing of human language by our natural language processing systems is evidenced by their still-unreliable performance on inferencing tasks, even using deep learning techniques. These tasks require the detection of subtle interactions between participants in events, of sequencing of subevents that are often not explicitly mentioned, and of changes to various participants across an event. Human beings can perform this detection even when sparse lexical items are involved, suggesting that linguistic insights into these abilities could improve NLP performance. In this article, we describe new, hand-crafted semantic representations for the lexical resource VerbNet that draw heavily on the linguistic theories about subevent semantics in the Generative Lexicon (GL). VerbNet defines classes of verbs based on both their semantic and syntactic similarities, paying particular attention to shared diathesis alternations. For each class of verbs, VerbNet provides common semantic roles and typical syntactic patterns. For each syntactic pattern in a class, VerbNet defines a detailed semantic representation that traces the event participants from their initial states, through any changes and into their resulting states. The Generative Lexicon guided the structure of these representations. In GL, event structure has been integrated with dynamic semantic models in order to represent the attribute modified in the course of the event (the location of the moving entity, the extent of a created or destroyed entity, etc.) as a sequence of states related to time points or intervals. We applied that model to VerbNet semantic representations, using a class's semantic roles and a set of predicates defined across classes as components in each subevent. We will describe in detail the structure of these representations, the underlying theory that guides them, and the definition and use of the predicates. We will also evaluate the effectiveness of this resource for NLP by reviewing efforts to use the semantic representations in NLP tasks.

BioVerbNet: a large semantic-syntactic classification of verbs in biomedicine.

BACKGROUND: Recent advances in representation learning have enabled large strides in natural language understanding; However, verbal reasoning remains a challenge for state-of-the-art systems. External sources of structured, expert-curated verb-related knowledge have been shown to boost model performance in different Natural Language Processing (NLP) tasks where accurate handling of verb meaning and behaviour is critical. The costliness and time required for manual lexicon construction has been a major obstacle to porting the benefits of such resources to NLP in specialised domains, such as biomedicine. To address this issue, we combine a neural classification method with expert annotation to create BioVerbNet. This new resource comprises 693 verbs assigned to 22 top-level and 117 fine-grained semantic-syntactic verb classes. We make this resource available complete with semantic roles and VerbNet-style syntactic frames. RESULTS: We demonstrate the utility of the new resource in boosting model performance in document- and sentence-level classification in biomedicine. We apply an established retrofitting method to harness the verb class membership knowledge from BioVerbNet and transform a pretrained word embedding space by pulling together verbs belonging to the same semantic-syntactic class. The BioVerbNet knowledge-aware embeddings surpass the non-specialised baseline by a significant margin on both tasks. CONCLUSION: This work introduces the first large, annotated semantic-syntactic classification of biomedical verbs, providing a detailed account of the annotation process, the key differences in verb behaviour between the general and biomedical domain, and the design choices made to accurately capture the meaning and properties of verbs used in biomedical texts. The demonstrated benefits of leveraging BioVerbNet in text classification suggest the resource could help systems better tackle challenging NLP tasks in biomedicine.

A neural classification method for supporting the creation of BioVerbNet.

BACKGROUND: VerbNet, an extensive computational verb lexicon for English, has proved useful for supporting a wide range of Natural Language Processing tasks requiring information about the behaviour and meaning of verbs. Biomedical text processing and mining could benefit from a similar resource. We take the first step towards the development of BioVerbNet: A VerbNet specifically aimed at describing verbs in the area of biomedicine. Because VerbNet-style classification is extremely time consuming, we start from a small manual classification of biomedical verbs and apply a state-of-the-art neural representation model, specifically developed for class-based optimization, to expand the classification with new verbs, using all the PubMed abstracts and the full articles in the PubMed Central Open Access subset as data. RESULTS: Direct evaluation of the resulting classification against BioSimVerb (verb similarity judgement data in biomedicine) shows promising results when representation learning is performed using verb class-based contexts. Human validation by linguists and biologists reveals that the automatically expanded classification is highly accurate. Including novel, valid member verbs and classes, our method can be used to facilitate cost-effective development of BioVerbNet. CONCLUSION: This work constitutes the first effort on applying a state-of-the-art architecture for neural representation learning to biomedical verb classification. While we discuss future optimization of the method, our promising results suggest that the automatic classification released with this article can be used to readily support application tasks in biomedicine.

Temporal Annotation in the Clinical Domain.

This article discusses the requirements of a formal specification for the annotation of temporal information in clinical narratives. We discuss the implementation and extension of ISO-TimeML for annotating a corpus of clinical notes, known as the THYME corpus. To reflect the information task and the heavily inference-based reasoning demands in the domain, a new annotation guideline has been developed, "the THYME Guidelines to ISO-TimeML (THYME-TimeML)". To clarify what relations merit annotation, we distinguish between linguistically-derived and inferentially-derived temporal orderings in the text. We also apply a top performing TempEval 2013 system against this new resource to measure the difficulty of adapting systems to the clinical domain. The corpus is available to the community and has been proposed for use in a SemEval 2015 task.

Assessment of software testing and quality assurance in natural language processing applications and a linguistically inspired approach to improving it.

Significant progress has been made in addressing the scientific challenges of biomedical text mining. However, the transition from a demonstration of scientific progress to the production of tools on which a broader community can rely requires that fundamental software engineering requirements be addressed. In this paper we characterize the state of biomedical text mining software with respect to software testing and quality assurance. Biomedical natural language processing software was chosen because it frequently specifically claims to offer production-quality services, rather than just research prototypes. We examined twenty web sites offering a variety of text mining services. On each web site, we performed the most basic software test known to us and classified the results. Seven out of twenty web sites returned either bad results or the worst class of results in response to this simple test. We conclude that biomedical natural language processing tools require greater attention to software quality. We suggest a linguistically motivated approach to granular evaluation of natural language processing applications, and show how it can be used to detect performance errors of several systems and to predict overall performance on specific equivalence classes of inputs. We also assess the ability of linguistically-motivated test suites to provide good software testing, as compared to large corpora of naturally-occurring data. We measure code coverage and find that it is considerably higher when even small structured test suites are utilized than when large corpora are used.

Towards comprehensive syntactic and semantic annotations of the clinical narrative.

OBJECTIVE: To create annotated clinical narratives with layers of syntactic and semantic labels to facilitate advances in clinical natural language processing (NLP). To develop NLP algorithms and open source components. METHODS: Manual annotation of a clinical narrative corpus of 127 606 tokens following the Treebank schema for syntactic information, PropBank schema for predicate-argument structures, and the Unified Medical Language System (UMLS) schema for semantic information. NLP components were developed. RESULTS: The final corpus consists of 13 091 sentences containing 1772 distinct predicate lemmas. Of the 766 newly created PropBank frames, 74 are verbs. There are 28 539 named entity (NE) annotations spread over 15 UMLS semantic groups, one UMLS semantic type, and the Person semantic category. The most frequent annotations belong to the UMLS semantic groups of Procedures (15.71%), Disorders (14.74%), Concepts and Ideas (15.10%), Anatomy (12.80%), Chemicals and Drugs (7.49%), and the UMLS semantic type of Sign or Symptom (12.46%). Inter-annotator agreement results: Treebank (0.926), PropBank (0.891-0.931), NE (0.697-0.750). The part-of-speech tagger, constituency parser, dependency parser, and semantic role labeler are built from the corpus and released open source. A significant limitation uncovered by this project is the need for the NLP community to develop a widely agreed-upon schema for the annotation of clinical concepts and their relations. CONCLUSIONS: This project takes a foundational step towards bringing the field of clinical NLP up to par with NLP in the general domain. The corpus creation and NLP components provide a resource for research and application development that would have been previously impossible.

Normalization and standardization of electronic health records for high-throughput phenotyping: the SHARPn consortium.

RESEARCH OBJECTIVE: To develop scalable informatics infrastructure for normalization of both structured and unstructured electronic health record (EHR) data into a unified, concept-based model for high-throughput phenotype extraction. MATERIALS AND METHODS: Software tools and applications were developed to extract information from EHRs. Representative and convenience samples of both structured and unstructured data from two EHR systems-Mayo Clinic and Intermountain Healthcare-were used for development and validation. Extracted information was standardized and normalized to meaningful use (MU) conformant terminology and value set standards using Clinical Element Models (CEMs). These resources were used to demonstrate semi-automatic execution of MU clinical-quality measures modeled using the Quality Data Model (QDM) and an open-source rules engine. RESULTS: Using CEMs and open-source natural language processing and terminology services engines-namely, Apache clinical Text Analysis and Knowledge Extraction System (cTAKES) and Common Terminology Services (CTS2)-we developed a data-normalization platform that ensures data security, end-to-end connectivity, and reliable data flow within and across institutions. We demonstrated the applicability of this platform by executing a QDM-based MU quality measure that determines the percentage of patients between 18 and 75 years with diabetes whose most recent low-density lipoprotein cholesterol test result during the measurement year was <100 mg/dL on a randomly selected cohort of 273 Mayo Clinic patients. The platform identified 21 and 18 patients for the denominator and numerator of the quality measure, respectively. Validation results indicate that all identified patients meet the QDM-based criteria. CONCLUSIONS: End-to-end automated systems for extracting clinical information from diverse EHR systems require extensive use of standardized vocabularies and terminologies, as well as robust information models for storing, discovering, and processing that information. This study demonstrates the application of modular and open-source resources for enabling secondary use of EHR data through normalization into standards-based, comparable, and consistent format for high-throughput phenotyping to identify patient cohorts.

The MiPACQ clinical question answering system.

The Multi-source Integrated Platform for Answering Clinical Questions (MiPACQ) is a QA pipeline that integrates a variety of information retrieval and natural language processing systems into an extensible question answering system. We present the system's architecture and an evaluation of MiPACQ on a human-annotated evaluation dataset based on the Medpedia health and medical encyclopedia. Compared with our baseline information retrieval system, the MiPACQ rule-based system demonstrates 84% improvement in Precision at One and the MiPACQ machine-learning-based system demonstrates 134% improvement. Other performance metrics including mean reciprocal rank and area under the precision/recall curves also showed significant improvement, validating the effectiveness of the MiPACQ design and implementation.

Towards temporal relation discovery from the clinical narrative.

Disease progression and understanding relies on temporal concepts. Discovery of automated temporal relations and timelines from the clinical narrative allows for mining large data sets of clinical text to uncover patterns at the disease and patient level. Our overall goal is the complex task of building a system for automated temporal relation discovery. As a first step, we evaluate enabling methods from the general natural language processing domain - deep parsing and semantic role labeling in predicate-argument structures - to explore their portability to the clinical domain. As a second step, we develop an annotation schema for temporal relations based on TimeML. In this paper we report results and findings from these first steps. Our next efforts will scale up the data collection to develop domain-specific modules for the enabling technologies within Mayo's open-source clinical Text Analysis and Knowledge Extraction System.

Nominalization and alternations in biomedical language.

BACKGROUND: This paper presents data on alternations in the argument structure of common domain-specific verbs and their associated verbal nominalizations in the PennBioIE corpus. Alternation is the term in theoretical linguistics for variations in the surface syntactic form of verbs, e.g. the different forms of stimulate in FSH stimulates follicular development and follicular development is stimulated by FSH. The data is used to assess the implications of alternations for biomedical text mining systems and to test the fit of the sublanguage model to biomedical texts. METHODOLOGY/PRINCIPAL FINDINGS: We examined 1,872 tokens of the ten most common domain-specific verbs or their zero-related nouns in the PennBioIE corpus and labelled them for the presence or absence of three alternations. We then annotated the arguments of 746 tokens of the nominalizations related to these verbs and counted alternations related to the presence or absence of arguments and to the syntactic position of non-absent arguments. We found that alternations are quite common both for verbs and for nominalizations. We also found a previously undescribed alternation involving an adjectival present participle. CONCLUSIONS/SIGNIFICANCE: We found that even in this semantically restricted domain, alternations are quite common, and alternations involving nominalizations are exceptionally diverse. Nonetheless, the sublanguage model applies to biomedical language. We also report on a previously undescribed alternation involving an adjectival present participle.