Text Modification for Bulgarian Users Slavina Lozanova∗ Ivelina Stoyanova† Svetlozara Leseva† [email protected] [email protected] [email protected] Svetla Koeva† Boian Savtchev‡ [email protected] [email protected] ∗ AssistNet, Sofia, Bulgaria † Department of Computational Linguistics, IBL, BAS, Sofia, Bulgaria ‡ Cambridge Language Assessment Centre BG015, Sofia, Bulgaria

Abstract cation devices (Paul, 1998; Conrad, 1979; Mussel- man, 2000; Traxler, 2000; Vermeulen AM, 2007), The paper discusses the main issues re- which shows that reading skills are influenced garding the reading skills and compre- by complex social, linguistic and communication- hension proficiency in written Bulgarian related factors rather than by the sensory disability of people with communication difficulties, alone. and deaf people, in particular. We consider The paper explores reading comprehension of several key components of text compre- Deaf people1 who use Bulgarian Sign Language hension which pose a challenge for deaf (BulSL) as their primary language. Various re- readers and propose a rule-based system search studies both in Bulgaria and abroad have for automatic modification of Bulgarian shown that hearing-impaired BulSL users have texts intended to facilitate comprehension poorer reading skills than their hearing peers. Var- by deaf people, to assist education, etc. In ious methods for text modification have been ex- order to demonstrate the benefits of such a plored to the end of obtaining texts that corre- system and to evaluate its performance, we spond to the proficiency of the readers. Most of have carried out a study among a group of the modification methodologies have been focused deaf people who use Bulgarian Sign Lan- on simplifying the original texts and decreasing guage (BulSL) as their primary language their complexity (Inui et al., 2003). Our approach, (primary BulSL users), which compares however, focuses not on simplification, but on the the comprehensibility of original texts and adaptation of the structure of the original texts to their modified versions. The results shows the linguistic properties of BulSL. a considerable improvement in readability The paper is organized as follows. Section 2 when using modified texts, but at the same discusses the reading skills of BulSL users, paying time demonstrates that the level of com- attention to children’s and adult education in Bul- prehension is still low, and that a complex garia focused on the acquisition of Bulgarian and set of modifications will have to be imple- the relationship between BulSL and verbal Bulgar- mented to attain satisfactory results. ian. After outlining the main principles which un- 1 Introduction derlie text adaptation aimed at fostering text com- prehensibility in the target population, we present The individual development of deaf people de- a rule-based method for automatic modification of pends on a complex of factors, which include the Bulgarian written texts. The method applies a set cause and the degree of hearing loss, the age of of linguistic transformations and produces modi- hearing loss onset, educational background, lan- fied versions of the texts, which are better suited guage and communication methods, cultural iden- to the needs of BulSL users (Section 3). Section tification, disability preconceptions. Hearing loss 4 describes an experiment devised to explore the leads to a limited spoken language input, delays reading comprehension of BulSL users of original in language acquisition and communication dif- and modified texts. Section 5 draws conclusions ficulties. Deaf children and adults demonstrate lower reading achievements than hearing people 1Capitalized ’Deaf’ is used to denote the community of deaf people who use Sign Language as their primary lan- regardless of the degree of hearing loss, and the guage. The term emphasizes the socio-cultural model of use (or lack) of high-performing hearing amplifi- Deafness rather than the medical view of hearing impairment.

39 Proceedings of the 2nd Workshop on Predicting and Improving Text Readability for Target Reader Populations, pages 39–48, Sofia, Bulgaria, August 4-9 2013. c 2013 Association for Computational Linguistics and outlines some directions for future work. provide the basis for developing advanced meth- ods for automatic text modification directed to im- 2 Reading Skills of Hearing-Impaired proving text readability for deaf BulSL users. People 2.2 Sign Language and Verbal Language 2.1 Education Research has shown that Deaf children of Deaf Previous research has shown that deaf students lag parents (DCDP) with sign language as their pri- behind their hearing peers in reading comprehen- mary mode of communication outperform their sion, because they experience difficulties with vo- deaf peers of hearing parents (DCHP) on differ- cabulary (Paul, 1996), syntax (Kelly, 1996), and ent academic tests, including reading tests (May- the use of prior knowledge and metacognition berry, 2000). Several studies have found a positive (Trezek et al., 2010). In addition, reading com- correlation between the advanced American Sign prehension difficulties are often linked to lack of Language (ASL) skills of deaf students and their general knowledge due to inadequate education higher reading skills (Hoffmeister, 2000; Padden and limited access to information (Lewis and Jack- and Ramsey, 2000). Evidence is not conclusive as son, 2001). Two independently performed studies to how sign languages relate to verbal languages (Albertini and Mayer, 2011; Parault and William, and what influence they have on the acquisition 2010) have found out that deaf college students’ of general communication skills and knowledge reading skills are below those of six-graders2. about the world. MacAnally et al. (1999) support the hypothe- The extensive research on sign languages in the sis that using less complicated and more accessi- last fifty years worldwide has shown that they are ble reading materials, consisting of language con- independent linguistic systems which differ from structions close or similar to sign language struc- verbal languages (Stokoe, 1960; Stokoe, 1972; ture can facilitate reading comprehension and mo- Sutton-Spence and Woll, 2003). Being an inde- tivate deaf people to read. In support of this claim pendent language, a sign language affects the way Berent (2004) points out that deaf students would in which its users conceptualize the world, accord- read more smoothly if the subject, verb, and object ing to the principle of linguistic relativity, first for- are in a simple SVO (subject-verb-object) word mulated by Sapir and Whorf (Lee, 1996). Due order. These studies provide evidence in favour to the fact that sign languages are very different of text adaptation that reflects features of the sign from verbal ones, many Deaf people attain a cer- language and the development of modified teach- tain level of proficiency in a verbal language at the ing materials. state of interlanguage4 (Selinker, 1972) but that Bulgarian education for the deaf is based en- level is not sufficient to ensure successful social tirely on the oral approach and no systematic effort integration. has been invested into exploring total communi- 2.3 Readability of Written Texts for Native cation and bilingual approaches (Lozanova, 2002; Users of Sign Language Saeva, 2010). Even in the specialized schools for the deaf, where sign language communication Readability is measured mainly on the basis of vo- occurs naturally, BulSL has not been integrated cabulary and sentence complexity, including word into the school curriculum. Besides, the linguis- length and sentence length: the higher the letter, tic analysis of BulSL has been limited to mere syllable and word count of linguistic units, the descriptions and presentation of signs: Bulgar- greater the demand on the reader. Some syntactic ian Sign Language dictionaries (1966, 1996); Sign structures also affect readability – negative and in- Language Dictionary in Civil Education (Stoy- terrogative constructions, passive voice, complex anova et al., 2003); Specialized Multimedia BulSL sentences with various relations between the main dictionary3 (2005). clause and the subordinates, long distance depen- In order to improve education and the reading dencies, etc. Besides, readability improves if the and communication skills of deaf people, a com- information in a text is well-organized and effec- prehensive study of BulSL is necessary, that will 4The term ’interlanguage’ denotes the intermediate state in second language acquisition characterized by insufficient 211-12-year-olds. understanding and grammatical and lexical errors in language 3http://www.signlanguage-bg.com production.

40 tively presented so that its local and global dis- close in meaning to the original. course structure is obvious to the reader (Swann, 1992). • Modified text should be grammatically cor- Text modification is often understood as simpli- rect and structurally authentic by preserving fication of text structure but this may result in an as much as possible of the original textual and inadequately low level of complexity and loss of syntactic structure. relevant information. Moreover, using a limited • In general, modified text should be character- vocabulary, avoiding certain syntactic structures, ized by less syntactic complexity compared such as complex sentences, is detrimental to the with the original text. However, the purpose communication and learning skills. of the modification is not to simplify the text The efforts towards providing equal access to but rather to make the information in it more information for Deaf people lack clear principles accessible and understandable by represent- and uniformity. Firstly, there is no system of cri- ing it in relatively short information chunks teria for evaluation of text complexity in terms of with simple syntax without ellipses. vocabulary, syntactic structure, stylistics and prag- matics. Further, no standard framework and re- • It should be possible to extend the range quirements for text modification have been estab- of modifications and include other compo- lished, which limits its applications. nents which contribute to readability or intro- duce other functionalities that facilitate read- 3 Text Modification of Bulgarian ing comprehension, such as visual represen- Language modification for improved readability is tations. not a new task and its positive and negative aspects 3.2 Stages of Text Modification have been extensively discussed (BATOD, 2006). At present we apply a limited number of modifica- One of the most important arguments against text tions: clause splitting, simplification of syntactic modification is that it requires a lot of resources in structure of complex sentences, anaphora resolu- terms of human effort and time. An appealing al- tion, subject recovery, clause reordering and inser- ternative is to employ NLP methods that will facil- tion of additional phrases. itate the implementation of automatic modification for improved readability of written texts aimed at 3.2.1 Preprocessing the BulSL community. The preprocessing stage includes annotation with the minimum of grammatical information nec- 3.1 General Principles of Text Modification essary for the application of the modification Several studies have observed different aspects of rules. The texts are sentence-split, tokenized, text modification: splitting chosen sentences with POS-tagged and lemmatized using the Bulgarian existing tools (Petersen and Ostendorf, 2007), Language Processing Chain5 (Koeva and Genov, ’translating’ from complex to simplified sentences 2011). Subsequently, clause splitting is applied with statistical machine translation methods (Spe- using a general method based on POS tagging, cia, 2010), developing text simplification systems lists of clause delimiters – clause linking words (Candido et al., 2009), etc. (Siddharthan, 2011) and multiword expressions and punctuation, and a compares a rule-based and a general purpose gen- set of language specific rules. erator approaches to text adaptation. Recently the We define a clause as a sequence of words be- availability of the Simple English Wikipedia has tween two clause delimiters where exactly one fi- provided the opportunity to use purely data-driven nite verb occurs. A finite verb is either: (a) a sin- approaches (Zhu et al., 2010). The main oper- gle finite verb, e.g. yade (eats); (b) or a finite verb ation types both in statistical and in rule-based phrase formed by an auxiliary and a full verb, e.g. approaches are: change, delete, insert, and split shteshe da yade (would eat); or (c) a finite copu- (Bott and Saggion, 2011). lar verb phrase with a non-verbal subject comple- Although text modification is a highly language ment, e.g. byaha veseli (were merry). dependent task, it observes certain general princi- We identify finite verbs by means of a set of ples: rules applied within a window, currently set up to • Modified text should be identical or very 5http://dcl.bas.bg/services/

41 two words to the left or to the right: stack, and the CC is not identified yet, we look Rule P1. A single finite verb is recognized by the for a coordinating clause linking word or phrase POS tagger. (Some smoothing rules are applied to which marks a coordinate CO. detect the verb forms actually used in the context Rule C5. If a coordinate CO is on the top of – e.g. forms with reflexive and negative particles). the stack, we look for another coordinating clause Rule P2. If auxiliaries and a lexical verb form linking word or phrase which marks a coordinate occur within the established window, they form a CO. single finite verb phrase. (This rule subsumes a Rule C6. If a coordinate CO is on the top of number of more specific rules that govern the for- the stack and no coordinate CO is found, we look mation of analytical forms of lexical verbs by at- for a punctuation clause delimiter (e.g. a comma) taching auxiliary verbs and particles.) which functions as a CC element. Rule P3. If an auxiliary (or a copular verb) but not Rule C7. If no clause boundary has been identi- a lexical verb form occurs within the established fied between two finite verbs, we insert a clause window, the auxiliary or copula itself is a single boundary before the second finite verb. finite verb. Rule C8. All COs from the stack should have a Rule P4. If a modal and/or a phase verb and a lex- corresponding CC. ical verb form occur within the established win- Rule C9. The part of the sentence to the right of dow, they form a single finite verb phrase. the last finite verb until the end of the sentence Rule P5. If a modal (and/or a phase) verb but should contain the CCs for all COs still in the not a lexical verb form occurs within the estab- stack. lished window, the modal verb itself is a single fi- nite verb. 3.2.2 Empty subject recovery A clause is labeled by a clause opening (CO) The detection, resolution, and assignment of func- at the beginning and a clause closing (CC) at the tion tags to empty sentence constituents have be- end. We assume that at least one clause boundary come subject of interest in relation to parsing – an opening and/or a close – occurs between any (Johnson, 2002; Ryan Gabbard and Marcus, 2004; pair of successive finite verbs in a sentence. Each Dienes and Dubey, 2003), in machine translation, CO is paired with a CC, even if it might not be information extraction, automatic summarization expressed by an overt element. (Mitkov, 1999), etc. The inventory of empty cate- gories includes null pronouns, traces of extracted We distinguish two types of COs with respect to syntactic constituents, empty relative pronouns, the type of clause they introduce: coordinate and etc. So far, we have limited our work to subject subordinate. Most of the coordinating conjunc- recovery. tions in Bulgarian are ambiguous since they can A common feature of many, if not all, sign lan- link not only clauses, but also words and phrases. guages (BulSL among others) is that each sentence On the contrary, most of the subordinating con- requires an overt subject. Moreover, each subject junctions, to the exception of several subordina- is indexed by the signer by pointing to the denoted tors which are homonymous with prepositions, person or thing if it is present in the signing area, particles or adverbs, are unambiguous. or by setting up a point in space as a reference Clause closing delimiters are sentence end, to that person or thing, if it is outside the sign- closing comma, colon, semicolon, dash. ing area, and referring to that point whenever the The following set of clause splitting rules are respective person or object is mentioned. In or- applied (C1-C9): der to avoid ambiguity, different referents are as- Rule C1. The beginning of a sentence is a coordi- signed different spatial points. Deaf people find nate CO. it difficult to deal with complex references in writ- Rule C2. A subordinating clause linking word or ten texts where additional disambiguating markers phrase denotes a subordinate CO. are rarely available. Being a pro(noun)-drop lan- Rule C3. If a subordinate CO is on the top of the guage, Bulgarian allows the omission of the sub- stack, we look to the right for a punctuation clause ject when it is grammatically inferable from the delimiter (e.g. comma) which functions as a CC context. element. So far the following rules for subject recovery Rule C4. If a subordinate CO is on the top of the have been defined and implemented:

42 Rule SR1. In case the verb is in the first or sec- Rule R2. The possessive pronoun is replaced with ond person singular or plural and the clause lacks a a prepositional phrase formed by the preposition nominative personal pronoun that agrees with the na (of ) and the identified antecedent. finite verb, a personal pronoun with the respective Rule R3. A relative pronoun of type one is agreement features is inserted in the text. replaced with the identified antecedent. Rule SR2. In case the verb is in the third person Rule R4. The relative pronoun chiyto (whose) singular or plural and the clause lacks a noun or is replaced with a prepositional phrase formed a noun phrase that a) precedes the verb; and b) by the preposition na (of) and the identified agrees with the verb in person, number and gen- antecedent. der, the closest noun (a head in a noun phrase) in Rule R5. The relative pronoun kakavto (such the preceding clause that satisfies the agreement that) is replaced by a noun phrase formed by features of the verb is inserted in the text. (The a demonstrative pronoun and the identified precision of the rule for singular verbs is low.) antecedent takava chanta (that bag). 3.2.3 Anaphora Resolution 3.2.4 Simplification of Complex Sentences With respect to text modification regarding Complex sentences are one of the main issues anaphora resolution, we focus on a limited types for deaf readers because in BulSL, as well as in of pronominal anaphors – personal, relative and other sign languages, they are expressed as sep- possessive pronouns. arate signed statements and the relation between Bulgarian personal pronouns agree in gender them is explicit. and number with their antecedent. Possessive pro- (Van Valin and LaPolla, 1997) observe that the nouns express a relation between a possessor and elements in complex sentences (and other con- a possessed item, and agree both with their an- structions) are linked with a different degree of tecedent (through the root morpheme) and with the semantic and syntactic tightness, which is re- head noun (through the number and gender fea- flected in the Interclausal Relations Hierarchy. tures of the inflection). For instance in the sen- The clauses in a sentence have different degree of tence Vidyah direktora v negovata kola (I saw the independence, which determines whether they can director in his car), the possessive pronoun negov be moved within the sentence or whether they can indicates that the possessor is masculine or neuter form an individual sentence. singular and the inflection -a – that the possessed Temporally related events in BulSL most often is feminine gender, singular. The agreement with are represented in a chronological order, and the the possessor is a relevant feature to text modifica- relation between them is expressed by separate tion. Some relative pronouns koyto (which) (type signs or constructions (Example 1). one) agree with their antecedent in gender and number while others (type two) – chiyto (whose) Example 1. agree with the noun they modify and not with their Zabavlyavayte se, dokato nauchavate i novi antecedent. neshta. We have formulated the following rules for Have fun while you learn new things. anaphora resolution: Rule AR1. The antecedent of a personal or a pos- Signed sentence: sessive pronoun is the closest noun (the head in the Vie se zabavlyavate. Ednovremenno nauchavate noun phrase) within a given window to the left of novi neshta /ednovremenno/. the pronoun which satisfies the agreement features You have fun. Simultaneously, you learn new of the pronoun. things /simultaneously/. Rule AR2. The antecedent of a relative pronoun (the sign ’simultaneously’ can be repeated at the is the nearest noun (the head in the noun phrase) end of the sentence again) in the preceding clause that satisfies the agreement features of the pronoun. Chambers et al. (2007) and Tatu and Srikanth The following rules for modification of (2008) identify event attributes and event-event anaphora can be used: features which are used to describe temporal re- Rule R1. The third personal pronoun is replaced lations between events. Attributes include tense, with the identified antecedent. grammatical aspect, modality, polarity, event

43 class. Further, the event-event features include the razreshenie. following: before, includes, begins, ends, simul- taneously, and their respective inverses (Cham- Based on the report from the investigation, bers et al., 2007), as well as sameActor (bi- after the person has been notified about all nary feature indicating that the events share the evidence and after /he/ has been heard, the same semantic role Agent), eventCoref (binary at- authorities have issued a permit. tribute capturing co-reference information), one- Sent (true when both events are within the same Modified: sentence), relToDocDate (defining the temporal Litseto e bilo uvedomeno za vsichki dokazatelstva. relation of each event to the document date) (Tatu Litseto e bilo izslushano. and Srikanth, 2008). Sled tova vaz osnova na doklada ot razsledvaneto, (Pustejovsky et al., 2003) also introduce tempo- organat mozhe da dade razreshenie. ral functions to capture expressions such as three years ago, and use temporal prepositions (for, The person has been notified about all evi- during) and temporal connectives (before, while). dence. Three types of links are considered: TLINK (tem- The person has been heard. poral link between an event and a moment or pe- After that based on the report from the investiga- riod of time); SLINK (subordination link between tion, the authorities may issue a permit. two events); and ALINK (aspectual link between aspectual types). The structure of the complex sentences is sim- 3.3 Evaluation of System Performance plified by clause reordering that explicitly reflects The evaluation of performance is based on the the chronological order of the described events. Bulgarian part of the Bulgarian-English Clause- The preposition or postposition of clauses with Aligned Corpus (Koeva et al., 2012) which temporal links if, before, after, etc. may not match amounts to 176,397 tokens and includes several the actual causal order. In such cases the order of categories: administrative texts, fiction, news. The clauses is simply reversed based on rules of the overall evaluation of the system performance is as- type: sessed in terms of the evaluation of all subtasks Temporal link sled kato /when, after/ (Section 3.2) as presented in Table 1. The evalu- Construction CL1 temporal link CL2 ation of finite verbs and anaphora recognition, as Modification(s) CL2. Sled tova /then/ CL1. well as subject identification is performed manu- 3.2.5 Post-editing ally on a random excerpt of the corpus. Clause splitting is evaluated on the basis of the manual Post editing aims at providing grammatically cor- annotation of the corpus. We assess the precision rect and semantically complete modified text. and recall in terms of full recognition and partial Clause reordering might lead to inappropriate use recognition. In the first case the entire verb phrase, of verb tenses. Coping a subject from the previous clause, anaphora, or dropped subject is recognized sentence might require a transformation from an correctly, while in the latter – only a part of the indefinite to a definite noun phrase. Thus, several respective linguistic item is identified. We ac- checks for grammaticality and text cohesion are count for partial recognition since it is often suf- performed and relevant changes to verb forms and ficient to produce correct overall results, e.g. par- noun definiteness are made. Specific expressions tial verb phrase recognition in most cases yields are introduced to highlight temporal, causative, correct clause splitting. conditional and other relations and to serve as con- nectives. 4 Experiments and Evaluation of Example 2 shows a fully modified text. Readability of Modified Texts

Example 2. 4.1 Outline of the Experiment Original: Vaz osnova na doklada ot razsledvaneto, sled kato 4.1.1 Aims and Objectives litseto e bilo uvedomeno za vsichki dokazatelstva The objective of the experiment was to conduct a i sled kato e bilo izslushano, organat e izdal pilot testing of original and modified texts in order

44 Task Precision Recall F1 compliance with the principles of textual and fac- Finite verb phrases 0.914 0.909 0.912 tual accuracy and integrity, and appropriate com- (full) plexity. The result from the automatic modifica- Finite verb phrases 0.980 0.975 0.977 tions has been manually checked and post-edited (partial) to ensure grammaticality. Clauses borders 0.806 0.827 0.817 4.2.2 Methodology Clauses (begin- 0.908 0.931 0.919 The testing is conducted either online via tests in ning) e-form (predominantly), or using paper-based ver- Anaphora (full) 0.558 0.558 0.558 sions. Respondents are given texts of each type, Anaphora (partial) 0.615 0.615 0.615 i.e. two original and two modified texts. Each Subject (full) 0.590 0.441 0.504 text is associated with two tasks, which have to be Subject (partial) 0.772 0.548 0.671 completed correctly after the reading. The tasks seek to check the level of understanding of the Table 1: Evaluation of different stages of text main idea, details, purpose, implication, temporal modification relations (the sequence of events), and the ability to follow the text development. to determine and confirm the need of text modi- • Task-type 1: Sequence questions. The re- fication for deaf people whose primary language spondents have to arrange text elements (sen- is BulSL and the verbal language is acquired as a tences and clauses) listed in a random se- second language. quence into a chronological order. The task The rationale was to identify and distinguish be- covers temporal, causative, conditional, and tween levels of comprehension of original and au- other relations, and its goal is to test reading tomatically modified texts. comprehension which involves temporal and logical relations and inferences. 4.1.2 Respondents’ Profile The participants were selected regardless of their • Task-type 2: Multiple response questions degree and onset of hearing loss. The experiment (MRQ) for testing general reading com- targeted the following group of people: prehension. MRQ are similar to Multiple choice questions (MCQs) in that they provide • Socially active adults (18+); a predefined set of options, but MRQ allow any number and combinations of options. • BulSL users; Text Type Version # # # • People with developed reading skills. sen- clauses tem- tences poral 4.2 Pilot Test Design Methodology and shifts Implementation 1 News Original 2 6 2 4.2.1 Text Selection 2 News Modified 5 6 0 We decided to use original and modified versions 3 Admin Original 1 4 2 of journalistic (e.g. news items) and administra- 4 Admin Modified 4 4 0 tive (e.g. legal) texts. The guiding principle was to select texts that are similar in terms of length, Table 2: Structure of the test complexity, and difficulty. The selected news refer to topics of general in- 4.2.3 Structure of the Test terest such as politics in neighbouring countries, The test consists of four different texts, each of culture, etc. The administrative texts represent them with two subtasks – for checking the com- real-life scenarios, rather than abstract or rare le- prehension of temporal relations and the logical gal issues. In general, selected texts do not include structure of the events in the text (type 1), and gen- domain-specific terms and professional jargon. eral comprehension (type 2). Regarding text modification the main objective The number of sentences, clauses and temporal was to preserve the meaning of the original text in shifts for each text is presented in Table 2.

45 4.3 Analysis of Results 5 Conclusions

19 deaf adults proficient in BulSL have taken part As the pilot test suggests, the limited number of in the pilot test study. The results are presented in modifications is not sufficient to compensate for Table 3 and on Figure 1. the problems which deaf people experience with reading. A wider range of text modifications are Task Type Version correct all % necessary in order to cover the problematic areas of verbal language competence. Other issues in- 1.1 News Original 5 19 26.32 clude the use of personal and possessive pronouns, 2.1 News Modified 9 19 47.37 in particular clitics, which are often dropped, the 3.1 Admin Original 6 19 31.58 correct use of auxiliary verbs and analytical verb 4.1 Admin Modified 10 19 52.63 forms. Additional problems such as adjective and 1.2 News Original 7 19 36.84 noun agreement, subject and verb agreement, etc. 2.2 News Modified 9 19 47.37 need to be addressed specifically, since these have 3.2 Admin Original 7 19 36.84 a very different realization in sign languages (e.g., 4.2 Admin Modified 10 19 52.63 subject and verb are related spatially). It should be emphasized that there has not been Table 3: Results of chronological order sub- any systematic effort for studying BulSL so far. tasks (1.1-4.1) and general comprehension sub- The detailed exploration of the linguistic proper- tasks (1.2-4.2) ties of BulSL in relation to Bulgarian can give a deeper understanding about the problems in the We recognize the fact that the small number acquisition of Bulgarian and in particular, the of respondents does not provide sufficient data reading difficulties experienced by deaf readers. to draw conclusions regarding the improvement Directions for future work include: of readability when using modified texts. How- ever, the results show a significant improvement • To explore the relationship between reading (t = 2.0066 with p = 0.0485 < 0.05) in the over- comprehension and social, educational and all comprehension (chronological order and gen- other factors; eral understanding) when using the modified texts • To explore the dependence between reading in comparison with the original texts. skills and proficiency in BulSL; • To analyze problems in relation to vocabulary with relation to reading; • To build a detailed methodology for testing of reading comprehension; • To explore further the potential of text modi- fication with respect to BulSL in relation to the comparative analyses of the features of BulSL and verbal .

References Figure 1: Results in % of correct answers for orig- J. Albertini and C. Mayer. 2011. Using miscue analy- sis to assess comprehension in deaf college readers. inal and modified texts Journal of Deaf Studies and Deaf Education, 16:35– 46. Still, the improvement in readability after the BATOD. 2006. Training materials for language modi- text modification is very low and not sufficient to fication. provide reliable communication strategies and ac- Gerald Berent. 2004. Deaf students’ command of En- cess to information. Further work will be aimed at glish relative clauses (Paper presented at the annual more precise methodology for testing the reading convention of Teachers of English to Speakers of skills of deaf people. Other Languages).

46 Stefan Bott and Horacio Saggion. 2011. Spanish text Penny Lee. 1996. The Logic and Development of simplification: An exploratory study. In XXVII Con- the Linguistic Relativity Principle. The Whorf The- greso de la Sociedad Espaola para el Procesamiento ory Complex: A Critical Reconstruction. John Ben- del Lenguaje Natural (SEPLN 2011), Huevla, Spain. jamins Publishing.

Arnaldo Candido, Erick Maziero, Caroline Gasperin, Margaret Jelinek Lewis and Dorothy Jackson. 2001. Thiago A. S. Pardo, Lucia Specia, and Sandra Television literacy: Comprehension of program con- Aluisio. 2009. Supporting the adaptation of texts tent using closed-captions for the deaf. Journal of for poor literacy readers: a text simplification ed- Deaf Studies and Deaf Education, 6(1):43–53. itor for Brazilian Portuguese. In Proceedings of the Fourth Workshop on Innovative Use of NLP for Slavina Lozanova. 2002. Predpostavki za razvitie Building Educational Applications, pages 34–42. na bilingvizma kato metod za obuchenie na detsa s uvreden sluh. Spetsialna pedagogika. Nathanael Chambers, Shan Wang, and Dan Juraf- R. I. Mayberry. 2000. Cognitive development of sky. 2007. Classifying temporal relations between deaf children: The interface of language and percep- events. In Proceedings of the ACL 2007 Demo and tion in cognitive neuroscience. In Child Neuropsy- Poster Sessions, Prague, June 2007, pages 173–176. chology, Volume 7 of handbook of neuropsychology, R. Conrad. 1979. The deaf schoolchild. London: pages 71–107. Amesterdam: Elsvier. Harper and Row. P. McAnally, S. Rose, and S. Quigley. 1999. Reading practices with deaf learners. Austin, TX: Pro-Ed. Peter Dienes and Amit Dubey. 2003. Deep syntac- tic processing by combining shallow methods. In Ruslan Mitkov. 1999. Anaphora resolution: the Proceedings of the 41st Annual Meeting of the Asso- state of the art; working paper, (based on the col- ciation for Computational Linguistics, volume 1. ing’98/acl’98 tutorial on anaphora resolution). R. J. Hoffmeister. 2000. A piece of the puzzle: ASL Carol Musselman. 2000. How do children who can’t and reading comprehension in deaf children. In hear learn to read an alphabetic script? a review of C. Chamberlain, J. P. Morford, and R. I. Mayberry, the literature on reading and deafness. Journal of editors, Language acquisition by eye, pages 143– Deaf Studies and Deaf Education, 5:9–31. 163. Mahwah, NJ: Earlbaum. C. Padden and C Ramsey. 2000. American Sign Kentaro Inui, Atsushi Fujita, Tetsuro Takahashi, Ryu Language and reading ability in deaf children. In Iida, and Tomoya Iwakura. 2003. Text simplifi- C. Chamberlain, J. P. Morford, and R. I. Mayberry, cation for reading assistance: A project note. In editors, Language acquisition by eye, pages 165– PARAPHRASE ’03 Proceedings of the second in- 189. Mahwah, NJ: Earlbaum. ternational workshop on Paraphrasing, volume 16, pages 9–16. Association for Computational Linguis- S. J. Parault and H. William. 2010. Reading motiva- tics Stroudsburg, PA, USA. tion, reading comprehension and reading amount in deaf and hearing adults. Journal of Deaf Studies and Mark Johnson. 2002. A simple pattern-matching al- Deaf Education, 15:120–135. gorithm for recovering empty nodes and their an- tecedents. In Proceedings of the 40th Annual Meet- Peter Paul. 1996. Reading vocabulary knowledge and ing of the Association for Computational Linguis- deafness. Journal of Deaf Studies and Deaf Educa- tics. tion, 1(1):3–15. Peter Paul. 1998. Literacy and deafness: The develop- L. Kelly. 1996. The interaction of syntactic compe- ment of reading, writing and literate thought. Need- tence and vocabulary during reading by deaf stu- ham, MA: Allyn & Bacon. dents. Journal of Deaf Studies and Deaf Education, 1(1):75–90. Sarah Petersen and Mari Ostendorf. 2007. Natural language processing tools for reading level assess- S. Koeva and A. Genov. 2011. Bulgarian language ment and text simplification for bilingual education. processing chain. In Proceedings of Integration of University of Washington Seattle, WA. multilingual resources and tools in Web applica- tions. Workshop in conjunction with GSCL 2011, James Pustejovsky, Jose´ Casta no, Robert Ingria, Roser University of Hamburg. Saur´ı, Robert Gaizauskas, Andrea Setzer, Graham Katz, and Dragomir Radev. 2003. TimeML: Robust Svetla Koeva, Borislav Rizov, Ekaterina Tarpomanova, Specification of Event and Temporal Expressions in Tsvetana Dimitrova, Rositsa Dekova, Ivelina Stoy- Text. Technical report, AAAI. anova, Svetlozara Leseva, Hristina Kukova, and An- gel Genov. 2012. Bulgarian-english sentence- and Seth Kulick Ryan Gabbard and Mitchell Marcus. 2004. clause-aligned corpus. In Proceedings of the Second Using linguistic principles to recover empty cate- Workshop on Annotation of Corpora for Research gories. In Proceedings of the 42nd Annual Meeting in the Humanities (ACRH-2), Lisbon, 29 November on Association For Computational Linguistics, page 2012, pages 51–62. Lisboa: Colibri. 184191.

47 S. Saeva. 2010. Gluhota i Bilingvizam. Aeropres BG. Zhemin Zhu, Delphine Bernhard, and Iryna Gurevych. 2010. A monolingual tree-based translation model L. Selinker. 1972. Interlanguage. International Re- for sentence simplification. In Proceedings of The view of Applied Linguistics, 10:209–231. 23rd International Conference on Computational Linguistics, Beijing, China, pages 1353–1361. Advaith Siddharthan. 2011. Text simplification using typed dependencies: A comparison of the robustness of different generation strategies. In Proceedings of the 13th European Workshop on Natural Language Generation (ENLG) , pages 211, September.

L. Specia. 2010. Translating from complex to simpli- fied sentences. In 9th International Conference on Computational Processing of the Portuguese Lan- guage (Propor-2010), Porto Alegre, Brazil, vol- ume 6001 of Lecture Notes in Artificial Intelligence, pages 30–39. Springer.

William Stokoe. 1960. Sign language structure: An outline of the visual communication systems of the american deaf. Studies in linguistics: Occasional papers, 8.

William Stokoe. 1972. Semiotics and Human Sign Languages. NICI, Printers, Ghent.

Ivelina Stoyanova, Tanya Dimitrova, and Viktorija Tra- jkovska. 2003. A handbook in civil education with a sign language dictionary. In Social and Educational Training for Hearing Impaired youths: A Handbook in Civil Education with a Sign Language Dictionary. Petar Beron, Sofia. (in Bulgarian).

Rachel Sutton-Spence and Bencie Woll. 2003. The Linguistics of : An Introduc- tion. Cambridge University Press, 3rd edition.

W. Swann. 1992. Learning for All: Classroom Diver- sity. Milton Keynes: The Open University.

Marta Tatu and Munirathnam Srikanth. 2008. Experi- ments with reasoning for temporal relations between events. In COLING ’08 Proceedings of the 22nd In- ternational Conference on Computational Linguis- tics, volume 1, pages 857–864.

C. B. Traxler. 2000. The stanford achievement test, 9th edition: National norming and performance stan- dards for deaf and hard-of-hearing students. Journal of Deaf Studies and Deaf Education, 5:337348.

Beverly Trezek, Ye Wang, and Peter Paul. 2010. Read- ing and deafness: Theory, research and practice. Clifton Park, NY: Cengage Learning.

Robert Van Valin and Randy LaPolla. 1997. Syntax: Structure, Meaning, and Function. Cambridge Uni- versity Press.

Schreuder R Knoors H Snik A. Vermeulen AM, van Bon W. 2007. Reading comprehension of deaf chil- dren with cochlear implants. Journal of Deaf Stud- ies and Deaf Education, 12(3):283–302.

48