2021.conll-1.52.pdf - ACL Anthology

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 674–683November 10–11, 2021. ©2021 Association for Computational Linguistics

674

The Influence of Regional Pronunciation Variation on Children’s Spellingand the Potential Benefits of Accent Adapted Spellcheckers

Emma O’NeillADAPT Centre

University College [email protected]

Joe KennyZeekoIreland

[email protected]

Anthony VentresqueLero Research Centre

University College [email protected]

Julie Carson-BerndsenADAPT Centre

University College [email protected]

AbstractA child who is unfamiliar with the correctspelling of a word often employs a “sound itout” approach: breaking the word down intoits constituent sounds and then choosing lettersto represent the identified sounds. This oftenresults in a misspelling that is orthographicallyvery different to the intended target. Recently,efforts have been made to develop phoneticbased spellcheckers to tackle the more deviantnature of children’s misspellings. However, lit-tle work has been done to investigate the po-tential of spelling correction tools that incorpo-rate regional pronunciation variation. If a childmust first identify the sounds that make up aword, it stands to reason their pronunciationwould influence this process. We investigatethis hypothesis along with the feasibility andpotential benefits of adapting spelling correc-tion tools to more specific language variants -particularly Irish Accented English. We usemisspelling data from schoolchildren acrossIreland to adapt an existing English phonetic-based spellchecker and demonstrate improve-ments in performance. These results not onlyprompt consideration of language varieties inthe development of spellcheckers but also con-tribute to existing literature on the role of re-gional accent in the acquisition of writing pro-ficiency.

1 Introduction

Children’s spelling has long been considered to berooted in phonetics; being influenced by pronun-ciation and speech sound similarity (Read, 2018).Misspellings are often ‘creative’ in that they reflectthe phonetic judgements of the child and stem fromtheir efforts to use letters to represent the sounds oftheir spoken language. As such, these misspellingsare often unique and can deviate heavily from the

correct spelling making the automatic correction ofsuch errors a non-trivial task.

Nowadays, a popular focus in early literacy ed-ucation is the teaching of phonics with schoolsacross the English speaking world incorporatingthe method into their curricula (National Councilof Curriculum and Assessment, 2019; Bowers andBowers, 2017). Phonics based approaches to read-ing and writing involve teaching the relationshipbetween letters and sounds. For example, the word‘cat’ would be broken down into the letters ‘c’, ‘a’,‘t’ and the corresponding phonemes /K/, /AE/, /T/ 1.When tackling the spelling of an unfamiliar word,children are then encouraged to adopt a “sound itout” approach by identifying the phonetic sequenceof the word and the letters which represent thesesounds. This approach is heavily relied upon bylow achieving spellers (Daffern and Critten, 2019).

Despite the phonetic nature of children’s spellingand the employment of a literacy education method-ology that is intrinsically linked to spoken words,little work has been carried out towards developingspellcheckers that account for systematic pronun-ciation variation. Much of the existing researchon the interaction between spoken variation andliteracy acquisition typically focuses on childrenwith an African American English (AAE) dialectand the phonological and morphological deviationfrom what is referred to as Mainstream AmericanEnglish (MAE) (see Section 2.2). In their system-atic review, Snell and Andrews (2017) indicate thatthere is insufficient research on the relationship be-tween regional accent or dialect and written Englishliteracy in England. From the works reviewed, they

1Throughout this paper we refer to phonemes using ARPA-bet notation consistent with the CMU Pronouncing Dictionary(Weide, 1998) which is used in the spelling correction model.

675

surmise that there is no straightforward relationshipbetween literacy achievement and language back-ground and state that regional dialect has only aminor impact on writing. However, throughout thereview, difficulties with spelling were not consid-ered related to pronunciation variation and insteada result of the complexity between English orthog-raphy and pronunciation that affects all childrenregardless of their accent. Whilst we agree thatall children will face challenges in acquiring writ-ten literacy, we seek to explore whether regionalvariation might influence the sort of problems en-countered by students and the types of misspellingsthey produce as a result. Indeed, Terry (2006) sup-poses that “while all children must learn to negoti-ate mismatches between speech and print in orderto become good readers and writers, this processmay be particularly problematic for children whosespoken language differs substantially from stan-dard written forms”. In much the same way asone might employ a targeted approach to teachinga foreign language based on the learner’s nativelanguage, it could prove beneficial to children withregional language variants to be taught reading andwriting skills with such variation in mind and tohave access to tools designed specifically to handletheir particular variant.

This paper seeks to investigate the hypothesisthat regional variation in pronunciation influencesmisspelling productions in children’s writing. Todo this, we explore whether an accent adaptedspellchecker outperforms a baseline system whichwould suggest some underlying commonality be-tween the misspellings produced by speakers of thesame language variant. In doing so we also investi-gate the feasibility of developing such a system andthe potential benefits it would have for speakersof a regional variant. An existing English pho-netic based spelling correction tool is fine-tunedusing misspelling data collected from schoolchil-dren across Ireland. We demonstrate that this IrishAccented English (IAE) adapted model achievesbetter results on our test set of IAE misspellings.Furthermore, we demonstrate that these gains inperformance are not just a result of additional train-ing, by fine-tuning the baseline model with nonIAE misspelling data which exhibits lesser perfor-mance gains. These results would suggest that therelationship between regional pronunciation varia-tion and written literacy merits further exploration.

2 Related Work

2.1 Spelling Correction MethodsKukich (1992) discussed spelling errors as belong-ing to one of two types; typographic and cognitive.The former occurs when the writer knows the cor-rect spelling of a word but makes an error whenproducing it. For example, entering a differentcharacter than intended by mistakenly pressing anadjacent key on the keyboard. These types of er-rors tend only to deviate from the target spellingby a single edit operation (substitution, deletion,insertion or transposition) (Damerau, 1964) andso are easily corrected by the character edit dis-tance methods employed by traditional spellcheck-ers. In contrast, cognitive errors result from a lackof knowledge of how to correctly spell a word. Mis-spellings that result from an effort to capture thesound sequence of a word fall under a subset ofcognitive errors labelled phonetic errors and theytypically deviate substantially from the target word(Kukich, 1992).

Probabilistic correction models, like that of Brilland Moore (2000) or Church and Gale (1991), ex-hibited improved performance on cognitive errorsby modelling the likelihood of multiple edit opera-tions. The correction of phonetic errors in particu-lar has been tackled by incorporating pronunciationinformation as opposed to just orthographic repre-sentations. For example, by using weighted editdistances that consider the phonetic similarity be-tween graphemes (Veronis, 1988), by assigning analphanumeric code intended to capture phoneticfeatures (Russell and Odell, 1918), or by convert-ing a misspelling to its corresponding phonemesequence using letter-to-sound rules (Fisher, 1999;Toutanova and Moore, 2002). In their recent survey,Hládek et al. (2020) noted the emerging popularityof encoder-decoder architectures and deep neuralnetworks that treat the spelling correction problemas one of statistical machine translation.

Spelling correction tools targeted specifically to-wards children have also been recently developed.Downs et al. (2020) released Kidspell: a child-oriented, rule-based, phonetic spellchecker. Theirsystem makes use of the phonetic rules of Englishto map letters to keys which aim to capture accuratephonetic representations. Candidate suggestionsfor spelling correction are generated by identifyingwords with matching or similar phonetic keys. Sim-ilarly, S-capade (O’Neill et al., 2020), our own En-glish spellchecker for children, produces candidate

676

corrections that have the same or similar phonemesequences to that predicted of the misspelled word.It generates a phoneme representation of the mis-spelling and calculates the weighted edit distancebetween the misspelling and a candidate correc-tion based on the likelihood of phonemes beinginserted, deleted, or substituted for others. The na-ture of S-capade’s edit distance method allows foradaptation of the phoneme weights and we use thisas a baseline for fine-tuning on IAE speakers’ datain this work.

2.2 Spoken Variation and LiteracyAcquisition

Early research into factors affecting children’s lit-eracy acquisition surmised that speaking a “non-standard” or “non-mainstream” language varietyhindered reading and writing performance withSchwartz (1982) coining the term ‘dialect inter-ference’. More recently, an alternative (thoughnot mutually exclusive) explanation for the asso-ciation between non-mainstream productions andliteracy achievement has become prevalent in theliterature. This involves the idea that children whoproduce more non-mainstream forms, particularlyin contexts where this would not be appropriate (i.e.in the classroom), potentially have less linguisticawareness in general. This lack of awareness orflexibility likely extends to other aspects includ-ing phonological awareness which is regarded asintegral to literacy acquisition (Terry and Scarbor-ough, 2011). Thus, it is not the use of these non-mainstream forms or non-standard dialect usage ingeneral that negatively impacts reading and writingskills, but rather, high frequency of usage is an in-dicator of an underlying linguistic weakness thatalso impacts literacy.

A number of works have examined the literacyskills of AAE speaking children - specifically thosewith frequent non-mainstream forms in their spo-ken production (Charity et al., 2004; Terry, 2006;Terry and Scarborough, 2011; Terry, 2012; Terryand Connor, 2010; Connor and Craig, 2006). Thisresearch confirms that children who frequently usethe AAE variant typically have more trouble withlearning to read and write. In particular, Terry andConnor (2010) demonstrated that words with di-alect sensitive features caused spelling issues inboth struggling and typically achieving readers.Beyond the examination of AAE, similar support-ing research on spoken language and education

has been carried out concerning language variantsof Dutch, Arabic, Greek, and Carribean creoles(Driessen and Withagen, 1999; Saiegh-Haddad,2003; Siegel, 1999; Yiakoumetti, 2007). This isevidently an issue not limited to the AAE variety oreven to varieties of English. As such, even thoughthis work focuses on IAE, the adaptation of lan-guage tools to spoken variants has potential benefitacross the world’s languages.

3 Methodology

3.1 The Dataset

The dataset used in this experiment was compiledfrom a collection of surveys of schoolchildrenacross Ireland carried out by Zeeko (Everri andPark, 2018). A subset of these surveys containedopen-ended questions of the form “Why do youthink...?” or “What did you like about...?” andparticipants responded in a free-text format. In to-tal, survey responses from 628 students containedthese free-text answers and were analysed for non-word spelling errors. The students ranged in agefrom 7 to 17 years with an average age of 10 years.The full distribution of ages across respondents canbe seen in Figure 1.

7 8 9 10 11 12 13 14 15 16 170

50

100

150

200

Age

Cou

nt

Figure 1: The age distribution of survey respondents.

Whilst responses were received from locationsacross the country, 50.6% of respondents were fromschools in County Dublin. As such, later analyseson the influence of regional pronunciation are heav-ily weighted to Dublin varieties of English. Col-lecting additional and more geographically varied

677

data is considered an area of future work.Non-word misspellings were extracted from the

responses and a human annotator judged their cor-responding real-word targets. Where a target couldnot be identified from context, the misspelling wasremoved. This resulted in a corpus of 232 pairsof misspellings and real-word targets henceforthreferred to as the IAE-corpus. Examples of somemisspelling:target pairs can be seen in Table 1.

Misspelling Targetachuly actuallybekos becausedifrunt differentegicasinol educationalmishon missionsichweshen situation

Table 1: Example misspellings and real-word targetsfrom the IAE-corpus.

3.2 S-capade’s Phoneme Distance Matrix

The candidate corrections and their rankings sug-gested by the S-capade tool are determined usinga weighted edit distance measure. A graphemeto phoneme tool (CMUSphinx, 2016) between thepredicted phoneme sequence of the misspelling andthat of the candidate. In an approach similar to thatof Wagner and Fischer (1974), sequences are opti-mally aligned to give the lowest total edit distancewhich is calculated by summing the cost of eachsubstitution, deletion and insertion.

The costs of these operations are taken from aphoneme distance matrix that models the similar-ity between phonemes. Since similarity can beconsidered a function of confusability (Gallagherand Graff, 2012), this distance matrix was gener-ated based on the confusability of phonemes bothacoustically and distributionally (Kane and Carson-Berndsen, 2016; O’Neill and Carson-Berndsen,2019). If two phonemes are likely to be confused,then they are considered highly similar and thushave a low distance score and low substitution cost.

Deletions and insertions are treated as substitu-tions of a phoneme with the empty string and viceversa. The distance values between phonemes andthe empty string were based on existing literatureon epenthesis and ellision (Collins and Mees, 2013;Fourakis and Port, 1986; Gimson and Ramsaran,1970; Itô, 1989; Yip, 1987).

If a misspelling has the same phoneme sequence

as its real-word target then the edit distance willbe 0 and the target will likely be the top rankedcandidate correction. However, if the misspelling’spredicted phoneme sequence does not match anyreal-word candidates, the suggested correction willbe the candidate deemed most similar as a result ofits low edit distance score (O’Neill et al., 2020).

The values in the distance matrix are not spe-cific to one spoken variety and instead modelgeneral phonemic confusability and phonologicalprocess of English. However, different realisa-tions of phonemes in a variant of English mightlead to differences in the degree of similarity be-tween phonemes. For example, consider the /TH/phoneme which, in IAE and particularly in Dublinvarieties, often undergoes fortition and is realisedas an alveolar stop rather than a dental fricative(Hickey, 2004). This could potentially prompt achild adopting a “sound it out” approach to spellingto encode the /T/ phoneme instead. Indeed, withinthe IAE dataset (see Section 3.1), we see possibleexamples of this effect as in *someting (something)and *tink (think). As such, fine-tuning the values inthe distance matrix using misspellings produced byIAE speakers should capture these variant specificeffects and result in better performance of the cor-rection tool on spelling errors from such speakers.

3.3 Irish Accented English Fine-Tuning

Given the limited size of the dataset, a k-fold crossvalidation approach was used to tune and evaluatethe performance of an IAE adapted spelling correc-tion tool. The IAE-corpus was separated randomlyinto 10 folds. Each fold was held out as a test setwhilst the other 9 folds were used to fine-tune thedistance matrix of the S-capade tool.

In the best case performance, the weighted editdistance between a misspelling and its correspond-ing real-word target would be lower than that ofthe other potential candidate corrections. Thus, theweights associated with the substitutions, deletions,and insertions observed in these misspelling:targetpairs should decrease whilst those observed be-tween misspelling:non-target pairs should increase.To do this we implement a shallow neural networkperforming logistic regression.

For each misspelling in the training set, the S-capade base model is used to generate the 10 mostlikely candidate corrections based on weightededit distance. The predicted phoneme sequenceof the misspelling is obtained using a grapheme

678

Figure 2: Visualisation of generating sparse input vectors from substitution counts.

to phoneme tool (CMUSphinx, 2016) trained onthe CMU Pronouncing Dictionary (Weide, 1998).These sequences are then aligned with those ofeach candidate as can be seen in Table 2. In thisexample we observe the following substitutions:

• AE : AE

• K : empty string

• SH : CH

• AH : UW

• L : L

• IY : IY

Note that the absence of an edit operation istreated as a substitution of a phoneme with itselfand insertions and deletions are treated as sub-stitutions involving the empty string. Each mis-spelling:candidate pair is then represented as asparse vector of length 1600 with each value repre-senting the number of occurrences of all possiblephoneme substitutions (see Figure 2 for a visuali-sation of this process).

Spelling Phoneme Sequenceactually AE K SH AH L IYachuly AE CH UW L IY

Table 2: Comparison of the phoneme sequences be-tween ‘actually’ and ‘achuly’.

Each training set has between 208-209 mis-spellings and each misspelling has 10 candidatecorrections. If not already present in the list of can-didates, the real-word target is added. Some candi-date lists contain multiple instances of the target asa result of different pronunciations within the dic-tionary. As such, we have, for each training set, be-tween 250-266 instances of true misspelling:targetpairs and between 1856-1883 instances of negativemisspelling:non-target pairs.

In a similar process to that shown in Figure 2, theoriginal phoneme distance matrix used by S-capadewas converted into a weight vector of length 1600.A single-layer neural network is used to performlogistic regression in an effort to classify the train-ing instances as either true misspelling:target pairs(0 label) or negative misspelling:non-target pairs(1 label). This allows us to tune the weights usingbackpropagation to minimise loss. By multiplyingeach training instance vector by the weight vectorwe obtain the weighted edit distance score used byS-capade to rank candidates. A bias is added anda sigmoid transform is then applied to determinethe probability that a training instance is a negativemisspelling:non-target pair (see Equation 1 whereW is the weight vector, Xi is the training vector i,and b is the bias).

yi = σ(W ∗Xi + b) (1)

679

MRR Recall@1 Recall@3 Recall@5 Recall@10S-capade Base Model 0.623 0.543 0.694 0.746 0.789

Irish Accented English Model 0.693 0.629 0.746 0.781 0.828Holbrook British English Model 0.673 0.599 0.728 0.784 0.823

Table 3: Mean Reciprocal Rank and Mean Recall@K for the three models.

The weight vector is then tuned using backprop-agation and cross-entropy loss (see Equation 2where N is the number of training instances, y isthe true label, and y is the probability of a negativemisspelling:non-target pair).

L = − 1

N

N∑n=1

[ynlogyn−(1−yn)log(1−yn)] (2)

The resulting tuned weight vector is then reshapedinto a 40x40 distance matrix for use with the S-capade spelling correction tool.

3.4 Non IAE Fine-Tuning

In order to determine whether improved perfor-mance of the IAE fine-tuned model was a resultof adaptation to the types of errors produced un-der the influence of a regional variant and not justthe product of increased training, we sought totest a model tuned using misspelling data fromchildren with a different English variant. For thispurpose we chose the Holbrook corpus (Mitton,1985) - a collection of misspellings extracted fromthe writings of British schoolchildren (Holbrook,1964). For the sake of comparison with the IAE-corpus, 232 misspellings were randomly sampledfrom the Holbrook corpus for the purposes of fine-tuning. The methodology discussed previously inSection 3.3 was similarly applied with the excep-tion of the k-fold cross validation. Instead all 232misspellings were supplied as a single training set.The results of the logistic regression weight tuningwere then passed to the S-capade spellchecker tobe used as edit operation costs and the Holbrookfine-tuned model was tested on the IAE-corpus ofmisspellings. We expect this model to considerablyoutperform the base model since it is fine-tuned onchildren’s misspelling data but not to perform aswell as the IAE-tuned model since it is not adaptedto the specific features of IAE (although BritishEnglish and IAE will share a number of features incommon).

4 Results

The effectiveness of a spelling correction toollies in its ability to suggest appropriate candidatecorrections and in ranking the true target highlyamongst them. As such, we determine the effec-tiveness of our models based on whether or not thetarget spelling occurred in the top K candidates (Re-call@K) and on the Mean Reciprocal Rank (MRR).Since there is only ever one relevant item in the listof candidates, namely the intended target word, Re-call@K is 1 if the target is in the top K candidatesand 0 otherwise. MRR (see Formula 3) is a mea-sure of how highly the target typically ranks in thelist of candidates with values closer to 1 indicatinghigher average target rankings.

MRR =1

|Q|

|Q|∑i=1

1

ranki(3)

Each IAE-tuned distance matrix is supplied tothe S-capade spelling correction tool (replacing theoriginal base model matrix) which is then run onthe misspellings in the corresponding test fold. Toassess the performance of a spellchecker fine-tunedusing IAE data, we report the mean value acrossall test folds for each of our evaluation measuresand compare with those obtained using the baselineS-capade model and on the Holbrook fine-tunedmodel. The results can be seen in Table 3 with thebest performance across the three models indicatedin bold.

5 Discussion

From the results in Table 3, we can see that the IAE-tuned model outperforms both the baseline modeland the Holbrook tuned model in almost all of ourevaluation metrics (with a comparable result forRecall@5). Accuracy, (Recall@1), increased by8.6% absolute over the base model and Recall@10increased by 3.9% absolute. These increases wouldmean that the system is more likely to suggest theintended target word as the top suggestion and, inan interactive spellchecking environment, is morelikely to present the user with a list of candidate

680

Figure 3: Resultant phoneme distance matrix from IAE fine-tuning.

corrections that contains the target. Thus, the expe-rience of a child with an IAE variant would improvethrough the model’s adaptation. The performancegains of the IAE-tuned model is preliminary evi-dence that there exists some commonality or sharedfeatures characteristic of the types of misspellingsproduced by IAE speaking children. These featuresare not all shared by the British misspellings in theHolbrook corpus which did not attain as high per-formance gains particularly in MRR and accuracy.

In theory, the resultant distance matrix afterIAE fine-tuning should capture this commonalityand the values should relate to aspects of the IAEphonological system and phonetic realisations. Fig-ure 3 shows this tuned distance matrix, averagedacross the ten folds, visualised as a heat map wherea darker colour indicates a lower distance and thushigher degree of similarity between phonemes. Ac-counting for the limited size of the IAE corpus andthe fact that not all features of IAE will manifestin misspellings, the distance matrix does appearto exhibit the influence of IAE phonology. Manyobserved values could be considered general En-glish features. For example, the /NG/ phoneme in

a target being likely substituted for /N/ (as in themisspelling *bein (being)) or the high degree ofvariability among vowels. However, some valuesare specific to IAE and Dublin English in particu-lar (Hickey, 2004). For instance, the most likelyphoneme to be substituted for /TH/ in the targetis /T/. This feasibly results from the fortition ofdental fricatives to alveolar stops in IAE promptingspellings such as *someting (something) and *tink(think). /AY/ onset raising is also common andcould explain the misspelling *niss (nice) and thetarget /AY/ phoneme’s most likely substitutes beingthe more central vowels /AH/, /EY/, and /IH/.

The comparative performance of the IAE-tunedspelling correction tool and the qualitative analy-sis of the resultant distance matrix would suggestthat regional variation in pronunciation does some-what influence misspelling productions in chil-dren’s writing. The approach taken to fine-tunethe existing model, albeit simple, has proven to beeffective in adapting a correction tool to IAE. Thisbodes well for future research and developmentin this area on a larger scale. The adaptation ofliteracy acquisition aids and educational tools to

681

a particular language variant could offer more tar-geted support to children who may otherwise faceobstacles in their learning. It is hoped that this workwill prompt more extensive research into the inter-action between spoken language and literacy andhow this can be incorporated into the developmentof language tools in order to ensure all children canbenefit regardless of language variant.

6 Limitations and Future Work

This work is intended as a preliminary investigationinto the idea that regional pronunciation variationmanifests in the spelling errors of children and thatan adapted spelling correction model is both fea-sible and beneficial to speakers of such variants.Whilst initial results support the original hypoth-esis, it is important to note the limitations of theapproach presented here. The IAE-corpus used fortuning and testing is limited in size and scope. Itis a relatively small sample of misspellings and isbiased towards IAE speakers from Dublin. Thecollection of a larger dataset that is more repre-sentative of Ireland as a whole is a planned areaof future work pending experimental design andethical approval. The weighted edit distance ap-proach employed by S-capade only accounts forone-to-one phoneme mappings and is context inde-pendent. Thus, potentially influential features ofIAE may not be adequately captured. For exam-ple, in Dublin English, short vowels are typicallylengthened when they occur before /R/ (Hickey,2004). This context dependent feature might beencoded in misspellings but the current model hasno way of distinguishing the specific _/R/ environ-ment. Should this prove to be a significant short-coming, a more complex method of spelling cor-rection would need to be applied. A grapheme tophoneme tool trained on the CMU pronouncing dic-tionary is used to predict the phoneme sequencesof misspellings. This does not always accuratelyrepresent the child’s intended phoneme sequence.It could be possible to improve the predictions bytraining the tool on misspelling data.

7 Conclusion

We have presented preliminary evidence of the in-fluence of regional pronunciation variation on chil-dren’s misspellings. By fine-tuning an English pho-netic spellchecker to misspellings produced by IAEspeakers we demonstrate improved performanceon similar test data. Qualitative analysis of the re-

sulting model suggests it was able to learn somephonological features of the language variant. Fu-ture work will seek to further explore this relation-ship through the expansion of the IAE misspellingdataset since its limited size and bias toward oneparticular county was identified as a limitation.

Ethical Considerations

Zeeko obtained consent to collect and analysethe survey responses for research purposes. Thedataset used in this paper consists of misspellingsfrom these responses extracted with express per-mission from Zeeko. It is anonymised and containsno personally identifiable information. We presentinitial work towards a spellchecking applicationintended to benefit children with regional pronun-ciation variation who may previously have experi-enced difficulties using existing phonetic spellingcorrection tools. This paper also argues that fur-ther research exploring the relationship betweenregional accent and written literacy is required inan effort to destigmatise such language variationand offer targeted support to children’s learning.

Acknowledgements

The ADAPT Centre for Digital Content Tech-nology (www.adaptcentre.ie) is funded un-der the SFI Research Centres Programme (Grant13/RC/2106_P2). Lero - the SFI Research Cen-tre for Software (www.lero.ie) is funded un-der the SFI Research Centres Programme (Grant13/RC/2094_P2). We would like to thank the Zeekopersonnel who have run the online wellbeing work-shops and surveys - and all the schools who havebeen involved with Zeeko over the years.

References

Jeffrey S Bowers and Peter N Bowers. 2017. Beyondphonics: The case for teaching children the logic ofthe english spelling system. Educational Psycholo-gist, 52(2):124–141.

Eric Brill and Robert C Moore. 2000. An improvederror model for noisy channel spelling correction. InACL, pages 286–293.

Anne H Charity, Hollis S Scarborough, and Dar-ion M Griffin. 2004. Familiarity with school en-glish in african american children and its relationto early reading achievement. Child development,75(5):1340–1356.

682

Kenneth W Church and William A Gale. 1991. Proba-bility scoring for spelling correction. Statistics andComputing, 1(2):93–103.

CMUSphinx. 2016. Grapheme-to-phoneme tool basedon sequence-to-sequence learning.

Beverley Collins and Inger M Mees. 2013. Practicalphonetics and phonology: A resource book for stu-dents. Routledge.

Carol McDonald Connor and Holly K Craig. 2006.African american preschoolers’ language, emergentliteracy skills, and use of african american english:A complex relation.

Tessa Daffern and Sarah Critten. 2019. Student andteacher perspectives on spelling. Australian Journalof Language and Literacy, 42(1):40–57.

Fred J. Damerau. 1964. A technique for computer de-tection and correction of spelling errors. Commun.ACM, 7(3):171–176.

Brody Downs, Oghenemaro Anuyah, Aprajita Shukla,Jerry Alan Fails, Sole Pera, Katherine Wright,and Casey Kennington. 2020. Kidspell: A child-oriented, rule-based, phonetic spellchecker. In Pro-ceedings of The 12th Language Resources and Eval-uation Conference, pages 6937–6946.

Geert Driessen and Virgie Withagen. 1999. Languagevarieties and educational achievement of indigenousprimary school pupils. Language Culture and Cur-riculum, 12(1):1–22.

Marina Everri and Kirsty Park. 2018. Children’s onlinebehaviours in Irish primary and secondary schools.Technical report, Zeeko, Nova UCD.

William M Fisher. 1999. A statistical text-to-phonefunction using ngrams and rules. In ICASSP, vol-ume 2, pages 649–652.

Marios Fourakis and Robert Port. 1986. Stop epenthe-sis in english. Journal of Phonetics, 14(2):197–221.

Gillian Gallagher and Peter Graff. 2012. The role ofsimilarity in phonology. Lingua, 2(122):107–111.

Alfred Charles Gimson and Susan Ramsaran. 1970. Anintroduction to the pronunciation of English, vol-ume 4. Edward Arnold London.

Raymond Hickey. 2004. The phonology of irish en-glish. Handbook of varieties of English, 1:68–97.

Daniel Hládek, Ján Staš, and Matúš Pleva. 2020. Sur-vey of automatic spelling correction. Electronics,9(10):1670.

David Holbrook. 1964. English for the rejected: Train-ing literacy in the lower streams of the secondaryschool.

Junko Itô. 1989. A prosodic theory of epenthesis. Nat-ural Language & Linguistic Theory, 7(2):217–259.

Mark Kane and Julie Carson-Berndsen. 2016. Enhanc-ing data-driven phone confusions using restrictedrecognition. In INTERSPEECH, pages 3693–3697.

Karen Kukich. 1992. Techniques for automaticallycorrecting words in text. Acm Computing Surveys(CSUR), 24(4):377–439.

Roger Mitton. 1985. A collection of computer-readable corpora of english spelling errors. Cogni-tive Neuropsychology, 2(3):275–279.

National Council of Curriculum and Assessment. 2019.Primary Language Curriculum. Government of Ire-land, Dublin, Ireland.

Emma O’Neill and Julie Carson-Berndsen. 2019. Theeffect of phoneme distribution on perceptual similar-ity in English. INTERSPEECH, pages 1941–1945.

Emma O’Neill, Robert Young, Elsa Thiaville, Muire-ann MacCarthy, Julie Carson-Berndsen, and An-thony Ventresque. 2020. S-capade: Spelling cor-rection aimed at particularly deviant errors. In In-ternational Conference on Statistical Language andSpeech Processing, pages 85–96. Springer.

Charles Read. 2018. Children’s creative spelling.Routledge.

Robert C. Russell and Margaret King Odell. 1918.Soundex. US patent 1,261,167.

Elinor Saiegh-Haddad. 2003. Linguistic distance andinitial reading acquisition: The case of arabic diglos-sia. Applied Psycholinguistics, 24(3):431.

Judith I Schwartz. 1982. Dialect interference in the at-tainment of literacy. Journal of Reading, 25(5):440–446.

Jeff Siegel. 1999. Creoles and minority dialects in ed-ucation: An overview. Journal of Multilingual andMulticultural Development, 20(6):508–531.

Julia Snell and Richard Andrews. 2017. To what extentdoes a regional dialect and accent impact on the de-velopment of reading and writing skills? CambridgeJournal of Education, 47(3):297–313.

Nicole Patton Terry. 2006. Relations between dialectvariation, grammar, and early spelling skills. Read-ing and Writing, 19(9):907–931.

Nicole Patton Terry. 2012. Examining relationshipsamong dialect variation and emergent literacy skills.Communication Disorders Quarterly, 33(2):67–77.

Nicole Patton Terry and Carol Connor. 2010. Africanamerican english and spelling: How do secondgraders spell dialect-sensitive features of words?Learning Disability Quarterly, 33(3):199–210.

Nicole Patton Terry and Hollis S Scarborough. 2011.The phonological hypothesis as a valuable frame-work for studying the relation of dialect variation toearly reading skills.

683

Kristina Toutanova and Robert C Moore. 2002. Pro-nunciation modeling for improved spelling correc-tion. In Proceedings of the 40th Annual Meetingon Association for Computational Linguistics, pages144–151.

Jean Veronis. 1988. Computerized correction of phono-graphic errors. Computers and the Humanities,22(1):43–56.

Robert A Wagner and Michael J Fischer. 1974.The string-to-string correction problem. JACM,21(1):168–173.

Robert L Weide. 1998. The CMU pronouncing dictio-nary.

Androula Yiakoumetti. 2007. Choice of classroom lan-guage in bidialectal communities: to include or toexclude the dialect? Cambridge journal of educa-tion, 37(1):51–66.

Moira Yip. 1987. English vowel epenthesis. NaturalLanguage & Linguistic Theory, pages 463–484.