Seite - 131 - in Document Image Processing

J. Imaging 2018,4, 15 presented inFigures1and3. It is composedof853pages thatwereautomaticallydividedinto lines, giving a total number of 20,356 lines. In the standard trainingpartition, the vocabulary size is of about11,000wordswithasetof106characters (the105different characters thatappear in the text of the trainingpartitionandoneextracharacter thatappears in the textof thevalidationpartition), including10numbers, 72upperandlowercase letterswithandwithoutaccents, 5punctuationmarks, 1blankspaceand18special symbols. Thefirst15,010 linesarepubliclyavailableon thewebsiteof the PatternRecognitionandHumanLanguageTechnology(PRHLT)researchcenter [20]. In thiswork, weusedthispubliclyavailablepartition. Thefirst9000 lineswereusedfor training theopticaland languagemodels, thenext1000 forvalidationandthe last5010 lines for testing. Figure3.Page515of theRodrigodataset. In theRodirgocorpus, therearemanyrarewordsandwords in theirarchaic formsyieldinga large amountofOOVwords.Moreover, thiscorpuscontainsscarceOOVcharacters(suchas:\, p´, g¯, andw) thatdonotbelongto the trainingset.OOVwordsgenerally includewords thatappear indistinct form inthetrainingandtest sets (e.g.,portugalandportug¯l), abbreviationsandwordshyphenateddifferently in the trainingandtest sets. Table1presentsasummaryof the informationcontainedin thepartitionsof theRodrigocorpus used in thiswork at the three lexical units studied: words, sub-words and characters. This table presents for each lexicalunit the total amount, thevocabulary size (differentunits), theamountof OOVunitsandtheoverlappingbetweentheOOVcontainedin thevalidationandtestpartitions, i.e., theamountofOOVunitscontainedin the testpartition thatarepresent in thevalidationpartition. 131