• No results found

The Term Generation Method was developed to meet requirements described above. Together with definition extraction and taxonomy induction the collection of tools and applications will in the following be referred to as Dresden Ontology Generator for Directed Acyclic Graphs (DOG4DAG). A basic assumption underlying the term generation pipeline is the one on the morphology and importance of terminology. It has been assumed that terms are statistically relevant noun phrases, which are the units of text possibly being subject or object in a sentence. The judgement on the importance of such phrases is made according to the occurrence statistics of a phrase in the selected document or document set (the one from which the terms are extracted) in relation to the occurrence statistics of the phrase in a domain specific or general reference corpus.

Method summary

We extract terms from English text, which we tokenize before POS-tagging, sen- tence identification, noun phrase and local abbreviation detection. As POS tagger

we use Ling-Pipe-Tagger 1trained on MEDLINE and the TNT tagger (Brants, 2000)

trained on the Wall Street Journal corpus. We generally regard phrases with pattern

[adj|verb] ∗ [f ill]{2}[noun]+ as noun phrases, where f ill are fill words like of, the, for, etc.

We first great for each noun phrase a candidate concept. In DOG4DAG, all lex- ically overlapping concepts are also grouped. Two concepts overlap, if any two of their lexical representations are similar. They are regarded as similar if they show a Hamming distance of less than 20% of the length of the shorter term label. The Ham- ming distance between two strings denotes the number of position the two strings differ from each other. We align strings from the beginning and include the length of overlapping tails in the distance. This grouping is not performed if concepts only have common abbreviations. Nested terms, i.e. noun phrases within longer noun phrases, are expanded as separate candidate concepts. DOG4DAG not only retrieves the maximal length terms, but also nested terms that follow noun phrase pattern specified above.

We rank candidate concepts, under consideration of all lexical variants and abbre- viations, according to their relative importance by the tf-idf (term frequency-inverse document frequency) measure, a weighting method commonly used in information retrieval. It captures the importance of a term in a set of documents in relation to a corpus. As corpus we used all scientific abstracts listed in PubMed.

The method with all modules (taggers and revisions) is illustrated in Figure 3.1. As representation for text a tree structure is used where each node represents a non- overlapping sub-string of the text. Nodes correspond to tagged regions, e.g. tokens, sentences. Nodes can hold several types of tags, e.g. tokens will have Part-Of-Speech tags if they represent a word (Section 6.1).

In the following the steps of the term recognition pipeline (Figure 3.1) will be shortly described in the same order they are applied:

The term generation pipeline

1. Prepare text: The initial step in the term recognition pipeline is the extraction of text. As input for the term generation the text from PDF documents, the snip- pets from Yahoo search results, and title and abstract from PubMed documents is used. The texts are converted in the TextTree representation (Section 6.1), which is the data structure that can be efficiently manipulated by the various

ElivagarTagger(Section 6.2) implementations.

2. Tokenization: The tokenizer separates text in lexical units, such as word to- kens, white spaces, (opening/closing) brackets, punctuations, (opening/closing) quotes. Each of those will be marked up.

3. Part-of-Speech tagging: Part-Of-Speech denotes the grammatical classification or the category a word can be assigned to in the context of a phrase, sentence or paragraph. This categories contain classes like noun, adjective, adverb, verbal. For each word token the Part-Of-Speech category gets identified and marked up. 4. Sentences segmentation: Sentence borders are identified to structured text as

series of sentences.

5. Noun phrase tagging: Depending on the used Part-Of-Speech tagger noun phrases are being identified and marked-up following defined consecutive pat- terns of Part-Of-Speech tags.

6. Abbreviation tagging: Local and common abbreviations are being identified and marked up. To find local abbreviations we adapted an implementation from Schwartz and Hearst (2003) to find the best long form.

7. Local frequency: The frequency of occurrence is obtained for each noun, noun phrase and abbreviation.

8. Concept extraction: Concept representations are created for the nouns and noun phrases obtained from text.

9. Syntactic filtering: Concepts are filtered to exclude known false predictions from the result set. Those include document type specific vocabulary, as “Introduc- tion”, “Methods”, “Conclusion”, etc., but also wrong classification of words as nouns by the Part-Of-Speech tagger.

10. Syntactic grouping: Concepts are grouped together if they share lexical repre- sentations.

11. Nested concepts: Concepts get derived from composed noun phrases. Each sin- gle noun phrase is added as candidate term.

12. Global frequency: From big corpora, i.e. all approximately 18M PubMed ab- stracts, the frequency of occurrence is obtained for each noun, noun phrase and abbreviation associated with a concept.

13. Probability revisions (PVALUE): Given the local and global frequency values the probability of a concept occurring in the local document set is calculate and assigned to a concept. A hypergeometric distribution of random variable was assumed (for details see Section 3.5.3 (Hypergeometric distribution)).

14. TFIDF revision: Given the local and global frequency values the tf-idf value is calculated and assigned to a concept.

15. Sorting candidate concepts: Candidate concepts are ranked by the scores ob- tained from the probability (PVALUE) or TFIDF revisions.

Technical implementation

The whole term recognition pipeline was implemented using the programming lan- guage JAVA and was encapsulated in an Axis2 generated web service allowing syn- chronous and asynchronous requests. This allows the seamless integration into other application as done for the web-based Term Generation Platform (Section 7.5), the OBO-Edit Ontology Ontology Generation Plug-in (Section 7.2), the Protégé Term Generation Plug-in (Section 7.3), and the Go3R Ontology Editor (Section 7.4). The generation of terms based on 250 PubMed abstracts takes approximately 1-2 sec- onds.

As source for PubMed abstracts the GoPubMed internal search API was used. The web service is running on an standard application server with 4 cores and 4GB of main memory together with the services for Definition Generation and Ontol- ogy Look-up. The algorithms are modularised and share the not definition specific components, like tokenization, noun phrase extraction, DOM like data structures, abbreviation detection etc. with the Definition Generation module. All system com- ponents are managed using MAVEN 2.0, a software project management and com- prehension tool. The module was developed using the Open Source IDE Eclipse.

Applications

The method has been integrated in several applications, among other the two most used ontology editors OBO-Edit (Section 7.2) and Protégé (Section 7.3). The other tools are the Go3R Ontology Editor (Section 7.4) and the web-based term generation platform (Section 7.5).

Fig. 3.2. Illustration of the endosome biogenesis(from Jean Gruenberg & Harald Stenmark. The biogenesis of multivesicular endosomes. Nature Reviews Molecular Cell Biology 5, 317-323, April 2004) Endocytosis is the process by which cells incorporate materials from the outside of the cell. The early endosomes located at the cell membrane receives vesicles coming in the cell and releases endosomal carrier vesicles (ECVs) or multivesicular bodies (MVB) which are further received by the late endosome. The late endosomes pass the material on to lysosomes.