Developing An Unsupervised Grammar Checker For Filipino .

PACLIC 30 Proceedingsword form that changes the word’s part-ofspeech category; and compounding – ‘whereindependent words are concatenated in some wayto form a new word’ (Bonus, 2003). See Table 1for some of the different forms of the root wordkain nankinakainanTranslationwill just eatjust eatfeedwill feedeat (something)ate (something)eating (something)(somebody) eatingeating/dinner tableeating placeeating place (wheredo-er will go later)eating place (wheredo-er is right now)foodpagkainAdjectivepalakainloves eatingTable 1: Different forms of kain ‘eat’There are also affixes that are separated by ahyphen (-) from its root word or morpheme (ex.mang-akit ‘to entice’ from the root akit ‘entice’).There are also cases wherein addition or insertionof an affix to a word could alter the spelling of itsbase form (ex. The prefix pang- palit ‘change’ pamalit ‘item for changing’). However, not allaffixes and reduplication can be applied to anyword. For instance, the root word luto ‘cook’ canuse ‘nag-‘ as prefix but kain ‘eat’ cannot. Itshould also be noted that there are assimilatedwords from English in Filipino wherein affixesare also appended to it (ex. magce-cellphone‘will use a cellphone’, i-file ‘to file (adocument)’). The Filipino language also has itsown set of compound words. There are two waysto combine words together, either with the use ofa hyphen (ex. halo-halo ‘(a type of Filipinodessert)’ from the word halo ‘mix’, and kisapmata ‘instant’ from the words kisap ‘blink’ &mata ‘eye’) or just combining them as is (ex.kapitbahay ‘neighbor’ from the words kapit‘hold onto’ & bahay ‘house’, and hanapbuhay‘livelihood’ from the words hanap ‘find’ &buhay ‘life’) (Paz, 2003).Another important linguistic phenomenon inFilipino is the rule: “Kung ano ang bigkas,siyang sulat” ‘Spell as you pronounce it’(Ortograpiyang Pambansa, 2013). As the rulestates, the words in Filipino are usually spelled asthey are pronounced with some exceptions. Thisphenomenon simplifies the way Filipino wordsare spelled out (ex. Filipinized form of‘computer’ as kompyuter) but also causes somespelling confusion which will be discussed in thenext section.3. Error TypesIn understanding the error types that exist inFilipino writing, three references were used: TheCambridge Learner Corpus (Nicholls, 1999),Wikapedia (2015), and a parallel corpus of 1252erroneous-and-correct word and phrase pairsfrom sentences written by Filipino universitystudents.The Cambridge Learner Corpus contains 16million words from English examination scriptsby learners of English containing different typesof errors. The corpus categorized the error typesinto general and specific errors. The proponentsnoticed that some error categories would have itsFilipino counterpart such as wrong form used,missing word/phrase, word/phrase needsreplacing, unnecessary word/phrase, punctuationerrors, countability errors, determiner agreement,incorrect verb inflection, spelling errors, andother error categories also exist in Filipino.Wikapedia (2015) is a booklet created bythe Presidential Communications Developmentand Strategic Planning Office of the Philippinescontaining correct usage of affixes, words, andphrases in Filipino which people may findconfusing. One example described in the bookwould be the use of ng, a function word definingpossession (ex. aso ng kapitbahay ‘dog ofneighbor’) and in a direct object phrase (ex.kumain ng mansanas ‘ate an apple’) vs the use ofnang which is commonly used before an adverb(ex. kumain nang mabilis ‘ate fast’). The usageof these two words is confusing because it ispronounced almost exactly the same. Otherexamples contained in the booklet are properusage of affixes and words, morphophonemics,usage of hyphens and spaces, and others.After analyzing the parallel corpus of 1252erroneous-correct word/phrase pairs, it is foundthat majority of the errors fall under spellingerrors, incorrect usage of affixes/reduplicationwhich is mostly caused by usage of hyphens andspaces, and wrong word usage.107

It is observed that one reason the studentsmade spelling errors is because of the way aword is pronounced which is usually simplifiedfor conversational use. Some of these simplifiedwords, see Table 2, are still not accepted informal Filipino writing which cause spellingerrors. Another cause of spelling errors is theconfusion whether to spell an English borrowedword in its English version or convert it to itsFilipinized spelling version.There were many instances of affix errorswhere the students were confused whether aword is an affix of a word, a separate word, or ifthere should be a hyphen between the affix andthe root word. A few of the affix errors also showthe confusion of students in selecting anappropriate affix of a verb when used for acertain focus and/or aspect. See Table 3.The students also committed severalmistakes in identifying which word to use incertain situations which is caused byunfamiliarity with Filipino syntax rules. SeeTable 4.Other errors that exist in the parallel corpusinclude the lack of space between words (ex. parin ‘still’ incorrectly written as parin), compoundwords that was separated by a space (ex. arawaraw ‘everyday’ incorrectly written as arawaraw) and punctuation errors where somecommas or periods are have’anong ontingnan ‘look’ ��researcher’Table 2: Spelling ErrorsCorrectMisspelledReasonWordasPangkainPang kainExtra Space‘usedforeating’TagtuyoTag-tuyoExtra Hyphen‘drought’IkawaloIka-waloExtra Hyphen‘eighth’i-predict ‘to ipredictMissingpredict’Hyphenmasmalaki masmalakiMissing Space‘bigger’inilagaysa nilagaysa Incorrectkahon ‘placed kahonAffix used forin a box’a verb focusTable 3: Affix ErrorsConfused between:ng ‘of’may ‘has (used beforenouns, verbs,adjectives andadverbs)’nang ‘(function wordbefore an adverb)’mayroon ‘has (usedbefore grammaticalparticles, personalpronouns, and adverbsof place’na ‘(type ofgrammatical particle)’suffix –ng ‘used inplace of na if wordpreceding it ends in avowelTable 4: Wrong Word Usage4. Overview of the Grammar CheckerThe grammar checker named Gramatika that isdiscussed in this paper utilizes the existingimplementation of the Lexbar application byTsao & Wible (2009) and extends it to covermore error types, some of which are unique inthe Filipino language. It uses n-grams as rules,commonly referred to as hybrid n-grams, fromgrammatically correct texts consisting of words,POS tags, and lemmas to detect grammaticalerrors and provide suggestions containingpossible corrections. The production of POStags, and lemmas can be produced by existing6POS taggers and morphological analyzers forFilipino making the system unsupervised suchthat new grammatically correct texts can be fedthrough these systems and to Gramatika to easilyincrease the number of grammar rules.6See Rabo & Cheng (2006) and Bonus (2003)108

PACLIC 30 Proceedings4.1 Rules LearningEven though Gramatika also uses hybrid n-gramssimilar to Lexbar’s (Tsao & Wible, 2009) andslightly similar to EdIt’s (Huang et al., 2011), theapproach in deriving the hybrid n-grams isdifferent. Gramatika uses a clustering approachas opposed to Lexbar’s pruning and EdIt’scollocations-based approaches. The n-gram sizesused as rules range from 2 to 7. For example,given an incorrect phrase para sa bata anglaruan ni iyon. ‘?that? toy is for the kid’, ifGramatika has the hybrid 7-gram ‘para sa[NNC] [DTC] [NNC] na [PRO].’ 7, then it canimmediately suggest to change the word ni ‘(agrammatical particle used before a personalproper noun)’ to na ‘(a grammatical particle usedaround adjectives, pointing pronouns, andothers)’ which produces the corrected version:para sa bata ang laruan na iyon ‘that toy is forthe kid’ which is a more appropriate suggestionthan the suggestion produced by the trigram[NNC] ni [NNP] 8 to change iyon to a propernoun (ex. Mark)producing the correctedversion: para sa bata ang laruan ni Mark‘Mark’s toy is for the kid’. The use of larger ngram sizes increases the context from which asuggestion can be based from.In the clustering approach, all n-gramsequences are retrieved from grammaticallycorrect texts and are stored in the database.During the storing process, the frequency of allPOS tag sequences is counted. POS tag sequencesexceeding the threshold of 2 are retrieved and theword n-grams are grouped as clusters. For eachn-gram clusters, the module checks if there areany token slot that can be generalized to POSlevel. For example, if a cluster has the instancesnagpunta sa bayan ‘went to the town’ andbumisita sa bahay ‘visited the house’, the firstand third tokens can be generalized because itmeets the minimum difference threshold of 2.This produces the hybrid n-gram [VBTS] sa[NNC] which can be used to flag the phraseumupo sa silya ‘sat on the chair’ asgrammatically correct or used to detectgrammatical errors. The n-gram rules are storedin the database as sequences of words, POS tags,lemmas, and a Boolean sequence denoting whichtoken slots are generalized. This is done to allowGramatika to provide word-specific suggestions7Based from the Rabo & Cheng (2006) tag set, NNC common noun, DTC determiner for common nouns, PRO pronoun pointing to an object8NNP proper nounand to also identify the appropriate transformedword to a specific POS -lemma mapping.4.2 Error DetectionIn detecting grammatical errors and producingsuggestions based on the hybrid n-grams, aweighted Levenshtein edit distance algorithm isused. This algorithm is commonly used in spellchecking to compute how many edits it will taketo convert a potentially misspelled word to acorrect word in the dictionary. It has also beenused by EdIt (Huang et al., 2011) in providingcorrections by substitution, insertion, anddeletion. In Gramatika, the edit distancealgorithm is extended to detect errors andprovide suggestions correctable by ,unmerging, and merging. The error types thatexists in Filipino are grouped based on the sixsuggestion types, see Table 5.CorrectionSubstitutionError TypesAffix/Formerrors,wrongword/punctuation usage (includespreposition, determiners, andothers)SpellingMisspelled words, misuse/lack ofCorrection hyphensInsertionMissing words and sUnmerging Incorrectlymergedwordsrequiring unmerging of words orremoval of hyphensMergingIncorrectlyunmergedwordrequiring removal of space orinsertion of hyphen between textsTable 5: Correction and Error TypesIn producing suggestions, Gramatika parsesthe input, which is POS and lemma-tagged, inton-grams starting from size 7 down to 2. For eachinput n-gram, it retrieves hybrid n-gram rules“similar” to the input n-gram from the database.A rule is considered “similar” to an input n-gramif at least n–2 POS tokens of it are equal to thePOS tokens in the input n-gram. Three sizes ofthe rules are also retrieved for each input n-gram:rules that are of equal size to the input n-gram tobe used for substitution and spelling correctionsuggestions, rules that are one token size largerto produce insertion and

Filipino is the rule: “Kung ano ang bigkas, siyang sulat” ‘Spell as you pronounce it’ (Ortograpiyang Pambansa, 2013).As the rule states, the words in Filipino are usually spelled as they are pronounced with some exceptions.This phenomenon simplifies the way Filipi