Named Entity Recognition is a crucial component in bio-medical text mining.In this paper a method for disease Named Entity Recognition is proposed which utilizes sentence and token level features based on Conditional Random Field's using NCBI disease corpus. The feature set used for the experiment includes orthographic,contextual,affixes,ngrams,part of speech tags and word normalization.Using these features,our approach has achieved a maximum F-score of 94% for the training set by applying 10 fold cross validation for semantic labeling of the NCBI disease corpus. For testing and development,F-score of 88% and 85% were reported.
The increasing amount of bio-medical literature
requires more robust approaches for information
retrieval and knowledge discovery because every
single day more information is published than
humans can read. Unique challenges
specifically to bio-medical Named Entity Recognition
(NER) are caused due to its structure,since
biomedical Named Entities(NEs) consist of
symbols and abbreviations to infer relationships, thus
the length of Bio-medical NEs are not
consistent,which is the primary reason why Bio-NER
have low performance compared to general
purpose NER
For disease NER our methodology follows the traditional machine leaning approach. Figure. I depicts the work-flow of our methodology. Firstly, raw text is obtained from training, testing and development set,then pre-processing is carried out to remove characters and symbols such as underscore character, full stop etc. After pre-processing various features are extracted as described in section 2.1. The features are fed into a sequential CRF as described in section 2.2. Thus, structured output in the form of annotated named entities is obtained. This section provides details about feature extraction and classification. 2.1
Feature extraction plays a vital role in the classification accuracy of machine learning classifier as well as the NER system. The selection of relevant feature set improves the classification performance of Bio-NER. Table.1 shows the list of features used for Bio-NER and their short descriptions is listed below Word normalization attempts to reduce different forms of words, such as nouns, adjectives,verbs, etc. to their root form . For word normalization, Porter stemmer has been used to reduce disease names to its root form. Below are few examples of disease names for word root reductions obtained with the Porter stemmer algorithm.
Endometrial cancer – endometri cancer
Neurological disease – neurolog diseas
Orthographic features are related to the
orthography of the text, such as Capitalization, Digits,
Numeric, Single Caps, All Caps,numerics and
punctuation. Such features are very effective in
boundary detection
IDASH: Whether a token/word contains an inner dash such as A-T, G6PD-deficient, Palizaeus-Merzbacher disease. 2IDAH: If the number of IDASH counts equals to 2 e.g. X-linked Emery-Dreifuss muscular dystrophy, Borjeson-ForssmanLehmann, word normalization
ł Orthographic features N
wi 2; wi 1; wi; wi+1; wi+2 Posw 2; P osw 1; P osw Posw+1; P osw+2
wi 2 =wi 1; wi =wi; wi=wi+1; wi +1 =wi+2 POSw 2; P OSw 1; P OSwi; POSw+1; P OSw+2 PREFIX(wi) SUFFIX(wi) 1 ALLCAPS: Is set to true if all the alphabets in a given token are capital examples includes DMD, BMD, FD, APC, FAP and HDD etc. TITLE: If the first alphabet in a token is capitalized such as Alzheimer disease, Huntington disease, Combined genetic deficiency of C6 and C7.
LOW: All the alphabets in a given word are in lower case e.g. myotonic dystrophy, idiopathic dilated cardiomyopathy, and facial lesions.
MIXED: If a given sequence of words contains both upper and lower case such as DMD defects, hypo myelination of the PNS, deficiency of active AVP.
ALPNUM: If a given words contains both numeric and text like abnormality of CYP27, C6 deficiency, achondrogensis 1B, abnormality of CYP27.
PARN: If a multi-word contains parenthesis such as Arginine vasopressin (AVP) deficiency, palmoplanter keratoderma (PPK) conditions, sporadic (nonhereditary) ovarian cancers BRACKS: Bracket is contained within a token,example includes hypoxanthine phosphoribosyl transferees [HPRT] deficiency. GREEKS: Greek letters such as I,II,III,IV etc,is contained within a token e.g. type IIA vWD,Type II ALD,type II Gaucher disease,type II GD and type III GD SLASH: Character / is contained within the multi-word token such as cleft lip/palate, CL/P, breast and/ovarian cancers, glucose/galactose malabsorption.
N-grams are defined by a sequence of n-tokens or words. The most common n-gram is uni-gram which,contains a single token.Other n-grams examples are bi-grams and tri-grams containing 2 and 3 tokens respectively. In this experiment unigram and bi-gram have been used,in this method all the digits within a word are replaced with d e.g,the uni-gram of 33 is dd, uni-gram of nt943 is ntddd.Bigram examples are ALD/Eighteen, skin tumor/caused, APC/protein, breast or ovarian cancer/novel, etc.
POS tags are helpful in defining boundary of a
phrase,inclusion of POS has been advocated by
Contextual features refer to the word preceding and following the NEs. Contextual features are the most important features in this experiment for semantic labeling of disease names. In our experiment four contextual features are selected. Two words preceded and two followed the named entities. E.g. for the term bactracin in colon carcinoma loss cells, colon carcinoma represents the named entities while bactracin in and loss cells represent the two preceding word and two following word. 2.2
CRF is a probabilistic model used for labeling
sequential data. It is widely used for POS
tagging and NER.
Our experiment is based on National Center
for Biotechnology Information (NCBI)
disease corpus, which is freely available at NCBI
website(http://www.ncbi.nlm.nih.
gov/CBBresearch/Dogan/DISEASE/
) NCBI corpus includes 793 abstracts, which
consist of 2783 sentences and a total of 6900
disease names
ł Specific Disease
1292 2959 781
264 556 20 121
Dev set 218 409 37 127
Men- 116 Table.3 shows contribution of features and its effect on performance of CRF. The feature set is mainly divided into Contextual(Cc),Normalized(Nm),Ngrams,Affixes(Ax),Part of speech (POS) and Orthographic(O). Performance evaluation has been carried out using the metrics precision, recall and F-score. Results obtained in Table.3 is based on applying 10 Fold cross validation on the training set. Orthographic features were taken as a benchmarks, which results in F-score of 0.53. This is considered as the lowest F-score reported in this experiment. Addition of normalized features resulted in an increasing the F-score by 21%.Further addition of POS tags increased the performance by 12%.With the addition of N-gram features the overall F-score achieved is 0.91.Finally,with the addition of affixes,the final F-score obtained is 0.94.Compared to other state of the art Bio-NER systems,such as BANNER,our system has a higher level of F-score using 10 fold cross validation on training set due to the selection of good features for disease NER. For result visualization we have plotted the fscore of individual classes. In Figure.2 the Fscore of individual dataset has been plotted. In Figure.2 DC denotes Disease Class,CM denotes Composite Mention,SD denotes Specific Disease and MD denotes Modifier.Figure II shows that the highest F-score have been reported by Modifier for Training,Testing and Development set respectively,followed by Specific Disease. The lowest F-score has been shown by Composite Mentions followed by Disease Class.One reason for the relatively poor performance of the Composite Mention is the inadequate training samples compared to the training samples of Specific disease and Modifiers,which exceed 1000.The relatively poor performance of the Disease Class is because it has been based on the second smallest training sample since, the performance of machine learning based techniques heavily depends on the number of training samples. 5
This paper presents a machine learning approach for human disease NER using NCBI disease corpus. The system takes the advantage of rich feature set which,helps in representation and distinguishing of related concepts and categories.Simple features including orthographic, contextual, affixes, bigrams, part of speech and normalized tokens without exploiting features such as head nouns, dictionaries etc.The model has achieved state of the art performance for semantic labeling of named entities using the NCBI disease corpus.Each feature set represent some knowledge about the named entity and hence, in order to evaluate the overall benefit for each feature, all possible combinations of feature additions need to be considered.