Building Roger: Technical Challenges While Developing a Bilingual

Corpus Management and Query Platform

Cosmin Strilețchi

, Mădălina Chitez

and Karla Csürös

Communications Department, Technical University of Cluj-Napoca, Cluj-Napoca, Romania

Department of Modern Languages and Literatures, West University of Timisoara, Timisoara, Romania

Keywords: Corpus Query Platform, Academic Writing, Bilingual Corpus, Big Data, ROGER.

Abstract: This paper presents an approach to a bilingual Corpus query system. ROGER has been designed and

implemented as a cross-platform distributed web application. The backend interface available to authenticated

administrators provides the digital tools for managing the database stored texts and associated metadata, and

also offers an extensive statistics mechanism that cover the data composition and usage (words, characters,

languages, study levels, genres, domains and n-grams). The frontend capabilities are offered to the registered

users allowing them to search for specific keywords and to refine the obtained results by applying a series of

filters. Current platform features include search terms and phrases, n-gram distributions and statistical

visualizations for performed queries. After inputting a search term / phase, the user may filter available texts

by: (i) language (English, Romanian); (ii) student genre (currently 20 genres); (iii) study year (1 through 4);

(iv) level (BA, MA or PhD); (v) discipline (currently 8 disciplines) and (vi) gender (male, female or

unknown). A series of solutions have been implemented to improve the response times of the intensely

computational procedures that manipulate big amounts of data.

1 INTRODUCTION

This paper presents a corpus search engine for the

bilingual comparable corpus ROGER (Corpus of

ROmanian Academic GEnRes), compiled by the

research group at CODHUS (Centre for Corpus

Related Digital Humanities) (Chitez et al, 2020) from

the West University of Timisoara. ROGER consists

of novice academic writing genres, in Romanian (L1)

and English (L2), collected from Romanian

universities, with the purpose of investigating student

writing practices in both their mother tongue and

English as a Foreign Language.

The corpus support platform addresses learner

corpus data in a multidimensional contrastive

framework: genres written by students in L1 versus

L2, genres written in different disciplines as well as

genres written at different study levels (Bachelor’s,

Master’s and Doctoral study programmes).

The present paper aims to introduce the ROGER

corpus platform to academic researchers and experts

https://orcid.org/0000-0002-5663-9159

https://orcid.org/0000-0001-9005-3429

https://orcid.org/0000-0002-9556-3724

in fields such as Corpus Linguistics (McEnery, 2019),

Academic Writing, Contrastive Analysis (Johnsson,

2003), Language for Specific Purpose studies

(Ackerley, 2021) and Computer-Assisted Language

Learning (Chitez & Bercuci, 2019). What

distinguishes the ROGER platform from other corpus

support platforms are several salient characteristics

(see also Section 2.1): (a) it is the first bilingual

novice / learner academic writing corpus with a

dedicated open source corpus query platform; (b) it

is the first corpus offering information about novice /

learner academic writing in the Romanian context; (c)

it is the first corpus support platform offering

discipline-specific information about novice / learner

academic writing ; (d) it is the first open source

corpus support platform that can be used for a

multitude of academic writing applications and

research studies: Romanian-English genre specific

contrastive studies, inter-disciplinary linguistic

contrastive studies, studies in English L2 disciplinary

and general writing.

390

Strilet

chi, C., Chitez, M. and Csürös, K.

Building Roger: Technical Challenges While Developing a Bilingual Corpus Management and Query Platform.

DOI: 10.5220/0011144500003266

In Proceedings of the 17th International Conference on Software Technologies (ICSOFT 2022), pages 390-398

ISBN: 978-989-758-588-3; ISSN: 2184-2833

The paper is structured as follows: In Section 2,

we describe the background of the ROGER project,

as well as various aspects of the compilation and

annotation processes. In Section 3, we focus on our

platform’s implementation details mentioning the

technical challenges and provided solutions. Section

4 covers the facilities offered to the ROGER

administrators and regular users. The conclusions are

emphasized in the final section.

2 A CORPUS SYSTEM

OVERVIEW

2.1 Background and Motivation

Numerous studies on student writing have used cross-

language comparisons (Donahue, 2002, 2009; Foster,

2006; Siepmann, 2006; Foster & Russell, 2002;

Kaiser, 2003; Trumpp, 1998) to detect rhetorical

patterns or linguistic interference. However, such

studies were not based on extensive linguistic

datasets, which can be employed to extract statistical

results. Beginning in about 2000, a new generation of

research studies emerged, using text corpora to study

texts more systematically and more efficiently. In the

case of academic writing studies, collecting corpora

is rather challenging, considering that authentic

materials (e.g. student papers) are not quite

accessible, teachers / tutors might not be willing to

collect student papers or text processing in corpus

format is time-consuming.

Until now, several well-known research projects

have resulted in reference corpora for academic-

writing research: (1) The ICLE range of corpora is,

certainly, the pioneering and most extensive learner

corpus project, serving as a valuable resource for

studies of Contrastive Interlanguage Analysis

(Granger, 1996); (2) the British Academic Written

English Corpus (BAWE): a large genre mapping

study by Hilary Nesi (2008; Alsop & Nesi, 2009) in

the UK led to a corpus of student texts in which the

authors identified about 100 different genres which

they group in 13 genre families (see Heuboeck,

Holmes & Nesi, 2009, pp. 46-50); (3) Michigan

Corpus of Upper-Level Student Papers (MICUSP), in

USA (O'Donnell & Römer, 2012; Römer &

O'Donnell, 2011): the corpus includes A-graded

upper-level papers in 16 disciplines at 4 levels of 7

paper types with 8 textual features; (4) Corpus of

Academic Learner English (CALE), in Germany,

aiming at developing a corpus-driven, text-centred

method based on linguistic criteria for the assessment

of writing proficiency in the academic register

(Callies & Zaytseva, 2013a, 2013b); (5) the Varieties

of English for Specific Purposes Database (VESPA),

compiled at the Université Catholique de Louvain, in

Belgium, with the purpose of contrasting academic-

writing ESP features from various mother tongue

backgrounds (Paquot, Hasselgård & Ebeling, 2013);

(6) The Romanian Corpus of Learner English

(RoCLE) compiled by Chitez (2014): it contains

academic writing papers (essays and literary texts)

written by Romanian students; (5) Corpus &

Repository of Writing (CROW), in USA, which is a

learner corpus planned for research, teaching,

mentoring and collaboration (Kwon et al., 2018).

From all these corpora, only few have also developed

their own corpus query platforms: MICUSP and

CROW are online free-access platforms, while ICLE

is a licence-based product. An alternative model is

offered by the BAWE corpus, which can be accessed

via SketchEngine (Kilgarriff et al. 2014), which has

been open-access until April 2022.

In this context, the ROGER corpus (Chitez et al.,

2021) and corpus support platform is unique in the

following components: online free-access corpus

query platform, bilingual corpus type, user-friendly

interface with statistical and feature extraction

options (e.g. n-grams).

2.2 Corpus Compilation

A variety of considerations were taken when

designing the ROGER corpus. The main target of the

ROGER project is the study of language use in

contemporary native Romanian (L1) and learner

English (L2) academic writing tasks by students

attending nine Romanian universities. The corpus

was collected over a four-year period (2018-2021)

with the help of 27 collaborators who were part of the

AWICNET (Academic Writing Collection Network)

subproject (AWICNET, 2019). AWICNET members

identified student contributors, explained the purpose

of the corpus collection process, obtained student

consent and collected the student texts, which were

both uploaded to the ROGER private cloud drive.

Student informants filled in a form that gathered

various metadata: demographic information,

educational background and writing practices, as well

as a GDPR section in which they consented to having

their data collected. Between 2018 and early 2020,

student gave their handwritten consent via printed

forms that were stored securely; since the beginning

of the COVID-19 pandemic, in observance with all

health protocols, all such forms were submitted by the

students in digital format via a Google Form which

Building Roger: Technical Challenges While Developing a Bilingual Corpus Management and Query Platform

391

we provided through our collaborators. The

previously mentioned metadata was stored separately

in *.xls format files. Since the metadata was collected

entirely in Romanian and consisted of student fill-in-

the-blank input, it was later uniformized and

translated into English. After the metadata was

processed, separate *.xls files were created to remove

any identifying or unessential information for the

ROGER application.

2.3 Corpus Annotation

Students submitted their texts in various digital

formats (primarily *.docx and *.pdf) or in paper

format (handwritten or printed). The digital variants

were converted to *.txt files via file conversion or

OCR systems, checked for transformation accuracy

and processed. As for the paper format texts, they

were scanned and transcribed by our team faithfully,

keeping all errors (e.g., grammar, spelling) made by

learners and preserving all diacritical marks in

Romanian texts.

The processed text files contain basic markup that

aims to anonymize the files and remove any

unnecessary information. Anonymization, or de-

identification, is achieved by replacing any personal

or identifiable information about the author of the

paper, the collaborator, or the university (e.g.,

<CONFIDENTIAL_NAME> replaces the names of

the student or the collaborator). We have decided to

replace all confidential data instead of simply deleting

it in order to preserve the layout of the paper.

Unnecessary information, which refers to any parts of

the paper that could interfere with the results of

statistical linguistic analyses (such as references,

tables, mathematical calculations, graphs, etc.), were

also marked up (e.g., <REF> replaces any in-text

citations, footnotes, and bibliographies).

Since the corpus comprises a very large amount of

data and metadata, building and processing the corpus

involved significant time and human resources from

our team of researchers, research assistants, interns

and volunteers working part-time or full-time on the

ROGER corpus. Due to this, we created a digital

guide for processing ROGER texts to guarantee

uniformity across the corpus.

3 IMPLEMENTATION DETAILS

The ROGER Corpus platform was designed and

implemented as a web distributed software

application, so that it will have no special

requirements (libraries, plugins, etc.) from the

devices (computers, laptops, mobile phones, tablets)

used for accessing it. The system was implemented

on a LAMP (Linux, Apache, MySQL, PHP / Perl /

Python) software stack.

3.1 Corpus Data Structure and Flow

At the core of the entire Corpus platform resides the

central database. The Input/Output operations that

store and retrieve data are crystallized in two web

accessible interfaces. The first one contains all the

facilities meant for the Corpus administrators while

the other one offers the tools that can be accessed by

the application’s visitors.

The ROGER Corpus system’s database is

supposed to store entire texts written by various

authors. As simple as it may sound, this involves

many aspects that must be taken care of.

The data associated with any Corpus application

is defined by its volume. The amount of stored data is

expected to be very large, and the immediate

consequence refers to increased access times. A

regular database structure can manifest unreasonably

large time intervals for writing and reading

operations. As opposed to this, a carefully planned

and structured database can significantly reduce the

input and output time intervals making the entire

system responsive.

3.1.1 Corpus Texts

The Corpus database is constructed around the corpus

texts. This content is stored in files and arrives in the

platform by being uploaded via the administration

interface.

A file contains a single text written in English or

Romanian containing plain text and is composed of

ANSI or UTF-8 characters.

Each file contains a useful payload (the Corpus

text itself) accompanied by secondary data delimited

by special markups.

A typical Corpus file looks like this.

<INTERNAL_IDENTIFIER>

<TITLE>Paper title</TITLE>

<MARKUP_1> markup1 content

<MARKUP_2> markup2 content

Actual text content with or without

<MARKUP_N> additional markups.

The only markup that is of interest while storing the

files’ content in the database is represented by the

<INTERNAL_INDENTIFIER>. Therefore, a corpus

file must pass through the following steps in order for

its content to be stored in the database.

▪ upload: the text file arrives on the server

ICSOFT 2022 - 17th International Conference on Software Technologies

392

▪ normalization: the file’s special characters

(escape sequences like \n, \r, \t, \0) along with

inappropriate spacing (newlines and multiple

space characters) are filtered

▪ processing: the internal identifier gets isolated,

validated, and extracted, the rest of the markups

and associated data are eliminated

▪ storing: the filtered content is deposited in the

database using the detected internal identifier

as primary key

Since our platform stores thousands of texts,

uploading them one by one would be a very time-

consuming process. Instead, bulk files upload is

allowed and the platform processes 30 files upload at

a time. An average Corpus file’s size is 100 Kbytes

so a bulk upload would consist in processing

approximately 3 Mbytes of data which is totally

reasonable.

Passing the uploaded files through all the steps

mentioned above requires between 2 and 40 seconds,

depending on the total size of processed data. To keep

the server connection alive and to avoid keeping the

administrator waiting in front of an unresponsive

page (the regular timeout for server response in 30

seconds) our application makes repeated AJAX calls

for each of the uploaded files.

After fully interpreting a file, a message

addressing the parsing success is displayed to the

administrator. The process continues until all the

uploaded files have been processed.

We did not increase the number of the uploaded

files even though their volume would allow this due

to the computational intensity of the operations

mentioned above.

3.1.2 Corpus Metadata

Each corpus text has a series of satellite metadata

information that categorizes it. This additional data

refers to the content’s

▪ genre

▪ discipline

▪ author’s gender

▪ author’s study level

▪ author’s study year

A traditional Content Management approach

would allow the administrator to fully configure each

Corpus text by introducing manually all the required

data. Such a process would be very slow and highly

ineffective in this platform’s situation due to the large

number of texts. The database is meant to store

thousands of records and uploading and configuring

manually the corresponding metadata would be a very

time-consuming activity (dozens of man-hours).

A new solution had to be invented to provide an

alternative and much more efficient route for storing

the texts along with the associated metadata. We

came up with creating a metadata *.xls file that

contains each text’s satellite information. The

correspondence between this file’s rows and the

corresponding text is made by an internal id

matching. Uploading and parsing the metadata file

would result in decorating the associated Corpus texts

with the corresponding categorizing information.

3.1.3 Corpus n-grams

A very important aspect related to any Corpus

database refers to the n-grams extracted from the

stored texts.

An n-gram’s relevance is given by its appearance

number in a certain context. The Roger Corpus

context is defined by:

▪ the language in which the n-gram was found

▪ the corresponding metadata classifiers (genre,

discipline, author’s gender, study level and

study year)

For computing the 2, 3, 4 and 5 n-grams, all the

adjacent groups of words are counted. An n-gram is

considered valid only if it occurs more than once.

Our database contains the following average n-

grams counters (Table 1).

Table 1: Roger Corpus n-grams counters.

n-gram Avg.

occurrences/ text

Avg. occurrences/

500 texts

rams 168.3 81,941

rams 129.3 64,650

rams 73.61 36,806

5-grams 49.5 24,798

Considering these numbers and keeping in mind

that the entire database contains thousands of texts,

the n-grams counters must be computed in a non-real-

time process. If stored accordingly, counting, and

displaying on-demand categorized n-grams becomes

a feasible task.

3.1.4 Corpus Statistics

A set of predefined statistics is defined by the Roger

Corpus requirements. They refer to

▪ texts counters for each discipline and genre,

both for Romanian and English writings

▪ total texts counters per language

▪ total words number per each language

Building Roger: Technical Challenges While Developing a Bilingual Corpus Management and Query Platform

393

▪ total characters number per each language

▪ total number of 2, 3, 4 and 5 n-grams

The numbers obtained for our current database are

reflected in Table 2. The n-grams are presented in

Table 1. and are excepted from this report.

Table 2: Roger Corpus statistics counters.

Item

Avg. counters/

500 texts

Romanian words 369,197.27

English words 432,025.79

Romanian characters 2,201,560.80

5-grams 2,578,191.17

Considering the values in the table above and the fact

that the counting SQL functions work slower as the

data increases in volume, computing the counters in a

non-real-time process is also recommended.

3.1.5 Database Storage

The database entities that define the metadata

characterizers are the following ones:

▪ corpus_genres (corpus_genre_id, name)

▪ corpus_disciplines (corpus_discipline_id,

name)

▪ corpus_genders (corpus_gender_id, name)

▪ corpus_study_levels (corpus_study_level_id,

name)

The database entity that stores the metadata is

defined by

▪ corpus_metadata_id (an autoincremented

value)

▪ internal_text_id (an external reference to the

corpus texts storing entity)

▪ corpus_genre_id (an external reference to the

corpus_genres entity)

▪ corpus_discipline_id (an external reference to

the corpus_study_levels entity)

▪ corpus_gender_id (an external reference to the

corpus_gender entity)

▪ corpus_study_level_id (an external reference

to the corpus_study_level entity)

Storing only the references to the actual entities

that store categorizing information was preferred as

opposed to directly memorizing the corresponding

name for two main reasons. Firstly, the text is

detached from its numerical abstraction making it

suitable for a possible future multilanguage

translation. Secondly an SQL syntax that retrieves

data from the metadata entity will work faster if the

filtering is done on numeric values than on textual

data.

The database entity used for storing the corpus

texts is composed of the following fields:

▪ corpus_text_id (an autoincremented value)

▪ filtered_content

▪ original_content (only kept for reference)

▪ internal_text_id

Considering the structure mentioned above, a text

can be fully identified and categorized by joining the

corpus texts and metadata entities.

The n-grams are stored in an entity defined by the

following fields:

▪ ngram_id (an autoincremented value)

▪ n_gramity (can have as possible values 2, 3, 4

or 5, representing the number of words in the

n-gram)

▪ n-gram (the exact words combination)

▪ corpus_genre_id (an external reference to the

corpus_genres entity)

▪ corpus_discipline_id (an external reference to

the corpus_study_levels entity)

▪

corpus_gender_id (an external reference to the

corpus_gender entity)

▪ corpus_study_level_id (an external reference

to the corpus_study_level entity)

Considering this structure, an n-gram can be

identified, categorized, and counted.

The statistics are memorized in a database entity

with the following fields:

▪ statistics_id (an autoincremented value)

▪ romanian_texts_counter

▪ english_texts_counter

▪ romanian_words_counter

▪ english_words_counter

▪ romanian_characters_counter

▪ english_characters_counter

▪ 2, 3, 4, 5 -grams counter

Being pre-computed and stored in this format, the

counters are easy to extract and to display both in the

administration and the user interfaces.

3.1.6 Big Data Processing Performance

Since our system involves manoeuvring large

amounts of data, the times measured for storing and

retrieving the data can make the difference between a

responsive system and a platform jammed by

background operations.

During the development process we worked with

a set of 50 texts, and we passed them through all the

ICSOFT 2022 - 17th International Conference on Software Technologies

394

processing phases. Inserting and selecting data from

the database had very small execution times, but once

we increased the volume of processed data, some

processing steps had to be rewritten.

The corpus text files upload depends only on the

size of each file and storing them on the server is a

very direct process that leaves nothing to the

optimization. Normalizing the content, processing,

and storing it in the database takes up to 2 seconds

for an average of 1571 words per file.

The metadata file corresponding to a set of 500

texts takes under 2 seconds for being uploaded and

under 10 seconds for parsing and processing. Each

row from the metadata file is transformed into an

insert or update in the metadata entity, and if

necessary, in corpus_genres, corpus_disciplines,

corpus_genders and corpus_study_levels. The

execution times are good considering the amount of

processed information.

The n-grams are isolated from an input text in less

than 1 second/text. An average of 105.17 n-grams

(pairs of 2, 3, 4 and 5 words) are detected per text file.

Inserting all the word pairs individually in the n-

grams database entity takes an average time of 5

seconds / text. Cumulated with the number of texts,

this leads to an average n-gram storing time of 5,000

seconds / 1000 files. This value represents a true

problem, and the solution came from cumulating the

insert operations into bulk inserts. All the insert

syntaxes are grouped by 400 and the execution times

were decreased dramatically, reaching the lowest

average of less than 2 seconds.

The obtained execution times and the associated

number of bulk inserts are represented in Figure 1.

Figure 1: The number of bulk inserts / execution time.

Counting the n-grams in a table that is composed

of several million records is a problematic issue.

Determining the number of times a certain group of

words occurs involves SQL syntaxes similar to this

one.

SELECT count(n_gram) as occurences,

n_gram FROM `table_name` WHERE

conditions group by n_gram order by

occurences desc;

At first, for an efficient memory usage, the

n_gram field was declared as varchar(50) […]. The

obtained execution times for such a routine took

approximately 70 seconds for a table containing

3,500,000 records. Obviously, this time interval is

totally un-practical, so we had to come up with a

better solution.

The first part of the solution was to renounce the

efficient memory usage in favour of the syntax

rapidity. The n_gram field was declared as char(50)

[…] and the response times dropped immediately to

an average of 25 seconds.

It’s a known fact that SQL syntaxes involving

SELECT and COUNT work much faster on

numerical values than on arrays of characters. A very

direct numerical representation of an array of

characters is given by applying a hashing function.

Having those values stored in the database and

rewriting the syntax as

SELECT count(hash) as occurences,

FROM `table_name` WHERE conditions

group by hash order by occurences

desc;

reduced the execution times to an outstanding 0.9

seconds for the same table with 3,500.000 records.

A small inconvenience appeared after this initial

success and was due to the necessity of including the

n_gram in the search. The following SQL syntax was

used with similar results.

select hash, n_gram, count(hash)

from `table_name` group by hash

order by count(hash) desc;

select T1.hash as hash_value,

count(T1.hash) as occurences,

T2.n_gram from `table_name` as T1

join `table_name` T2 on T1. id=T2.id

group by T1.hash order by occurences

desc

Both syntaxes averaged a satisfying 4.5-5 seconds

for 3,500,000 records. From these recorded times,

approximately 2.5 seconds are produced by ordering

the results. We tried to get the results unordered and

to sort them programmatically, but the processing

times increased with an average of 5 seconds so we

adopted the database implicit sorting mechanism.

Building Roger: Technical Challenges While Developing a Bilingual Corpus Management and Query Platform

395

4 CORPUS INTERFACES

The Roger Corpus platform aims to be very close to

the user and to trigger extensive background

computations with simple clicks.

4.1 Administration Interface

After authentication, an administrator can have access

to the following functionalities.

Texts and metadata management, where the

administrator can:

▪ Upload text files into the platform; this process

is followed by normalization, processing, and

database storage; an extensive report is

displayed to the administrator.

▪ Display texts; each text can be fully viewed,

edited and deleted.

▪ Set texts displaying criteria; the filtering can be

done by id, language, genre, discipline, study

level, study year and author gender;

▪ Upload metadata files; each upload is followed

by parsing and database storing operations in

the corresponding entities.

▪ View metadata; a synthesis of genres,

disciplines, study levels and genders is

displayed; the metadata ids without matching

texts and the texts without paired metadata is

displayed.

Overall statistics, where the administrator can:

▪ Generate overall statistics: by a press of a

button, the texts, texts by discipline, texts by

language, characters by language, words by

language and n-grams counters are computed;

the date the statistics were generated is

determined and displayed.

▪ View overall statistics: the values mentioned

above are displayed in web format.

Detailed n-grams statistics, where the administrator

can:

▪ View n-grams statistics: the n-grams can be

filtered by language, genre, discipline, study

level, study year, and author gender; the first

100 n-grams are displayed, and the full list of

n-grams (thousands of records) can be

downloaded in *.xlsx format.

4.2 User Interface

Any visitor that accesses the Roger Corpus platform

can benefit from accessing the following areas.

▪ About: a short description of the platform.

▪ Corpus documentation: a selection of

downloadable documents.

▪ Tutorials: Corpus user guides.

▪ Research: an area for publishing various results

related to the linguistic research performed by

the academic staff that manages the platform.

▪ Statistics: an area in which the texts, words,

characters and n-grams statistics are displayed.

▪ Contact

▪ Terms and conditions

After creating an account, in addition to the visitor

options, a registered user can:

▪ Perform Corpus searches: by mentioning a

sequence of words, the corresponding texts are

displayed as a list; the user can refine the initial

search by filtering the results by language,

genre, and discipline.

▪ Generate and download specific n-grams

statistics: by specifying the searched number of

words, language, genre, discipline, study level,

study year the first 100 n-grams are displayed;

the results can be downloaded in *.xslx format;

the user has to manifest his agreement to the

Roger Corpus platform’s policy before

downloading the selected n-grams;

5 CONCLUSIONS

This article presents a Corpus application that offers

the tools for handling a large collection of texts

written in English and Romanian. ROGER aims to

expose the accumulated database and to offer to the

registered users the possibility of filtering the stored

information by language, genre, discipline, author’s

study year and learning cycle.

Both the English and the Romanian sub-corpora

feature texts from eight different disciplines: (i)

Humanities; (ii) Economics; (iii) Political Sciences;

(iv) Engineering; (v) Computer Science; (vi) Law;

(vii) Mathematics; (viii) Social Sciences. In each

discipline, the students labelled the genre of their own

writings as follows: (i) Essay; (ii) Scientific paper;

(iii) Report; (iv) Bachelor thesis (BA); (v) Master’s

dissertation (MA); (vi) Case study; (vii) Summary;

(viii) Literary analysis; (ix) Review; (x) Others (to be

elaborated further). While samples of most genres can

ICSOFT 2022 - 17th International Conference on Software Technologies

396

be found in both languages, there are genres which

have only been collected either in English (e.g., CV,

interview, documentation etc.) or Romanian (e.g.,

summary, reading notes, portfolio etc.).

The exposed data serves the purpose of

investigating student writing practices in both their

mother tongue and English as a Foreign Language.

The platform was designed and implemented as a

cross-platform distributed web application.

The backend interface available to authenticated

administrators provides the digital tools for managing

the database stored texts and associated metadata.

Also, it offers an extensive statistics mechanism that

covers the corpus data composition, distribution, and

usage. The quantified aspects target the words,

characters, languages, study levels, genres, domains,

and n-grams.

The frontend capabilities are offered to the

registered users allowing them to search for specific

keywords and to refine the obtained results by

applying a series of filters.

As a technical challenge, manoeuvring high

volume data was a serious problem generator. The

registered processing times were sometimes bigger

than a regular web application can afford. A series of

optimizations were applied, including processing

fragmentation, SQL syntaxes cumulation and

database optimization techniques.

Overcoming the challenges brings additional

value to the resulting instrument, which becomes the

first European open-access corpus support platform.

The platform offers user-friendly search options to a

predefined original corpus (ROGER), whose

compilation itself is a major contribution to the

academic writing research community. The fact that

the corpus is bilingual (Romanian and English),

multi-disciplinary, multi-genre and multi-level makes

it a valuable asset for any interested user that can

access the platform for personal use, teaching,

research or professional development.

ACKNOWLEDGEMENTS

The ROGER platform has been developed in the

framework of the ROGER project

(https://roger.projects.uvt.ro/), developed at the West

University of Timisoara, Romania. ROGER was

financed by the Swiss National Science Foundation,

as part of the PROMYS program (project code

IZ11Z0_166537), for a five-year period (2017-2022),

through the research grant awarded to the project

coordinator, Dr Madalina Chitez.

REFERENCES

Ackerley, K. (2021). 4 Exploiting a genre-specific corpus

in ESP writing: students’ preferences and strategies In

M.Charles, & A. Frankenberg-Garcia (Eds.), Corpora

in ESP/EAP Writing Instruction: Preparation,

Exploitation, Analysis (pp. 78-99). New York:

Routledge.

Alsop, S., & Nesi, H. (2009). Issues in the development of

the British Academic Written English (BAWE) corpus.

Corpora, 4(1), 71-83.

AWICNET (2019), accessed March 1, 2022,

https://roger.projects.uvt.ro/news/aprilie-2019-lansare-

proiect-awicnet/

Callies, M. & Zaytseva, E. (2013a). The Corpus of

Academic Learner English (CALE) – A new resource

for the study and assessment of advanced language

proficiency. In Granger, S., Gilquin, G., & Meunier, F.

(Eds.), Twenty Years of Learner Corpus Research:

Looking back, Moving ahead (Corpora and Language

in Use - Proceedings Vol. 1). Louvain-la-Neuve:

Presses universitaires de Louvain.

Callies, M. & Zaytseva, E. (2013b).The Corpus of

Academic Learner English (CALE) – A new resource

for the assessment of writing proficiency in the

academic register. Dutch Journal of Applied

Linguistics, 2(1), 126-132.

Chitez, M., Rogobete, R., & Foitoş, A. (2020). Digital

Humanities as an Incentive for Digitalisation Strategies

in Eastern European HEIs: A Case Study of Romania.

In Curaj, A., Deca, L. & Pricopie, R. (Eds.), European

Higher Education Area: Challenges for a New Decade

(pp. 545-564). Springer: Cham.

Chitez, M., Bercuci, L., Dincă, A., Rogobete, R., & Csürös,

K. (2021). Corpus of Romanian Academic Genres

(ROGER). Roger-corpus.org. Retrieved [20, February,

2022], from https://roger-corpus.org/.

Chitez, M., & Bercuci, L. (2019). Data-driven learning in

ESP university settings in Romania: multiple corpus

consultation approaches for academic writing support.

In: F. Meunier, J. Van de Vyver, L. Bradley, L. & S.

Thouesny (Eds). CALL and complexity–short papers

from EUROCALL2019 (pp. 75-81), research-

publishing.net.

Chitez, M. (2014). Learner corpus profiles: the case of

Romanian Learner English. Linguistic Insights Series,

Vol. 173 (Series Editor: Mau-rizio Gotti). Bern, Berlin,

Bruxelles, Frankfurt am Main, New York, Oxford,

Wien: Peter Lang.

Donahue, C. (2002). The lycée to university progression in

French students’ development as writers. In Foster, D.,

& Russell, D. (Eds), Writing and Learning in Cross-

national Perspective: Transitions from Secondary to

Higher Education (pp. 134-191). Urbana, IL: National

Council of Teachers of English (NCTE) Press.

Donahue, C. (2009). "Internationalization" and

Composition Studies: Reorienting the Discourse.

College Composition and Communication, 61 (2), 212-

243. Available at: http://www.jstor.org/stable/

40593441.

Building Roger: Technical Challenges While Developing a Bilingual Corpus Management and Query Platform

397

Foster, D. (2006). Writing with authority: students' roles as

writers in cross-national perspective. Carbondale:

Southern Illinois University Press.

Foster, D., & Russell, D. R. (2002). Writing and learning in

cross-national perspective: Transitions from secondary

to higher education. New Jersey: Lawrence Erlbaum.

Granger, S. (1996). From CA to CIA and back: An

integrated approach to computerized bilingual and

learner corpora. Lund Studies in English, 88, 37–52.

Heuboeck, A., Holmes, J., & Nesi, H. (2009), The BAWE

Corpus Manual version II. http://www.reading.ac.uk/

internal/appling/bawe/BAWE.documentation.pdf

(accessed 20 February 2022).

Johansson, S. (2003). Contrastive linguistics and corpora.

In Granger, S., Lerot, J., & Petch-Tyson, S. (eds.),

Corpus-based Approaches to Contrastive Linguistics

and Translation Studies (pp. 31-44). Brill.

Kaiser, D. (2003). Nachprüfbarkeit versus Originalität.

Fremdes und Eigenes in studentischen Texten aus

Venezuala und Deutschland. In Ehlich, K., & Steets, A.

(Eds.), Wissenschaftlich schreiben – lehren und lernen

(pp. 305-325). Berlin: Walter de Gruyter.

Kilgarriff, A., Baisa, V., Bušta, J., Jakubíček, M., Kovář,

V., Michelfeit, J., Rychlý, P. & Suchomel, V. (2014).

The Sketch Engine: ten years on. Lexicography, 1(1),

7-36.

Kwon, H., Partridge, R. S., & Staples, S. (2018). Building

a local learner corpus: Construction of a first- year ESL

writing corpus for research, teaching, mentoring, and

collaboration. International Journal of Learner Corpus

Research, 4(1), 112-127.

McEnery, T. (2019). Corpus linguistics. Edinburgh

University Press.

Nesi, H. (2008). BAWE: an introduction to a new resource.

In A. Frankenberg- Garcia, T., Rkibi, M. Braga da

Cruz, Carvalho, R., Direito, C., & Santos-Rosa, D.

(Eds). Proceedings of the Eighth Teaching and

Language Corpora Conference (pp. 239–46). Lisbon,

Portugal: ISLA.

O’Donnell, M.B. & U. Römer. (2012). From student hard

drive to web corpus (Part 2): the annotation and online

distribution of the Michigan Corpus of Upper-level

Student Papers (MICUSP). Corpora, 7(1), 1–18.

Paquot, M., Hasselgård, H., & Oksefjell Ebeling, S. (2013).

Writer/reader visibility in learner writing across genres:

A comparison of the French and Norwegian

components of the ICLE and VESPA learner corpora.

In S. Granger, G. Gilquin, & F. Meunier, Twenty Years

of Learner Corpus Research (Corpora and Language in

Use - Proceedings; 1), Presses Universitaires de

Louvain: Louvain-la-Neuve. Retrieved from

https://www.uclouvain.be/cps/ucl/doc/cecl/documents/

Paquot_al_LCR2011.pdf.

Römer, U., & O'Donnell, M. B. (2011). From student hard

drive to web corpus (part 1): The design, compilation

and genre classification of the Michigan Corpus of

Upper-level Student Papers (MICUSP). Corpora, 6(2),

159-177.

Siepmann, D. (2006). Academic writing and culture: An

overview of differences between English, French and

German. Meta, 51, 131-150.

Trumpp, E. C. (1998). Fachtextsorten Kontrastiv:

Englisch-Deutsch-Französisch. Gunter Narr Verlag.

ICSOFT 2022 - 17th International Conference on Software Technologies

398