Concept retrieval-based semantic annotation Performing the semantic annotation of a document D consists of nding mappings between text chunks of D (i.e., sequences of adjacent terms), and the concepts provided by a knowledge resource (KR) that best semantically describes the contents of D. As concepts of a KR are usually expressed as noun phrases, text chunks are usually associated to these syntactic structures. We assume that 4 http://www.ebi.ac.uk/Rebholz-srv/CALBC/
This paper presents a study about the use of the concept retrieval annotation
method for parallel corpora. The concept retrieval annotation method (CRA)
consists of considering concepts as documents and text chunks as queries [
In this work, we will apply CRA in a bi-lingual scenario. For this purpose, we make use of the provided lexicons at CLEFER workshop. More speci cally, we have made use of the English and Spanish lexicons. In this extended abstract, we rst summarize the main features of CRM as a cross-lingual annotator, and then obtained results over the two provided parallel corpora, EMEA and MEDLINE R . there exists a function lexKR(C) that returns the set of strings describing the concept C. This set of strings can contain di erent lexical variants of C, synonyms of these variants, and a short de nition of the concept.
Given a text chunk T = (w1 wn), and the concept C of the KR, the
retrieval score of C w.r.t. to T is calculated as follows [
The function inf o(s) provides the information the string s brings, which is calculated as inf o(s) = Pw2s log(p(wjBackground)).
The retrieval of candidate concepts is e ciently performed by using an inverted le where each entry is a vocabulary word, and the hit list contains the concept strings containing the word. In this way, the text chunk T is executed as a query over this inverted le, and the retrieved concept strings are ranked according to sim(T; C). Finally, the top-ranked concepts that best cover the words in T are included in the semantic annotation of T .
We propose a simple strategy to annotate texts given a multi-lingual KR: { Build a di erent inverted le for each language supported by the KR. { De ne a series of simple lexical rules to generate variants from one language to the other (e.g., prote na ! protein). { Fetch the query to each inverted le with the variants corresponding to its language. { Return the set of all concepts retrieved by each lexicon.
In the multi-lingual scenario we also have to estimate the word probabilities
in large text collections for each language. Fortunately, there exist several
publicly available resources providing such word estimations 5. We have performed
Word Sense Disambiguation (WSD) based on the MRD (Machine Readable
Dictionary) method [
Table 1 shows the main features of the annotated collections. Annotated collections provided at CLEFER are indicated with SSC (Silver Standard Corpus), and they are in English. Annotations are calculated as the number of text chunks having associated some concept. The average size of an annotation is the average of the number of words of annotated text chunks. We also measure the percentage of ambiguous annotations, which are those having more than one entity type associated to the text chunk. In general, English collections generate more ambiguous annotations than the Spanish ones. However, this result is 5 http://invokeit.wordpress.com/frequency-word-lists/
Annotated Collection Documents/Units Annotations Ann. Avg. size Ambiguity mainly due to the higher noise of the English lexicon. We noticed that many ambiguous annotations were derived from contextual descriptions. To alleviate this problem, we developed a simple heuristic to detect these false ambiguities, and accordingly edited the lexicons. Thus, collections EMEA EN/ES (Ed) and MEDLINE EN/ES (Ed) have been annotated using the edited lexicons. Notice that for the EMEA English collection, this heuristic has reduced notably the ambiguity degree of the annotations (compare the second and third rows of the table).
Table 2 shows the concept overlap at both collection and the aligned unit levels. As it can be noticed, the edition of the lexicon increases the overlap between the collections. This is because wrongly annotated concepts are unlikely to appear in the parallel collection. It must be also noticed that overlap at collection level is much higher than at unit level.
Collection
Collection level Unit level
This work has been partially funded by the Spanish National R&D Programme project with contract number TIN2011-24147 of the \Ministerio de Econom a y Competitividad", and the EU STREP project grant 296410 (\Mantra") under the 7th EU Framework Programme within Theme \Information Content Technologies, Technologies for Digital Content and Languages" [FP7-ICT-2011-4.1]. It has been also supported by Australian Federal and Victoria State Governments and the Australian Research Council through the ICT Centre of Excellence program, National ICT Australia (NICTA).