1. Chapter 1. Introduction
The Exponential Growth of Biomedical Research Data
The current capabilities of our biomedical research enterprise, exemplified by the
completion of Human Genome Project, enable researchers to quickly and routinely
survey the contents of entire molecular and cellular systems. This capability is generating
a revolution in biomedical research in various profound ways. One significant change is
the availability of staggering amounts of genomic and functional genomic data gathered
at a whole genome or whole cell scale. As the result of such tremendous technology
breakthroughs, the challenge for biomedical research is being shifted from experimental
data generation to the organization, curation and interpretation of these data (Lander ES
et al, 2001; Meldrum D et al, 2000).
Biomedical research literature can be considered to be a knowledgebase that
comprises the most complete status of our research enterprise. Reflecting the geometric
growth of available experimental data, the publication rate in biomedicine is also
increasing exponentially. There are currently more than 17 million biomedical articles
already represented in the National Library of Medicine’s biomedical literature database
MEDLINE, including more than 3 million articles published within last 5 years alone and
2,000 per day in 2006 (Hunter L et al, 2006; MEDLINE). Keeping abreast of this large
and ever-expanding body of information is increasingly daunting for researchers in order
to track and utilize what’s relevant to their interests, especially for new investigators. For
example, the pediatric tumor neuroblastoma is a common pediatric tumor but considered
to be quite rare overall, with approximately 600 new cases diagnosed in the US each
1
2. year. However, there are almost 25,000 research articles describing neuroblastoma,
making it virtually impossible for a new investigator to systematically assess historical
research on this topic.
Furthermore, researchers have the increasing need to get in touch with the
research fields outside their core competence. The commonly used PubMed system,
which provides a convenient query interface for MEDLINE, provides keyword search
and some concept mapping for researchers to narrow down the information they are
looking for (PubMed). However, its capabilities lack the precision (positive predictive
value), recall (sensitivity), granularity, and relevance ranking capabilities that many
typical but complex research queries have. One of the most popular demands that
general-purpose systems such as PubMed fail to satisfy is the ability to extract and
compile specific knowledge or facts out of literature records. For example, there is no
provision in PubMed-like systems to determine which genes have been studied thus far in
relation to a certain type of malignancy, other than to read through the set of articles
identified by PubMed using keywords defining the concepts “gene” and “cancer” (or the
type of cancer of interest), and then identifying the particular genes one article at a time.
With the exponentially increasing literature size, the process will not only be more time
consuming, but also be less reliable on getting the right articles. Consequently, the gap
between what is recognized and what is currently known is widening (Wren JD et al,
2004). Biomedical text mining techniques can help researchers meet this challenge by
developing automated systems to extract the relevant information out of the text and
organize it into a structured knowledgebase.
2
3. Data Integration Opportunities in Cancer Research
The general challenge of biomedical literature knowledge extraction is
confounded in cancer research, including an acute need to more systematically identify
linkages between genomic data and malignant phenotypes. Characterization of the
molecular aberrations responsible for the onset and progression of malignancy is a major
goal for cancer researchers, and genomic components of the aberrations, ranging from
base pair variance to chromosome deletion, are crucial determinants in this regard.
Despite the existence of some locus-, mutation- and disease-specific resources, there is
currently no central cancer knowledge database in the public domain integrating genomic
findings with phenotypic observations of tumors (Cairns J et al, 2000; Freimer N et al,
2003). While high-throughput screening efforts increasing allow researchers to identify
genome-wide mutational profiles for specific tumors, this information is largely diffusely
distributed and is mostly catalogued in a semi-structured manner throughout the
biomedical literature. Such decentralization is holding back the efforts towards making
rapid and comprehensive inferences of the genomic basis of malignancy onset and
progression in a manner that incorporates cumulative knowledge. Ideally, researchers and
clinicians would likely benefit from a comprehensive cancer knowledgebase that
consolidates experimental work (genome-level investigation), clinical observations
(descriptions of phenotype) and patient outcome (efficacy of treatment). Because the
biomedical literature represents a large proportion of this information, which is both
critically reviewed and eventually objective in its presentation of cancer research
information, means for more adequately extracting, normalizing and relating such diverse
collections of information in literature are crucial to solving this data integration problem
3
4. in cancer research.
Named Entity Recognition
The successful development of text mining technology has been increasingly
applied in biomedical research to assist with meeting the above-mentioned challenges.
There have been significant efforts from both computational linguists and
bioinformaticists within the past 5 years to develop automated biomedical text mining
(BTM) systems (Jensen LJ et al, 2006). BTM tasks include named entity recognition
(NER), information extraction (IE), document retrieval (DR), and literature-based
discovery (LBD). NER, which serves as the basis for most other BTM undertakings, is
the process of identifying mentions of biomedical entities (objects, such as genes and
diseases) in the text. Named entity recognition can be at first deceptively straightforward,
but it is has emerged as a challenging and considerable task in BTM research. NER
begins with the classification and definition of biomedical entities, which easily
consumes tremendous amount of effort because of the complex and lack-of-standard
nature in biomedical entities.
The process of identifying references to biomedical objects in text is usually split
into two steps: the identification of mentions of specific entity instances in text, such as
“the p53 gene” or “acute lymphoblastic leukemia”; and the assignment of these mentions
to a standard referent (normalization), such as classifying “the p53 gene” as a mention of
the official gene symbol “TP53”, or “ALL” as “acute lymphoblastic leukemia”. Many
biomedical entities either lack controlled vocabularies that can act as sufficient
nomenclature standards, or the instances in text are not expressed with the standards due
to historical reasons. Therefore, normalization is absolutely necessary for equating entity
4
5. values as appropriate, or placing values into a hierarchical or ontological framework (e.g.,
“ALL” as a form of “leukemia”. Much BTM research to date has focused upon molecular
entities that tend to be more discretely definable, such as genes and protein-protein
interactions, than phenotypic entities, which are harder to classify semantically
(BioCreAtIvE; McDonald R et al, 2005; Settles BA 2005; Zhou G et al, 2005).
NER methods include both rule-based and machine-learning approaches. Rule-
based approaches use sets of “rules”, alone or in combination, that pre-state signature
grammatical and especially character and word-based patterns within a string of text
being considered, and then return Boolean values as an output. For example, a rule to
identify a gene name could be “This word is a gene if it contains the consecutive letters
‘KIAA”, all of which are capitalized”. There can be some allowance for lexical
variations, such as capitalization, stemming, or punctuation, and some or all rules might
compare the text being considered to a term list, such as a pre-compiled list of known
tumor types. However, the performance of the approach can’t count on the completion of
the dictionary-type list in terms of both depth (the completion of the entity unique
identifiers) and breadth (the completion of the synonyms for each unique identifier)
because for most biomedical entities, the term lists are always changing and never
complete. For complexly formulated text, rule-based approaches typically require
considerable thought and exquisite biological knowledge. Advantages of this approach
are relatively high precision without the requirement for generating extensive training
material. However, disadvantages include high false negative rates, a performance
plateau that is increasingly difficult to overcome, and, for complex and heterogeneous
text, a tendency to generate low recall. Most first-generation systems and many domain-
5
6. focused current systems utilize rule-based approaches; when coupled with a term list, this
approach accomplishes both steps of the overall NER task at one time. However, rule-
based systems have enjoyed only modest success for biomedical applications, likely
because their performances have plateaued below rates acceptable for wide use by
researchers, or their application domains have been overtly narrow (Hanisch D et al,
2005; Fundel K et al, 2005; Chang JT et al, 2004; Finkel J et al, 2005).
Given the limitations of rule-based systems, a number of machine-learning
algorithms have been applied to improve the first step of the NER task. Generally, these
algorithms consider and then define sets of features within and surrounding entity
mentions that co-associate with the mentions. These can include orthographic features of
the text (e.g., suffixes, particular sequential combinations of characters or words,
capitalization patterns, etc.) and domain-specific features (e.g., term lists). For example,
the suffix “-ase” usually indicates a protein name, and the noun phrase immediately
preceding the word “gene” is often a gene name. Machine-learning approaches have
several advantages: at their purest, they require no domain knowledge; they can consider
thousands or millions of features simultaneously; they can provide confidence scores for
predictions; and they can consider the entire feature space simultaneously. However, the
success of machine-learning approaches is dependent upon two critical and costly factors.
First, ML systems require the establishment, quality, and representativeness of a set of
manually generated training material from which to “learn” features, a process that
requires considerable effort and does not generalize effectively. Second, the most
effective systems incorporate biological knowledge—either in the form of domain-
specific rules or definition of features that are domain-specific (such as specialized
6
7. lexicons)—that are likewise costly to implement (McDonald R et al, 2004; Coller N et al,
2000; Tanabe L et al, 2002).
It is most critical to let human set the examples of gold standards before machines
can learn from it. To better reduce the annotation ambiguity and disagreement, it is
crucial to define the target biomedical entities explicitly. Currently, most developed NER
systems take some version of pre-established conceptual definitions, by which annotators
could apply with very different standards. We have tried otherwise and put tremendous
effort in an iterative annotation process to develop literature-based definitions drawing
both the conceptual and textual boundaries.
Step 2 work (normalization) is syntactically easier since the identification of
textual boundaries is not necessary. However, it poses significant semantic challenges,
because the non-unique synonyms have to be disambiguated to find out the real intent.
And also, a comprehensive thesaurus like dictionary is necessary in order to match the
raw entity mentions to their unique identifiers. Classification techniques, rule-based
systems, and pattern-matching algorithms have been utilized to solve this issue, and some
approaches also take the contextual information to disambiguate the synonyms (Chen L
et al, 2005).
Information Extraction
Ideally, BTM systems extract and synthesize “facts” out of the literature that
combine entity mentions with relationships between and among the mentions established
in the literature. This work requires NER results, that is, the relationships between the
entities can only be extracted once the individual entities have been identified. Although
biomedically oriented research in this area is not as advanced as NER, BTM researchers
7
8. have recently been increasing their efforts on these challenges.
A most straightforward but powerful approach is co-occurrence. This approach
identifies the relationships between the involved biomedical entities based on their co-
occurrence in the articles, or by considering how close mentions are to each other within
a document. The assumption taken by the co-occurrence method is that if two (or more)
entity instances are co-mentioned in one single text record (or defined subset, such as a
sentence or a paragraph), these instances have some type of underlying biological
relationship. As it is possible that entity instances can coincidentally co-occur, systems
commonly use some parameters to rank the relationships, such as the frequency and
location of their co-occurrence. If two entity instances are repeatedly co-mentioned
together in close proximity, it is most likely that they are related. This approach tends to
perform with better recall but at the expense of precision because it has no intelligent
means for distinguishing specific from general relationships. For example, if the
information to be extracted is the causal relationship between gene A and disease
diagnostic labels, this approach will recognize relationships of any kind between gene A
and relevant diseases, including but not limited to direct or causal relationships. In order
to improve precision, some co-occurrence-based IE systems include additional
approaches, such as combining with a customized text-categorization system to
preferentially identify relevant articles or sentences. Co-occurrence-based IE systems are
usually used as exploratory tools making inferential calls since they can identify both
direct and indirect relationships between entity instances (Jessen TK et al, 2001; Alako
BT et al, 2005).
Another approach is to take advantage of natural language processing (NLP)
8
9. methodology that combines syntactic and semantic analysis of text. In this approach,
individual tokens in test are often first identified and then assigned part-of-speech labels,
in a process that has been converted to automation with high accuracy. Then a nested tree
like structure (either top-down or bottom-up) is developed in order to determine the
relationships between noun phrases or beyond, such as subjective and objective. After a
NER process is applied for assigning semantic labels to specific words and phrases, either
rule-based or machine-learning based processes can be used to extract relationships
between entity mentions. Although the syntactic parsing and the semantic labeling have
been carried out as separate steps by most NLP-based IE systems, results indicate that
better performance can be obtained by integrating the two steps, due in part to the often
complex relationships of biomedical entity mentions. This NLP-based approach can
achieve better precision, but lower recall, largely because of increased challenges in
identifying relationships across sentences. These approaches are also labor-intensive,
since either expert defined sophisticated extraction rules or manually annotated training
corpus are required (Rzhetsky A et al, 2004; Daraselia N et al, 2004; Yakushiji A et al,
2001).
Although there is some research touching base with n-ary relationships between a
set of biomedical entities, most IE systems currently classify binary relationships between
same-type entities. These systems most commonly focus on entities and relationships that
are easier to define, such as protein-protein/gene-protein interactions, protein
phosphorylation, other specific relations between genomic entities such as cellular
localizations of proteins, or interactions between proteins and chemicals. Few NER
systems have yet to be designed for relating phenotypic attributes, such as gene-disease
9
10. relationships (Temkin et al, 2003; McDonald R et al, 2005).
High-performance systems that can extract many types of relationships and also
distinguish among relationships beyond the sentence level are not yet achievable. This is
due largely to three contributing factors. First, biomedical text is complex and highly
variable in its structure and presentation. Second, many complicating factors need to be
considered, including co-reference (e.g, the use of pronouns), ambiguity in intent, and
variability in formulation. Finally, systems need to incorporate various approaches
simultaneously (e.g., tokenizers, POS taggers, NER systerms, parsers, disambiguators),
each of which contributes some measure of error that combines to significantly degrade
finalized output (Ding J et al, 2002).
Document Retrieval
DR systems typically identify and rank documents pertaining to a certain topic
from a large collection of text. Topics of interest might be derived from user-supplied
search terms or from pre-selecting specified types of documents. Most DR systems
feature keyword search capabilities; advanced keyword searching allows users to input a
combination of search terms and/or to perform advanced functions, such as including
logical operations or inducing limits to terms. Systems then commonly retrieve
documents containing or excluding certain terms that match the search criteria. This
method often retrieves irrelevant articles, and relevance-ranking functions are often
absent or primitive. More sophisticated DR systems go beyond this by applying distance
metrics, such as a vector-space model. With this model, every document is represented as
a vector, which is determined by measuring text-based features and/or document
metadata, such as a list of frequency-based weighted terms identified in each document.
10
11. The query vector, which is determined by the relative importance of each query term, is
then compared to document vectors to relevance rank the documents. The comparison
between document vectors can also calculate document similarity. PubMed is a well-
known DR system that is highly adapted for use as a query interface for MEDLINE.
PubMed uses both keyword searching and a vector model (Glenisson P et al, 2003).
Advanced DR systems integrate NER or other NLP methods in order to more
accurately assess document content and identify documents that mention certain
biomedical entity mentions. FABLE, MedMiner and Textpresso are examples of systems
that make retrieval decisions by extracting and considering knowledge from gene/protein
mentions in the documents (FABLE; Tanabe L et al, 1999; Muller HM et al, 2004).
Literature-Based Discovery
An ultimate goal of BTM is to assist with literature-based discovery. LBD can be
defined as a process that discovers testable novel hypotheses by inferring implicit
knowledge in biomedical literature. An early and often-cited example of LBD was from
researcher recognizance of facts from two unrelated bodies of biomedical text, describing
Raynaud’s disease, in which patients suffer from vasoconstriction, high blood viscosity
and platelet aggregability, and describing fish oil, indicating that besides its capability of
causing vasodilation, its active ingredient can also lower blood viscosity and platelet
aggregation. This connection was formed completely through extensive reading of the
literature, and later the relationship was proved experimentally. The model used in this
seminal example was very simple: if A leads to B, and B leads to C, then it is plausible
that A could lead to C. Based on this closed discovery process (to connect two previously
known relations), this researcher subsequently discovered a novel association between
11
12. migraine and magnesium deficiency (also proved experimentally) as well as additional
successes (Swanson DR 1986; Swanson DR 1988; Swanson DR 1990).
More challenging LBDs might arise from an open discovery process, which
attempts to derive relationships between two entities of interest through implicit
relationships in literature. For example, the process of identifying candidate genes for a
certain disease is an open discovery process. One example of this process would be to
first identify gene mentions co-occurring in the literature (gene set A) with mentions of a
disease of interest, next identifiying co-occurring gene mentions (gene set B) with known
disease genes, and then consider the overlap between the two sets of gene mentions as
candidate genes for the disease. There are two assumptions taken for this approach: Gene
set B is functionally related with known disease genes; Gene set A has some sort of
relations with the disease. One potential problem for this approach is that there are many
types of direct and indirect relationships identified in such a process, including the high
likelihood that a substantial number of false positives are generated. NLP-based IE can
certainly help narrow down the relationship types, but further research is needed to
improve the performance of such models. Also fundamentally, literature inevitably
contains conflicting and inaccurate statements, which is impossible for an automated
algorithm to adjudicate (Weeber M et al, 2005).
It is much likely that more reliable inference of novel hypotheses and research
directions from literature achieves success by integration of BTM results with other data
types, including from curated data sets and experimental data. Experts’ curation and
experimental evidence provides verification, filtering, and relevance ranking capabilities
from information derived from real biological relationships between entities. For
12
13. example, researchers have made novel discoveries by transferring text-mined
relationships of a protein to its orthologous proteins based on sequence-similarity
searches. The integration effort of BTM results with functional genomic data such as
microarray data has helped researchers rank significant genes as well as develop novel
hypotheses based on both experimental data and previously known knowledge in a large
scale, automated fashion (Yandell MD et al, 2002; Raychaudhuri S et al, 2002; Glenisson
P et al, 2004).
Significance
Along with the rapid expanding of experimental data, the exponential increase of
the biomedical research text makes it more and more difficult for researchers to track and
utilize the relevant information to their interests, especially for the domains outside their
core competence. Automated text mining systems can process the unstructured
information in the literature into structured, queryable knowledgebase. This dissertation
research has developed well-performed automated entity extractors based on the refined
manual annotation with iteratively defined literature-based entity definitions in genomic
variation of malignancy. Co-occurrence-based information extraction process was
applied to integrate with microarray expression data in the pursuit of determining
neuroblastoma research candidate genes. Both functional pathway analysis and RT-PCR
experiment validated the text mining’s contribution. This thesis demonstrated that in
addition to systematic curation of the textual information, biomedical text mining also
has inferential capability especially when combined with experimental data.
13
14. Introduction to the Thesis
Using the genomics of malignancy as a test bed, this thesis has touched upon
every aspect of BTM outlined above. Work regarding the BTM process developed and
employed will be discussed in detail in Chapter 2 and Chapter 3. This thesis has also
established important work regarding information extraction in this domain, which has
been applied to research regarding the pediatric tumor neuroblastoma (Chapter 3 and
Chapter 4). Integration of BTM-extracted information with expression array analytical
results to discover candidate genes for neuroblastoma research will be discussed in detail
in Chapter 4.
14
15. Chapter 2. Defining Biomedical Entities for Named Entity Recognition
Yang Jin
Mark A. Mandel
Peter S. White
Abstract
The performance of machine-learning based named entity recognition is highly
dependent upon the quality of the training data, which is commonly generated by manual
annotation of biomedical text representative of the target domain. The development of
robust definitions of biomedical entities of interest is crucial for highly accurate
recognition but is often neglected by text-mining applications. While the conceptual and
syntactic complexities of biomedical entities often generate ambiguities in assigning text
mentions to particular entity classes, entity definitions that exhibit as distinct semantic
and textual boundaries as possible are desired. We have created a highly generalizable
process for developing entity definitions specifying both conceptual limits and detailed
textual ranges for target biomedical entities. This process utilizes representative text and
manual annotators to initially define and iteratively refine definitions. The process was
tested within the knowledge domain of genomic variation of malignancy. This work
describes in detail the different types of challenges faced and the corresponding solutions
devised during the definition process. The resulting entity definitions were used to
annotate a training corpus for the development of automated entity extraction algorithms
and for use by the research community. We conclude that manual annotation consistency
is useful for the success of later biomedical text mining tasks, and that explicit, boundary-
defined entity definitions can assist with achieving this goal.
15
16. 1. Introduction
Automated information extraction techniques can assist in the acquisition,
management and curation of data. A necessary first step is the ability to automatically
recognize biomedical entities in text, as also known as named entity recognition (NER).
Development of named entity extractors for biomedical literature has progressed rapidly
in recent years. For example, a number of machine-learning algorithms currently exist for
identifying gene name instances in text (Collier N et al, 2000; Tanabe L et al, 2002;
GENIA; Hanisch D et al, 2005). However, a major shortcoming of many approaches is
that they often minimize efforts to define biomedical entities in an explicit fashion.
Rather, the tendency is often to ignore this step by adapting or refining existing semantic
standards as the target entities’ conceptual definitions, leaving interpretive details to
manual annotators. Additionally, existing standards often provide little or none of the
semantic depth required to establish concept boundaries with enough rigidity to provide
highly accurate extraction. This tends to create outstanding consistency problems in later
steps when training automated extractors and utilizing the extracted entity mentions for
particular applications, because non-literature based conceptual definitions often generate
significant annotation ambiguity problems due to the semantic as well as syntactic
complexities of biomedical entities in the literature. As a result, automated systems
derived from such systems tend to perform more poorly. For biologists in particular, high
precision is a necessary prerequisite for widespread acceptance of automated tools, in
order to establish a level of reliability acceptable to users.
Strongly believing the importance of establishing well-defined, literature-based
entity definitions with clear boundaries specially designed for biomedical NER practice,
16
17. the Biomedical Information Extraction Group at University of Pennsylvania (Penn
BioIE) has developed an iterative annotation process designed to establish a set of
“precise” entity definitions. These definitions are meant to clarify the conceptual
boundaries both semantically and syntactically, while also striking a balance between the
requirements of researchers, annotators, and computational scientists. This paper will
first describe the annotation process developed by the Penn BioIE group, and then
introduce the necessities and challenges of defining biomedical entities with specific
examples in the literature.
2. Overview of manual annotation process and entity classification
QuickTimeª and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
Figure 2-1. The processes of developing entity definitions and extractors
Figure 2-1 demonstrates the iterative process developed for establishing and
refining entity definitions, first through manual annotations and then in developing
extractors based on the manually annotated training data. The process begins with the
creation of an initial definition that establishes the general concept and scope of an entity
17
18. class, which is supplied by one or a group of domain experts. Commonly existing
standards and resources are explored and, if deemed suitable, adopted as nuclei for the
process. Subsequently, the domain expert(s) plays the role of adjudicating definition
discrepancies. Manual annotators are then trained with the initial versions of the entity
definitions, from which they manually annotate the selected training corpora. Invariably,
as the annotators encounter the wide diversity of semantic representations of specific
concepts, a need for iterative refinement of the entity definitions emerges. Often, text
encounters require major revisions or even restructuring of definitions to accommodate
such heterogeneity. Accordingly, definitions are continually refined during the analysis of
annotated texts and annotation disambiguation. The Penn BioIE group founded useful
frequent communication forums where the emerging definitions and identified exceptions
were fully discussed among annotators and researchers. Communication modalities
included weekly face-to-face meetings, email lists, and live chat. After annotation has
been executed, entity extractors were developed by implementation of machine-learning
algorithms utilizing probability models (we used Conditional Random Fields); the
manually annotated texts were utilized as both training and testing data for these
algorithms. Comparison of the annotations produced by the automatic extractors and
human annotators allows for evaluation of the extractor performance.
The target knowledge domain we chose was “Genomic Variation of Malignancy”,
conceptualized as a relationship among three entity classes: Gene, Variation and
Malignancy. As shown in Figure 2-2, the Gene and Variation entities comprise genomic
components of cancer while the Malignancy entity covers phenotypic aspects of
malignancy, including malignancy diagnostic labels and a number of malignancy
18
19. phenotypic attributes.
QuickTimeª and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
Figure 2-2. Entity classification scheme for the domain of genomic variation of malignancy
A total of 1442 MEDLINE abstracts were selected for exploration and annotation
in this study, one subset of which contained many different malignancy types to establish
breadth, and a second subset of which mentioned only one major malignancy
(neuroblastoma) to establish depth. As diagrammed in Figure 2-1, the manual annotation
process was first applied to the corpus with an electronic annotation tool, WordFreak
(http://sourceforge.net/projects/wordfreak). After the entity definitions were refined and
stabilized, the manually annotated data were then used to develop entity and attribute
extractors (McDonald RT et al, 2004, Jin Y et al, 2006). These automated extractors
19
20. performed with state-of-the-art accuracy, in part due to the careful design and
management of our annotation process. In the following paragraphs, we will discuss the
challenges we have encountered during the manual annotation process, and why we
believe that consistent entity definitions are critical for the success of later steps in
biomedical text mining.
3. The challenges of defining biomedical entities
Although we began this task believing we had clear ideas of what information
each entity should cover, it quickly proved challenging to develop detailed working
definitions. Our a priori notions of entity definition adequacy were that definitions
establish distinct and defensible boundaries both conceptually and textually, therefore
providing guidance to the annotators both semantically and syntactically. Solid entity
definitions are an essential foundation for the subsequent steps of developing machine-
learning algorithms and utilizing the extracted information for specific applications. First,
the performance of entity extractors is highly dependent not only on the selection of the
underlying algorithms, but also on the quality of the training data, which are entirely
based on the entity definitions. If the annotators cannot identify specific entity mentions
consistently on the basis of the definitions, it is hard to imagine that automated extractors
can replicate this task reliably. More importantly, without clear definitions, researchers
will certainly run into problems when trying to utilize the extracted mentions, since it will
be difficult to know the precise boundaries of the gathered information.
As mentioned earlier, we initially defined three major entities in the knowledge
domain of genomic variation of malignancy, based on existing ontological categories and
concepts. However, we quickly found that ontology-based definitions often don’t
20
21. precisely reflect what has been conceptualized throughout the biomedical texts
contributed by researchers worldwide. For example, a gene defined by NCI thesaurus is:
“A functional unit of heredity which occupies a specific position (locus) on a particular
chromosome, is capable of reproducing itself exactly at each cell division, and directs the
formation of a protein or other product.” If annotators use this definition for identifying
gene mentions in the text, they could quickly be confused by many situations such as
whether promoters should be included; how should gene family names be treated; how
about pronoun referents to genes, etc. Thus, we found the need to invoke text-based
working entity definitions, which are most effectively determined as annotators
proceeded with the entity recognition task in the training corpus. Every new mention of
an entity and every new context for a mention provided a test for the pre-developed entity
definition. If a definition could not explicitly lead the annotators to a “correct”, or at least
consistent decision in each case, the problematic mention required further examination,
interpretation, and possibly, refinement of the definition. Through such an iterative
process, we were able to develop fine-tuned entity definitions that provided distinct
boundaries both for semantic scope and contextual range.
The challenges that we encountered in refining our definitions can be grouped
into four categories: conceptual, syntactic, syntactic/semantic ambiguity, and inter-
annotator agreement. In the following paragraphs we will illustrate these types and give
examples of our devised solutions and their limits.
3.1 Conceptual definition challenges
As discussed earlier, an entity definition has to clarify both conceptual and textual
boundaries. Initial versions of our definitions were completely conceptual, based on our
21
22. understanding of biomedical categories. Surprisingly, more than half of the annotators’
difficulties with definitions fell into this category during the annotation process, and most
of them were reasonable as you can observe in the following paragraphs showing the four
most common challenges in this category. This reflects the semantic complexity and
diversity of biomedical entities, which often cannot be easily defined without some
ambiguity.
3.1.1 Sub-classification of entities
Based on the classification scheme stated above, our target knowledge domain
was initially divided into three major conceptual classes: gene, genomic variation, and
malignancy. However, this broad conceptual classification was far from sufficient for the
generation of highly accurate extractors. For example, according to the conceptual
definition, the malignancy concept covers all phenotypic information of cancer, including
a tumor’s diagnostic type, the tumor’s anatomical location and cellular composition, and
its differentiation status. Each of these types of information are presented in a variable
and often bewildering array of syntactic and contextual patterns, which increases entropy
and thus erodes the ability of machine-learning approaches to classify mentions. If
instead we further classified the mentions into sub-categories such as those described
above and annotated them as such, entropy is reduced and extractor performance can be
expected to improve. However, a major disadvantage of this approach is that, sub-
categorization introduces considerable additional annotation effort. Thus, the annotation
process requires first the establishment of a level of entity granularity that balances the
cost of manual annotation with the application value of the extracted data.
There are countless ways to further divide entities into their underlying
22
23. components. For our purpose, we decided to let the level of granularity be generated by
the annotation process. By beginning with broad classes and subdividing them as needed,
we considered that we would eventually approach an optimal balance between effort and
effectiveness. We considered it to be critical to determine how the text strings represented
subcategories in the real world of biomedical literature. Therefore we divided our
annotation efforts into two stages: data gathering and data classification, as demonstrated
in Figure 2-3 with a genomic variation entity example.
QuickTimeª and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
Figure 2-3. The text-based two-stage entity sub-classification process
In the example illustrated by Figure 2-3, annotation of our initial concept of
“Genomic Variation” proceeded through a preliminary stage of annotation before it was
divided into sub-categories, which we named “Data Gathering”. In this stage, all textual
mentions falling within or partially within our initial concept definition were annotated
regardless of syntax. When sufficient information was gathered, sub-categories were
23
24. defined based on their semantic and syntactic representations. In addition, by proceeding
with this exercise, the annotators became familiar with the concepts, definitions, and
emerging challenges of the tasks. By employing this method, the sub-classification
scheme began to approximate how the concepts were actually presented in the text.
3.1.2 Levels of specificity
Textual entity mentions referring to the same semantic types can range from very
general to quite specific, and not all levels of detail may be appropriate for a particular
project. A gene mention may refer to a specific gene instance in a single cell of a sample,
or to the wild type or a specific variation of the gene; or it may refer to gene families,
super families and generalized classes, which represent classes of genes. For instance,
“MAPK10” or “mitogen-activated protein kinase 10” is a family member of “MAPK”,
which itself belongs to a higher level family “protein kinase”. We made the decision to
include all levels of information for the gene entity except for the most general level such
as “gene”. That is, in the above example, all three levels of gene mentions are legitimate
and should be annotated as such.
The decision was based on a couple of considerations. First of all, gene class
information is valuable information to extract in later steps; although we don’t know
which specific gene it refers to, it does help us narrow down to a class of genes. Second,
if we only include the mentions describing genes at the instance level (the level that can
lead to a specific genomic element), we have to draw a line between gene classes and
instances. Because textual mentions for gene classes and instances are sometimes
interchangeable (researchers tend to use gene class names referring to gene instance
names and vice versa), it will be quite difficult for the automated extractors to distinguish
24
25. between the two. And finally, we exclude gene mentions at the most general level, which
contains no information content or application value to extract. In another words, all
information-containing levels of mentions are included.
3.1.3 Conceptual overlaps between entities
An ideal entity classification scheme should result in independent information
categories without any conceptual overlaps. Unfortunately, the subjective and adaptive
nature of biological objects makes this ideal especially difficult to achieve, especially
when defining two different but related entities. Even a basic concept such as “organism”
is difficult to define when considering entities such as viruses and viroids, self-replicating
machines with attributes necessary but not necessarily sufficient to qualify as life forms.
Because our gene and genomic variation concepts both fall within the genomic domain
and are closely associated, we were very careful to make a clear distinction. Eventually,
our gene entity evolved to encompass solely the names of genes and their downstream
products (i.e., RNAs and proteins), while the genomic variation entity covered specific
descriptions of genomic element variations.
Although our definitions of gene and genomic variation managed to eventually
establish a reasonable boundary between them, for other entities, we found it sometimes
impossible to avoid the conceptual overlapping problem. We encountered such problems
when trying to make a clear division between the entity classes symptom and disease. The
symptom entity was designed to capture subjective or objective evidence of disease, such
as headache, diarrhea or hyperglycemia, while the disease entity captured specific
pathological processes with a characteristic set of symptoms, such as Long QT Syndrome
or lung cancer. As with most cases, the distinction is often clear to domain experts unless
25
26. considerable scrutiny is requested, as it appears to be simple common sense that these
concepts represent two distinct and non-overlapping sets of information. However, when
presented with the broad contextual variation in use and, often, semantic intent, it actually
becomes quite difficult to draw a clear boundary between the two. We quickly found that
many terms can be considered as both symptoms and diseases, depending both upon
intent and the level of domain knowledge available. For example, “arrhythmia” itself is a
disease entity mention, representing a pathological process, but it is usually used as a
diagnostic label of a disease (symptom), such as long QT Syndrome. We certainly don’t
want to have two entity types heavily overlapping with each other, since that will make
the classification unnecessary. That is not the case for the symptom and disease entity
types, and their overlapping mentions are less than approximately 10% overall. Most
conceptually overlapping mentions cannot be put into either category without reading the
text. We leave it to the annotators to determine authors’ intent based on the context and
increasingly, they became quite good at minimizing the disagreement.
3.1.4 Domain-specific clarification
As biological entities tend to be conceptually subjective, we often found it to be
quite challenging and labor-intensive to establish consistent conceptual boundaries. The
process of defining the gene entity is a good example to illustrate this challenge. Initially,
we considered the task of defining a “gene” to be a straightforward task, as this concept is
considered by biologists to be a rather discrete object. The HUGO Gene Nomenclature
Committee (HUGO), the nomenclature body tasked with establishing official names for
human genes, defines a gene as “a DNA segment that contributes to phenotype/function.
In the absence of demonstrated function a gene may be characterized by sequence,
26
27. transcription or homology". On top of that, our gene entity is initially defined as the
nominal reference to a gene or its downstream product in biomedical text. However, as
annotations moved forward, annotators raised more and more questions, forcing us to
make difficult determinations on the boundaries as illustrated below.
An example of biological complexity is the many ways that a gene can contribute
to phenotype. Typically, genes functionally impact biological processes through their
downstream products, proteins. However, there are DNA segments on the genome which
are able to affect phenotype by regulating how genes are expressed in particular
biological contexts. Promoter and enhancer regions, which are distinct segments of DNA
(often far) removed from the DNA segment that directly contributes to an RNA and/or
protein product, are such example. These elements control whether and when the gene
itself is expressed. Although biologists disagree whether promoters should be considered
as genes or components of particular genes, annotators are required to make a decision on
the gene entity boundary limits. In this case, we considered our application domain to be
the most important determinant, as the main focus of our gene entity was to capture those
“traditional genes” that could be directly and consistently associated with a protein. Thus,
we limited our scope of genes to include only what we considered to be biologically
functional DNA segments which are translated into protein products.
There are many more cases that required further clarification of the gene entity
conceptual definition, such as how to deal with segments and multiplexes of
genes/RNAs/proteins. We realized that consistency was more valuable than trying to
establish universal truth, the former of which we considered to be the key to developing
well-performing automated extractors and increasing the application value of extracted
27
28. mentions.
3.2 Syntactic definition challenges
Even with precise conceptual definitions, we found that guidelines needed be made
regarding the textual boundaries of the entity mentions. Although many of these were
syntactical nuances, they were not necessarily trivial for the annotator disagreement. In
order to make consistent automated extractors, we determined that detailed annotation
guidelines were required to make manual annotations consistent between different
annotators. We designed our guidelines to be practical and based on actual contexts,
specifying to the annotators exactly what to do under any uncertain circumstances that we
had encountered.
3.2.1 Associating a text string to an entity mention
There are many different ways to associate a text string with an entity mention in
biomedical literature. In order to harvest consistent training data to develop highly
performed automated extractors, we needed to define a series of rules specifying how to
select text strings in the literature as legitimate entity mentions. We allowed entity
references to include more than one word, including punctuation, but not to cross
sentence boundaries.
Although the majority of the entity mentions were nouns, not all of them were.
For some entity mentions such as variation type, other part-of-speech forms were not
uncommon. For example, for genomic variation types that would likely be normalized as
the forms “insertion”, “deletion”, or “translocation”, those variation type mentions were
usually expressed as verbs: “inserted”, “deleted”, or “translocated”. Moreover,
malignancy attribute mentions were nearly always adjectives, such as “well-
28
29. differentiated”, “hereditary”, and “malignant”.
All modifiers in a noun phrase mention were considered to be included as part of
a mention, because not only can the modifiers provide very useful information to be
extracted, but also that some modifiers are indispensable parts of the standard terms. We
observed that this decision made it easier for both manual annotators and machine-
learning extractors to operate since it was difficult to define boundaries on what
modifiers to include in noun phrases. However, modifiers were not included for other
part-of-speech phrases, in order not to complicate the issue. For example, in a noun
phrase malignancy type mention “malignant squamous cell carcinoma”, both “malignant”
and “squamous cell” are the modifiers of “carcinoma”, and both provide very useful
information. “Squamous cell carcinoma” is also a commonly employed name of a type of
cancer. Our experience determined that it was difficult for annotators and impossible for
automatic extractors to draw consistent boundaries between modifiers on what should be
included as part of the legitimate mentions.
Lastly, we found it necessary to make entity-specific rules for some biological
entities. For example, the gene entity mentions commonly appeared in the text as “The
mycn gene…”, necessitating a decision as to whether the article “The” and the noun
“gene” should be included as part of the entity mention. We reasoned that the decision
should depend on how the extracted information was to be further processed and utilized.
Accordingly, we decided to include neither word, since all the extracted gene mentions
were to be subsequently mapped and normalized to official gene symbols.
3.2.2 Co-reference issue
Often a single entity is referred to in different ways in the same text, a situation
29
30. known as co-reference. Besides its standardized form, an entity instance can also be
referred to by aliases, acronyms, descriptions or pronoun references. For example, the
mycn gene has at least 10 aliases in the literature, including “n-myc”, “oded”, and “v-myc
avian myelocytomatosis viral related oncogene, neuroblastoma derived”. Moreover,
researchers commonly engineer their own acronyms as self-convenient but non-standard
and often unique aliases. Co-reference is generally recognized as a challenging task for
entity recognition and information extraction. To deal with this issue in manual
annotation, we have classified this problem into the following four categories and made
corresponding decisions for each of them.
A. Extended form vs. acronym
Regular expression: ___ ___ ___ (___)
Examples:
• …mitogen-activated protein kinase (MAPK)…-- gene entity mention
• …squamous cell carcinoma (SCC)… -- malignancy type entity mention
Our decision: Tag both the extended form and abbreviated form of the entity mention.
For the above examples, “MAPK” is co-referential with “mitogen-activated protein
kinase”, and “SCC” is co-referential with “squamous cell carcinoma”. Both extended
forms and acronyms would be tagged as corresponding entity instances in our system.
Our rationale: Both forms are interchangeable descriptions of entity mentions, and they
should be treated equally.
B. Alias description
Regular expression: …Y…X… or …Y (X)…
Examples:
30
31. • TrkA (NTRK1)…
• The N-myc gene, or MYCN…
Our decision: NTRK1 and MYCN are official name designations of the TrkA and N-myc
genes, and here they are being co-referenced accordingly. We decided to tag all different
expression forms of the entity instances, including standard/official nomenclatures,
aliases or descriptions. Like acronyms and their extended forms, these various names are
also tagged individually: in the first example, we tagged “TrkA” and “NTRK1”
separately and without the parentheses, not the combined string “TrkA (NTRK1)”.
Our rationale: Researchers often use unofficial nomenclatures for entity mentions, so we
can’t just annotate standard descriptions. However, they should be normalized later.
C. General vs. specific
Regular expression: X, a (the) Y…
Examples:
• C-Kit, a tyrosine kinase which plays an important role, …
• K-Ras is an oncogene. The Ras gene…
Our decision: In the examples above, the gene family name “Ras” and the superfamily
name “tyrosine kinase” are used to co-refer to the gene family instances “K-Ras” and “C-
Kit”. In such situations, our annotation guideline treated the general terms and more
specific terms completely independently, regardless of the co-referential relationship
between them. That is, depending on the conceptual definition, if the term was a
legitimate mention, it was tagged as an entity mention no matter what levels of specificity
it had. For those examples, since the gene entity definition included both gene instances
and family names, all four terms were tagged as gene entity mentions. We did not,
31
32. however, tag “oncogene”, nor did we extend the tag on “Ras” to include the following
word “gene”. These words, at the highest level of generality, convey no taggable
information.
Our rationale: Based on our decision on tagging all information-containing levels of
mentions and specifically for the examples listed, all gene instances, gene families and
superfamilies are determined legitimate mentions.
D. Pronoun reference
Regular expression: …X…PRONOUN (It, This, etc.)…
Examples:
• K-Ras is an oncogene. It is mutated in…
• Five point mutations were found in the MYC gene, and they were next to each
other.
Our decision: In the two examples, “It” is co-referential to “K-Ras”, and “they” is co-
referential to “point mutations”. We generally did not annotate pronouns, although they
may refer to legitimate entity mentions.
Our rationale: Pronoun co-reference is a challenging problem in text mining research,
which involves cross-sentence, whole-record level of relation extraction. Without deeper
parsing of the text, there is no value by extracting the pronoun itself.
3.2.3 Structural overlap between entity mentions
Entities can overlap not only conceptually, but also literally, with their textual
mentions in the literature. Annotation guidelines were developed for the following
32
33. situations:
A. Entity within entity – tag within tag
This refers to the situation that one entity mention is completely included in the
textual range of another. As the two intertwined entity mentions could belong to either
the same or different entities, we divided this category of problem into two sub-
categories. If the two mentions were in the same entity, only the subsuming entity
mention was tagged. For example, in “mitogen-activated protein kinase kinase kinase”,
there exist 7 distinct gene entity mentions: mitogen-activated protein; mitogen-activated
protein kinase; mitogen-activated protein kinase kinase; mitogen-activated protein kinase
kinase kinase; and three mentions of “kinase”. While this type of a situation was a source
of confusion among new annotators, we considered it both unnecessary and costly to tag
all possible mention permutations. As the mention with the largest range was always the
one being discussed, only the outermost mention was considered to be tagged as a gene
mention. In fact, this situation led to the adoption of a more generalized guiding
principle, where the annotation should reflect the author intent whenever possible
(although exceptions were encountered, such as poorly written abstracts where the intent
from the context occasionally and obviously differed from the actual word or phrase
used).
If two completely overlapping mentions instead belonged to different entity types,
we annotated both. These mentions were usually related, and they both often provided
valuable information. Some entities, such as malignancy attributes, often appeared as part
of another entity mention. For instance, “colon cancer” is a malignancy type mention, and
“colon” is a malignancy site mention. “Hirschsprung disease 1” is another example, that
33
34. “Hirschsprung disease” is a disease mention while the whole phrase is a gene mention.
B. Entity co-identity – double tagging
This category represents the situation that two entity mentions share the exact
same text. We annotated the same text twice with the two corresponding labels under
such circumstances. For example, in the phrase “deletion of the K-ras gene”, “K-ras” was
tagged as both a gene entity mention and a variation-location mention.
C. Discontinuous mentions – chaining
Sometimes mentions of several entities of the same type shared a common
substring. When written together in the text, the common part only occured once for the
first or last mention, and other mentions were only represented with the different parts.
For example, in the text “H-, K-, and N-ras…”, there are really three gene mentions: “H-
ras”, “K-ras” and “N-ras”, but a limitation of our annotation software prevented tagging
of discontinuous mentions as one parent mention (in the example above, only “N-ras”
could be tagged. For the other two discontinuous mentions, we developed a chaining,
procedure through which annotators were able to link the component parts (“H-” and
“K-” with “ras”) by inserting comments into the annotation in a standard format.
Chaining was strictly limited within one sentence in order not to complicate issues
for subsequent syntactic parsing of sentences. Employing the same logic, entity mentions
were not allowed to come across different sentences.
3.3 Syntactical vs. Semantic – ambiguity challenges
We considered ambiguity in mentions to be the most common and difficult
challenge in our annotation experience, as it truly reflects the limitation of human-
invented texts in fully communicating author intent. In biomedical text, we found it not
34
35. uncommon that an identical text string could represent completely different concepts, and
the frequency of ambiguity appeared to be much higher than for non-biological text. In
the following paragraphs, we will use mainly gene entity examples to illustrate the
illusive nature of this problem.
We found ambiguity to occur both within and outside gene entities. Genes have a
tradition of being independently named, with poor adherence to or awareness of
standards. People tended to make up new acronyms for gene names, as the result of
which, there are more gene names than the combinations of letters and numbers for short-
character symbols/aliases. Thus, there are lots of similarities between aliases just by
chance. Since each gene has multiple non-unique aliases with one unique gene symbol,
there exists very serious internal ambiguity problem among the aliases. Based on our
calculation, just for human genes alone, there are as many as 3% genes share the same
aliases and the numbers are number higher if including other species. Also, many species
have traditions of naming the genes the same, especially mouse and human (Chen L et al,
2005). For example, p90 is the common alias shared by the distinct gene symbols CANX
and TFRC. As a protein naming convention, p90 actually refers to the protein with
molecular weight 90. Therefore, it is not surprising that there are two proteins with the
same name.
When such gene mentions appear in literature, (often quite distant) context is the
only way to clarify which gene is in discussion, although sometimes it offers no
assistance. Another type of within gene entity ambiguity that we recognized was the
frequent apparent inability to distinguish a gene from its downstream products, based
purely on the text string of the mention. Although initially, our gene entity was designed
35
36. to capture only the nomenclatures of functional genomic elements, we soon discovered
that researchers were frequently using the same referents to represent a gene and also its
RNA and protein products in the literature. Without looking at the context, a gene
mention “mycn” had almost an equal probability to refer to a gene or its downstream
product, and both the gene and its mRNA were referred to as being “expressed” to create
a mRNA or a protein product, respectively. In addition, authors also tended to obscure
the conceptual boundaries between a gene and its downstream products. For example,
while a given protein X performs biological functions, we found it common that the
corresponding gene X was being described as performing this action. It became apparent
that while researchers were personally clear regarding distinctions, their descriptions did
not adequately convey these distinctions. In fact, in several cases, we found it impossible
to determine whether certain gene mentions referred to a gene or its RNA or protein
products even when considering the entire article. This overwhelming ambiguity problem
finally prompted us to reach the decision to include genes’ downstream products when
annotating gene entity mentions. Finally, we created one entity class gene but also
included labels for partially subdividing them, while making considerations for not being
able to perfectly divide mentions into the 3 classes. If it was not clear in the text whether
a mention referred to a gene or a protein, the mention was annotated as “gene.generic”, as
apposed to “gene.gene/RNA” or “gene.protein”.
Besides the challenges mentioned above, it was common to encounter gene entity
mentions that were easily be confused with objects belonging to other entity types, This
is because genes have been named with a wide variety of methods, from the use of lay
languages to the invention of specialized and often clever acronyms. For example, “Cat”
36
37. is an official gene symbol for the gene catalase, while it could also be used to refer to a
kind of animal. “NB” is the acronym of a well-known pediatric cancer neuroblastoma,
but it is also an official name of a gene locus putatively located on chromosome 1p36.
This cross-entity ambiguity problem was also commonly seen for other entity classes,
such as variation type. As an example, “Insertion” and “deletion” are well-defined
variation type mentions, but they are also frequently used to denote biological or clinical
actions. Regardless of the types of the ambiguity problems, the task for our manual
annotators was to make their best calls to identify the intended reference of the text
strings and annotate them as such. Sometimes annotators needed to take entire abstract
or, rarely, the entire article, into consideration in order to determine what particular
mentions truly represented. Depending on the nature of the biomedical entities and how
representative the training data was, the subsequent automatic extractors were able to
disambiguate problematic text strings to certain degree by taking local contextual features
into account.
3.4 Annotator perceptions
Even if perfect entity definitions and annotation guidelines could somehow be
created, there would still be variations among human annotators in understanding and
applying them during the annotation process, and we certainly encountered lively
discussion regarding some topics. Usually, manual annotation is done by different
annotators in order to get more files done within a shorter period of time, but the
downside is that it introduces more inconsistencies between annotators. Even with only
one annotator, there will be variability in application of guidelines.
We took two approaches to deal with this problem. First, annotators were told to
37
38. discuss anything unclear, and we promoted frequent discussion to determine a consistent
path. And also, a dual, sequential-pass manual annotation process was developed and
applied to better adjudicate different annotators’ work and produce training data as
consistent as possible. During this process, every document was annotated de novo by
one annotator and then subsequently checked by a second annotator, who is more
experienced and consistent, charged with identifying and revising any annotations
considered to be incorrect by first pass annotators. Edited items were then subject to
review by the group, and senior annotators used this editing process as an opportunity for
educating less experienced annotators if repeated error patterns were identified.
3.5 Publication-based errors
Typographical and grammatical errors, though infrequent, are inevitable, and
some of them were observed in entity mentions during our process. Due to the
considerations of copyright issues, we were not authorized to change the text in such
cases but instead skipped tagging the mentions with added comments.
4. Application
As a result of the generation and application of these carefully refined entity
definitions and annotation guidelines, 1442 MEDLINE abstracts were manually
annotated. Of these, 1157 files have been made publicly available (release 0.9, BioIE web
site). Since the release, the data has been widely used by the biomedical text mining
community for a variety of purposes, including entity recognition, normalization etc., and
the usage is likely to increase (Cohen KB et al, 2005).
Because of the consistency of the training data across the corpus, the developed
entity and attribute extractors perform with high precision and recall rates. Table 2-1
38
39. indicates the performance of three entity extractors built with this data (McDonald RT et
al, 2004; Jin Y et al, 2006).
Entity Precision Recall F-measure
Gene 0.864 0.787 0.824
Variation Type 0.8556 0.7990 0.8263
Location 0.8695 0.7722 0.8180
State-Initial 0.8430 0.8286 0.8357
State-Sub 0.8035 0.7809 0.7920
Overall 0.8541 0.7870 0.8192
Malignancy type 0.8456 0.8218 0.8335
Table 2-1: Entity extractor performance on evaluation data
5. Conclusion
Manual annotation is an indispensable step to create training data for developing
machine-learning automated extractors. In order to generate extractors that perform with
accuracies high enough to be acceptable to the biomedical research community,
consistently annotated training data is a prerequisite. Although we did not formally prove
it, our experience has been that investment of developing literature-based entity
definitions and annotation guidelines yields far better extracted information with distinct
conceptual boundaries, which in turn increases the opportunity for practical application.
We have concluded that rather than trying to construct unifying definitions that maximize
acceptance and minimize contention amongst domain experts, that a consistent and
generally arguable definition was preferable when making decisions to specify entity
boundaries and magnitudes. More important for us was to consider how the extracted
information will be used, and once determined, how to maintain consistency throughout
the training corpus.
39
40. Reference
Chen L, Liu H, Friedman C: Gene name ambiguity of eukaryotic nomenclatures.
Bioinformatics, 21: 248-256. (2005).
Cohen KB, Fox L, Ogren PV, Hunter L: Corpus design for biomedical natural language
processing. Proceedings of the ACL-ISMB workshop on linking biological literature,
ontologies and databases, pp. 38-45. Association for Computational Linguistics. (2005).
Collier N, Nobata C, Tsujii J: Extracting the names of genes and gene products with a
hidden Markov model. In Proceedings of the 18th International Conference on
Computational Lingustics, Saarbrucken, Germany. (2000).
GENIA: http://www-tsujii.is.s.u-tokyo.ac.jp/GENIA/ (2004).
Hanisch D, Fundel K, Mevissen HT, Ximmer R, Fluck J: ProMiner: rule-based protein
and gene entity recognition. BMC Bioinformatics. 6: S14. (2005).
Jin Y, McDonald RT, Lerman K, Mandel MA, Carroll S, Liberman MY, Pereira FC,
Winters RS, White PS: Automated recognition of malignancy mentions in biomedical
literature. BMC Bioinformatics, 7: 492. (2006).
McDonald RT, Winters RS, Mandel M, Jin Y, White PS, Pereira F: An entity tagger for
recognizing acquired genomic variations in cancer literature. Bioinformatics 22(20):
40
41. 3249-3251. (2004).
Penn BioIE: http://bioie.ldc.upenn.edu/index.jsp
Tanabe L, Wilbur W: Tagging gene and protein names in biomedical text,
Bioinformatics, 18:1124-1132. (2002).
41
42. Chapter 3. Automated Recognition of Malignancy Mentions in
Biomedical Literature
Yang Jin
Ryan T. McDonald
Kevin Lerman
Mark A. Mandel
Steven Carroll
Mark Y. Liberman
Fernando C. N. Pereira
R. Scott Winters
Peter S. White
Pulished: BMC Bioinformatics, 7:492, 2006
Abstract
Background: The rapid proliferation of biomedical text makes it increasingly
difficult for researchers to identify, synthesize, and utilize developed knowledge in their
fields of interest. Automated information extraction procedures can assist in the
acquisition and management of this knowledge. Previous efforts in biomedical text
mining have focused primarily upon named entity recognition of well-defined molecular
objects such as genes, but less work has been performed to identify disease-related
objects and concepts. Furthermore, promise has been tempered by an inability to
efficiently scale approaches in ways that minimize manual efforts and still perform with
high accuracy. Here, we have applied a machine-learning approach previously successful
for identifying molecular entities to a disease concept to determine if the underlying
probabilistic model effectively generalizes to unrelated concepts with minimal manual
intervention for model retraining.
42
43. Results: We developed a named entity recognizer (MTag), an entity tagger for
recognizing clinical descriptions of malignancy presented in text. The application uses
the machine-learning technique Conditional Random Fields with additional domain-
specific features. MTag was tested with 1,010 training and 432 evaluation documents
pertaining to cancer genomics. Overall, our experiments resulted in 0.85 precision, 0.83
recall, and 0.84 F-measure on the evaluation set. Compared with a baseline system using
string matching of text with a neoplasm term list, MTag performed with a much higher
recall rate (92.1% vs. 42.1% recall) and demonstrated the ability to learn new patterns.
Application of MTag to all MEDLINE abstracts yielded the identification of 580,002
unique and 9,153,340 overall mentions of malignancy. Significantly, addition of an
extensive lexicon of malignancy mentions as a feature set for extraction had minimal
impact in performance.
Conclusions: Together, these results suggest that the identification of disparate
biomedical entity classes in free text may be achievable with high accuracy and only
moderate additional effort for each new application domain.
Background
The biomedical literature collectively represents the acknowledged historical
perception of biological and medical concepts, including findings pertaining to disease-
related research. However, the rapid proliferation of this information makes it
increasingly difficult for researchers and clinicians to peruse, query, and synthesize it for
biomedical knowledge gain. Automated information extraction methods, which have
recently been increasingly concentrated upon biomedical text, can assist in the acquisition
and management of this data. Although text mining applications have been successful in
43
44. other domains and show promise for biomedical information extraction, issues of
scalability impose significant impediments to broad use in biomedicine. Particular
challenges for text mining include the requirement for highly specified extractors in order
to generate accuracies sufficient for users; considerable effort by highly trained computer
scientists with substantial input by biomedical domain experts to develop extractors; and
a significant body of manually annotated text—with comparable effort in generating
annotated corpora—for training machine-learning extractors. In addition, the high
number and wide diversity of biomedical entity types, along with the high complexity of
biomedical literature, makes auto-annotation of multiple biomedical entity classes a
difficult and labor-intensive task.
Most biomedical text mining efforts to date have focused upon molecular object
(entity) classes, especially the identification of gene and protein names. Automated
extractors for these tasks have improved considerably in the last few years [1-13]. We
recently extended this focus to include genomic variations [14]. Although there have
been efforts to apply automated entity recognition to the identification of phenotypic and
disease objects [15-17], these systems are broadly focused and often do not perform as
well as those utilizing more recently-evolved machine-learning techniques for such tasks
as gene/protein name recognition. Recently, Skounakis and colleagues have applied a
machine-learning algorithm to extract gene-disorder relations [18], while van Driel and
co-workers have made attempts to extract phenotypic attributes from Online Mendelian
Inheritance in Man [19]. However, more extensive work on medical entity class
recognition is necessary because it is an important prerequisite for utilizing text
information to link molecular and phenotypic observations, thus improving the
44
45. association between laboratory research and clinical applications described in the
literature.
In the current work, we explore scalability issues relating to entity extractor
generality and development time, and also determine the feasibility of efficiently
capturing disease descriptions. We first describe an algorithm for automatically
recognizing a specific disease entity class: malignant disease labels. This algorithm,
MTag, is based upon the probability model Conditional Random Fields (CRFs) that has
been shown to perform with state-of-the-art accuracy for entity extraction tasks [5, 14].
CRF extractors consider a large number of syntactic and semantic features of text
surrounding each putative mention [20, 21]. MTag was trained and evaluated on
MEDLINE abstracts and compared with a baseline vocabulary matching method. An
MTag output format that provides HTML-visualized markup of malignant mentions was
developed. Finally, we applied MTag to the entire collection of MEDLINE abstracts to
generate an annotated corpus and an extensive vocabulary of malignancy mentions.
Results
MTag performance
Manually annotated text from a corpus of 1,442 MEDLINE abstracts was used to
train and evaluate MTag. Abstracts were derived from a random sampling of two
domains: articles pertaining to the pediatric tumor neuroblastoma and articles describing
genomic alterations in a wide variety of malignancies. Two separate training experiments
were performed, either with or without the inclusion of malignancy-specific features,
which were the addition of a lexicon of malignancy mentions and a list of indicative
suffixes. In each case, MTag was tested with the same randomly selected 1,010 training
45
46. documents and then evaluated with a separate set of 432 documents pertaining to cancer
genomics. The extractor took approximately 6 hours to train on a 733 MHz PowerPC G4
with 1 GB SDRAM. Once trained, MTag can annotate a new abstract in a matter of
seconds.
For evaluation purposes, manual annotations were treated as gold-standard files
(assuming 100% annotation accuracy). We first evaluated the MTag model with all
biological feature sets included. Our experiments resulted in 0.846 precision, 0.831 recall,
and 0.838 F-measure on the evaluation set. Additionally, the two subset corpora
(neuroblastoma-specific and genome-specific) were tested separately. As expected, the
extractor performed with higher accuracy with the more narrowly defined corpus
(neuroblastoma) than with the corpus more representative for various malignancies
(genome-specific). The neuroblastoma corpus performed with 0.88 precision, 0.87 recall,
and 0.88 F-measure, while the genome-specific corpus performed with 0.77 precision,
0.69 recall, and 0.73 F-measure. These results likely reflect the increased challenge of
identifying mentions of malignancy in a document set demonstrating a more diverse
collection of mentions.
To determine the impact of the biological feature sets we included to provide domain
specificity, we excluded these feature sets to create a generic MTag. This extractor was
then trained and evaluated using the identical set of files used to train the biological
MTag version. Somewhat surprisingly, the extractor performed with similar accuracy
with the generic model, resulting in 0.851 precision, 0.818 recall, and 0.834 F-measure
on the evaluation set. These results suggested that at least for this class of entities, the
extractor performs the task of identifying malignancy mentions efficiently without the
46
47. use of a specialized lexicon.
Extraction versus string matching
We next determined performance of MTag relative to a baseline system that could be
easily employed. For the baseline system, the NCI neoplasm ontology, a term list of
5,555 malignancies, was used as a lexicon to identify malignancy mentions [22]. Lexicon
terms were individually queried against text by case-insensitive exact string matching. A
subset of 39 abstracts randomly selected from the testing set, which together contained
202 malignancy mentions, were used to compare the automated extractor and baseline
results. MTag identified 190 of the 202 mentions correctly (94.1%), while the NCI list
identified only 85 mentions (42.1%), all of which were also identified by the extractor.
We also determined the performance of string matching that instead used the set of
malignancy mentions identified in the manually curated training set annotations (1,010
documents) as a matching lexicon. This system identified 79 of 202 mentions (39.1%).
Combining the manually-derived lexicon with the NCI lexicon yielded 124 of 202
matches (61.4%).
A closer analysis of the 68 malignancy mentions missed by the string matching with
combined lists but positively identified by MTag determined two general subclasses of
additional malignant mentions. The majority of MTag-unique mentions were lexical or
modified variations of malignancies present either in the training data or in the NCI
lexicon, such as minor variations in spelling and form (e.g., “leukaemia” versus
“leukemia”), and acronyms (e.g., “AML” in place of “acute myeloid leukemia”). More
importantly, a substantial minority of mentions identified only by MTag were instances
of the extractor determining new mentions of malignancies that were, in many cases,
47
48. neither obvious nor represented in readily available lexicons. For example, “temporal
lobe benign capillary haemangioblastoma” and “parietal lobe ganglioglioma” are neither
in the NCI list or training set per se, or approximated as such by a lexical variant. This
suggests that MTag contributes a significant learning component.
Application to MEDLINE
MTag was then used to extract mentions of malignancy from all MEDLINE
abstracts through 2005. Extraction took 1,642 CPU-hours (68.4 CPU-days; 2.44 days on
our 28-CPU cluster) to process 15,433,668 documents. A total of 9,153,340 redundant
mentions and 580,002 unique mentions (ignoring case) were identified. Interestingly, the
ratio of unique new mentions identified relative to the number of abstracts analyzed was
relatively uniform, ranging from a rate of 0.183 new mentions per abstract for the first
0.1% of documents to a rate of 0.038 new mentions per abstract for the last 1% of
documents. This indicated that a substantial rate of new mentions was being maintained
throughout the extraction process.
The 25 mentions found in the greatest number of abstracts by MTag are listed in
Table 1. Six of these malignant phrases: pulmonary, fibroblasts, neoplastic, neoplasm
metastasis, extramural, and abdominal did not match our definition of malignancy. Of
these, only “extramural” is not frequently associated with malignancy descriptions and is
likely the result of containing character n-grams that are generally indicative of
malignancy mentions. The remaining five phrases are likely the result of the extractor
failing to properly define mention boundaries in certain cases (e.g., tagging “neoplasm”
rather than “brain neoplasm”), or alternatively, shared use of an otherwise indicative
character string (e.g., “opl” in “brain neoplasm” and “neoplastic”) between a true positive
48
49. and a false positive.
For comparison, we also determined the corresponding number of articles identified
both by keyword searching of PubMed and by exact string matching of MEDLINE for
each of the 19 most common true malignancy types (Table 1). Overall, MTag’s
comparative recall was 1.076 versus PubMed keyword searching and 0.814 versus string
matching. As PubMed keyword searching uses concept mapping to relate keywords to
related concepts, thus providing query expansion, the document retrieval totals derived
from this approach do not strictly compare to MTag’s approach. Furthermore, the exact
string totals would be inflated relative to the MTag totals, as for example the phrase
“myeloid leukemia” would be counted both for this category and for a category
“leukemia” with exact string matching, but would only be counted for the former phrase
by MTag. To adjust for these discrepancies, for MTag document totals listed in Table 1,
we included documents that were tagged with malignancy mentions that were both strict
syntactic parents and biological children of the phrase used. For example, we included
articles identified by MTag with the phrase “small-cell lung cancer” within the total for
the phrase “lung cancer”.
Comparison of these totals between MTag articles and PubMed keyword searching
revealed that MTag provided high recall for most malignancies. Interestingly, there are
three malignancy mention instances (“carcinoma”, “sarcoma”, “melanoma”) that have
more MTag-identified articles than for PubMed keyword searches. This suggests that a
more formalized normalization of MTag-derived mentions might assist both with
efficiency and recall if employed in concert with the manual annotation procedure
currently employed by MEDLINE. Furthermore, MTag’s document recall compared
49
50. quite favorably to exact string matching. Only two of the 25 malignancy mentions
yielded less than 60% as many articles via MTag than via PubMed exact string matching
(“bone neoplasms” and “lung cancer”). In these two cases, the concept-mapping PubMed
search identifies the articles with a broader range beyond the search terms. For example,
a PubMed search for the term “lung cancer” identifies articles describing “lung
neoplasms”, while for “bone neoplams”, articles focusing on related concepts such as
“osteoma” and “sphenoid meningioma” are identified by PubMed. Generally, MTag
recall would be expected to improve further after a subsequent normalization process that
maps equivalent phrases to a standard referent.
To assess document-level precision, we randomly selected 100 abstracts identified by
MTag each for the malignancies “breast cancer” and “adenocarcinoma”. Manual
evaluation of these abstracts showed that all of the articles were directly describing the
respective malignancies. Finally, we evaluated both the 250 most frequently mentioned
malignancies as well as a random set of 250 extracted malignancy mentions from the all-
MEDLINE-extracted set. For the frequently occurring mentions, 72.06% were considered
to be true malignancies; this set corresponds to 0.043% of all malignancy mentions. For
the random set, 78.93% were true malignancies. This suggests that such extracted
mention sets might serve as a first-pass exhaustive lexicon of malignancy mentions.
Comparison of the entire set of unique mentions with the NCI neoplasm list showed that
1,902 of the 5,555 NCI terms (34.2%) were represented in the extracted literature.
Software
50
51. MTag is platform independent, written in java, and requires java 1.4.2 or higher to
run. The software is freely available under the GNU General Public License at
http://bioie.ldc.upenn.edu/index.jsp?page=soft_tools_MalignancyTaggers.html. MTag
has been engineered to directly accept files downloaded from PubMed and formatted in
MEDLINE format as input. MTag provides output options of text or HTML file versions
of the extractor results. The text file repeats the input file with recognized malignancy
mentions appended at the end of the file. The HTML file provides markup of the original
abstract with color-highlighted malignancy mentions, as shown in Figure 1.
Discussion
We have adapted an entity extraction approach that has been shown to be successful
for recognition of molecular biological entities and have shown that it also performs with
high accuracy for disease labels. It is evident that an F-measure of 0.83 is not sufficient as
a stand-alone approach for curation tasks, such as the de novo population of databases.
However, such an approach provides highly enriched material for manual curators to
utilize further. As was determined by our comparisons with lexical string matching and
PubMed-based approaches, our extraction method demonstrated substantial improvement
and efficiency over commonly employed methods for document retrieval. Furthermore,
MTag appeared to be accurately predicting malignancy mentions by learning and
exploiting syntactic patterns encountered in the training corpus.
Analysis of mis-annotations would likely suggest additional features and/or heuristics
that could boost performance considerably. For example, anatomical and histological
descriptions were frequent among MTag false positive mentions. Incorporation of
lexicons for these entity types as negative features within the MTag model would likely
51
52. increase precision. Our training set also does not include a substantial number of
documents that do not contain mentions of malignancy; recent unpublished work from
our group suggests that inclusion of such documents significantly impacts extractor
performance in a positive manner.
Unlike the first iteration of our CRF model [14], the MTag application required only
modest computational effort (several weeks vs. several months) of retraining and
customization time (see Methods). To our surprise, the addition of biological features,
including an extensive lexicon for malignancy mentions, provided very little boost to the
recall rate. This provides evidence that our general CRF model is flexible, broadly
applicable, and if these results hold true for additional entity types, might lessen the need
for creating highly specified extractors. In addition, the need for extensive domain-
specific lexicons, which do not readily exist for many disease attributes, might be
obviated. If so, one approach to comprehensive text mining of biomedical literature might
be to employ a series of modular extractors, each of which is quickly generated and then
trained for a particular entity or relation class. Conversely, it is important to note that the
entity class of malignancy possesses a relatively discrete conceptualization relative to
certain other phenotypic and disease concepts. Further adaptation of our extractor model
for more variably described entity types, such as morphological and developmental
descriptions of neoplasms, is underway. However, the finding that biological feature
addition provided minimal gain in accuracy suggests that further improvements may be
more difficult to obtain than by merely identifying and adding additional domain-specific
features. Significantly, challenges in rapid generation of annotations for extractor
training, as well as procedures for efficient and accurate entity normalization, still
52
53. remain.
When combined with expert evaluation of output, extractors can assist with
vocabulary building for targeted entity classes. To demonstrate feasibility, we extracted
mentions of malignancy for all pre-2006 MEDLINE abstracts. Our results indicate that
MTag can generate such a vocabulary readily and with moderate computational resources
and expertise. With manual intervention, this list could be linked to the underlying
literature records and also integrated with other ontological and database resources, such
as the Gene Ontology, UMLS, caBIG, or tumor-specific databases [23-25]. Since
normalization of disease-descriptive term lists requires considerable specialized
expertise, the role of an extractor in this setting more appropriately serves as an
information harvester. However, this role is important, as such supervised lists are often
not readily available, due in part to the variability in which phenotypic and disease
descriptions can be described, and in part to the lack of nomenclature standards in many
cases.
Finally, to our knowledge, MTag is one of the first directed efforts to automatically
extract entity mentions in a disease-oriented domain with high accuracy. Therefore,
applications such as MTag could contribute to the extraction and integration of
unstructured, medically-oriented information, such as physician notes and physician-
dictated letters to patients and practitioners. Future work will include determining how
well similar extractors perform for identifying mentions of malignant attributes with
greater (e.g. tumor histology) and lesser (e.g. tumor clinical stage) semantic and syntactic
heterogeneity.
53
54. Conclusions
MTag can automatically identify and extract mentions of malignancy with high
accuracy from biomedical text. Generation of MTag required only moderate
computational expertise, development time, and domain knowledge. MTag substantially
outperformed information retrieval methods using specialized lexicons. MTag also
demonstrated the ability to assist with the generation of a literature-based vocabulary for
all neoplasm mentions, which is of benefit for data integration procedures requiring
normalization of malignancy mentions. Parallel iteration of the core algorithm used for
MTag could provide a means for more systematic annotation of unstructured text,
involving the identification of many entity types; and application to phenotypic and
medical classes of information.
Methods
Task definition
Our task was to develop an automated method that would accurately identify and
extract strings of text corresponding to a clinician’s or researcher’s reference to cancer
(malignancy). Our definition of the extent of the label “malignancy” was generally the
full noun phrase encompassing a mention of a cancer subtype, such that “neuroblastoma”,
“localized neuroblastoma”, and “primary extracranial neuroblastoma” were considered to
be distinct mentions of malignancy. Directly adjacent prepositional phrases, such as
“cancer <of the lung>”, were not allowed, as these constructions often denoted ambiguity
as to exact type. Within these confines, the task included identification of all variable
descriptions of particular malignancies, such as the forms “squamous cell carcinoma”
(histological observation) or “lung cancer” (anatomical location), both of which are
54
55. underspecified forms of “lung squamous cell carcinoma”. Our formal definition of the
semantic type “malignancy” can be found at the Penn BioIE website [26].
Corpora
In order to train and test the extractor with both depth and breadth of entity mention,
we combined two corpora for testing. The first corpus concentrated upon a specific
malignancy (neuroblastoma) and consisted of 1,000 randomly selected abstracts
identified by querying PubMed with the query terms “neuroblastoma” and “gene”. The
second corpus consisted of 600 abstracts previously selected as likely containing gene
mutation instances for genes commonly mutated in a wide variety of malignancies. These
sets were combined to create a single corpus of 1,442 abstracts, after eliminating 158
abstracts that appeared to be non-topical, had no abstract body, or were not written in
English. This set was manually annotated for tokenization, part-of-speech assignments,
and malignancy named entity recognition, the latter in strict adherence to our pre-
established entity class definition [27, 28]. Sequential dual pass annotations were
performed on all documents by experienced annotators with biomedical knowledge, and
discrepancies were resolved through forum discussions. A total of 7,303 malignancy
mentions were identified in the document set. These annotations are available in corpus
release v0.9 from our BioIE website [29].
Algorithm
Based on the manually annotated data, an automatic malignancy mention extractor
(MTag) was developed using the probability model Conditional Random Fields (CRFs)
[20]. We have previously demonstrated that this model yields state-of-the-art accuracy
for recognition of molecular named entity classes [5, 14]. CRFs model the conditional
55