Increasing numbers of students enrol in formal and informal courses taught in a foreign language. Studying a course from an unfamiliar university/program is di cult enough, but the di culties multiply when the transition to new course requirements is exacerbated by the necessity to learn course material in a foreign language. This paper describes Interlingua a platform where students can study textbooks in a foreign language supported by on-demand access to relevant reading material in their mother tongue. Interlingua automatically recognises important terminology within textbooks content, extracts structural models of textbooks and links sections and subsections across textbooks in di erent languages covering the same academic subject. The interface and architecture of Interlingua as well as the technologies underlying the platform are described.
1.1
"Inter-lingual" students
Two parallel trends exist in the current EU education, independent one from
another, originating from di erent conditions, yet leading to a shared outcome.
From the socio-economic perspective, EU promotes ever-increasing mobility,
especially when it comes to younger population. The Bologna process [
From the pedagogical (and technological) perspective, new forms of learning have emerged supported by information and communication technologies. They facilitate free and easy access to learning materials and promote more central and active role of a learner. Initiatives like Open Educational Resources (OER) and phenomena like Massive Open Online Course (MOOC) shape the new educational reality where students have more choice and exibility in terms of which textbook to read, which course to take, and which skill to acquire. As a result they become less dependent on the actual institution issuing a degree.
These two trends reinforce each other and jointly contribute to a muchdesired outcome of more scalable, sustainable and a ordable education. However, they also lead to a potential problematic situation that is occurring more often as more international students enrol in formal university courses and/or free MOOCs taught in a foreign language. Studying a course from an unfamiliar university/program is challenging enough. It might be taught on a new (more abstract or intensive) level. It might require a student to have prerequisite knowledge and skills that s/he has not acquired yet. Yet for foreign students, this transition to new course requirements is aggravated by the necessity to learn material in a foreign language that they did not use when taking the prerequisite courses. A foreign student inevitably faces a certain language barrier ampli ed by the mismatch in the background knowledge and terminology. Unfortunately, the current tradition of resolving these di culties is hardly e cient - teachers report that non-native students are allowed to bring dictionaries to regular classes and exams. As a result, we have an educational system that promotes student mobility on the level of policies, but does not su ciently support it in on the individual level.
An e ective remedy to this problem is the provision of international students with multilingual access to instructional material, where educational resources in a language of a course are accompanied by resources in their native language. There are two principle ways to achieve this: translation of the original resource and linking between two corresponding "inter-lingual" resources. The translation-based approach will not help solve the problem: manual translation is not scalable, and machine-based translation is not yet capable of producing results of adequate quality in a narrow academic domain. In this paper, we present a solution based on automated semantic linking of related educational resources across languages with the main focus on textbooks1. Probability theory and statistics has been chosen as the target domain. One reason for it is the popularity of this subject in teaching programs of many technical and social science degrees. Another reason is that this subject uses a lot of speci c terminology both new and borrowed from prerequisite parts of mathematics. 1.2
Textbooks are often considered non-structured information resources for the purpose of text analysis and information extraction. Yet, textbooks are created, structured and formatted by their authors (who are presumably domain experts) with a primary purpose to explain the knowledge in the domain to a novice. A textbook author uses his/her own understanding of the domain when structuring and formatting the content of the textbook to facilitate this explanation. As a result, the formatting and structural elements of a textbook (headers, table of
DeLux 32274)
content (ToC), index) not only shape the organisation of the textbook, but also
re ect the organisation of the domain as the author sees it. When extracted they
can be formally represented as a semantic model of the textbook itself and the
domain it teaches. We have experimented with automated extraction of such
models from multiple textbooks in di erent domains (probability theory and
statistics and information retrieval) and di erent representation formats (PDF
and ePUB) [
A textbook model can support a variety of tasks including inference and
reasoning about knowledge taught bu the textbook, tracking and modelling
of reader's progress with the textbook, enhanced interaction through
semantic search and adaptive navigation through the textbook. In the case of multiple
textbooks in the same domain, such models can be mapped between each other
providing a semantic bridge for linking related sections and fragments across
textbooks. An external reference model can be used to facilitate the mapping
and help resolving possible terminology con icts. In [
The rest of this paper is structured as follows. Section 2 gives an overview of related work in the eld of textbook model extraction and linking. Section 3 presents the details of the proposed technology and the architecture of Interlingua - the Web-based platform that students can use to access textbooks in a foreign language while getting recommendations of related reading resources in their mother tongue. Section 4 presents the interface of Interlingua and the ways it can support students. Section 5 concludes the paper with a discussion and a description of future work. 2
Creation of knowledge models from textbooks has been explored in a limited
way. Larrn~aga et al. [
Several projects experimented with semantic linking of relevant textbooks
written in the same language. Guerra et al. [
Architecturally, Interlingua consists of two large components. The o ine component performs the tasks of textbook modelling and linking, while the online component supports students' interaction with the content of linked textbooks. This section describes the details of the o ine component; the student interface is presented in the next section.
The overall process of adding a new textbook into the Interlingua content
repository is depicted by 1. The only action performed manually is the upload
of textbook les. A teacher decides which textbooks and in which languages
should be available for the students of the course. After a textbook le is
submitted, the textbook model is generated: the textbook is divided into sections
and sub-sections, its ToC and index are extracted, each page is identi ed, every
di erently formatted fragment is recognised and provided with a semantic label
if a corresponding rule exists (we keep working on expanding the rule base).
Index plays a special role in this process, as it provides a glossary of manually
selected terms that the author of the textbook deemed meaningful. Interlingua
extracts index terms and pages referenced by the terms and uses them as the
semantic anchors to link pages and sections of the textbook to the concepts of
the reference ontology and through them to other textbooks available in the
content repository. Finally, the self-assessment component uses the information
from DBPedia to generate multiple choice questions related to the index terms
explained by the current section of the textbook. Once processed and stored this
way, the textbook becomes accessible through the student interface. The rest of
the section describes individual phases of this process; however, we refer readers
to our two other papers for a detailed description of textbook modelling [
Raw content extraction. In this step, the PDF le of the textbook is parsed to extract the text (the characters), geometrical (X- and Y- coordinates on the pages) information and formatting (type, size, and features of fonts) styles of the document. The parsing of the PDF is done with the Apache PDFBox library2.
The parser reads the PDF content stream where the elements are sorted from left to right and top to bottom according to their position on a page. It merges smaller objects into bigger using a bottom-up approach. First, individual characters are processed. Then, characters are grouped into words, words into lines, and lines into pages. At each stage, the extractor compares the proximity of an object's coordinates with the coordinates' of its preceding neighbour to form a bigger object. Besides, the content itself, the parser also saves the styling information if every object, such as font size, family, properties (boldness, colour, etc.) as well as margins.
Style library construction. Once all text fragments are extracted and their styles are identi ed, the library of all the formatting styles used in the textbook is constructed. This library is used in the next step to recognise the di erent structural elements constituting the textbook. The main style of the textbook content is identi ed, as well as styles for important text fragments, headings of sections and subsections, etc. Geometric features of the text play a role in this process as well, as they allow to combine text into paragraphs, sections and chapters, allow to compute spacing and indentations, allow to recognise columns and captions, etc.
Logical element identi cation In this step, some auxiliary texts such as
headers, footers and page numbers are recognised and separated from the main
content. In addition, important structural elements of the textbook are
identied. The ToC section is recognised and parsed to construct an outline of the
textbook. Each content section and sub-section is identi ed, associated with
its header, its ToC entry, and its page interval. Finally, the textbook index is
processed. This section plays a special role, as it provides the main source of
information for subsequent textbook linking. A good textbook index is not just
a collection of words, but, essentially, a reference model produced by a domain
expert according to a prede ned set of rules. Every publisher guides the process
of index creation stipulating index length and style, suggesting what can be good
and bad candidates for index terms, advising on how to maintain consistency
when creating hierarchical indices, etc. [
Model construction Finally, with all the extracted information a semantic knowledge model of the textbook is constructed. The model allows to quickly obtained di erent knowledge views of the textbooks to provide rich information to the users, for example: hierarchical structure of sections, all index terms related to one section, all pages where an index term is explained, etc. The model is represented using SKOS3, which is a W3C standard based on RDF and OWL. It provides a model for expressing the basic structure and content of concept schemes such as thesauri, classi cation schemes, subject heading lists, taxonomies and other similar types of controlled vocabulary. 3.2
This component takes the model of a textbook and links it to two external reference models: the ISI Glossary and DBpedia. The ISI Glossary4 was created by The International Statistical Institute. It contains more than 3500 statistical terms combined into synsets and translated into 31 di erent languages. The glossary has been parsed and represented as a reference ontology. As a result, Interlingua can unambiguously recognise that, for example, in English term "zscore" is the same as "standard score". What is even more importantly, we can also identify all the ways this term is called in all 31 languages of the ISI glossary. DBpedia is a machine-understandable source of knowledge created based on automatically-extracted structured information from Wikipedia, which contains 4.58 million resources and is available in 125 languages. It plays two important roles in Interlingua. First, it provides and alternative domain-independent multilingual reference model. Although, ISI glossary has a much better granularity and coverage statistical terms, DBPedia ensures that the overall approach of Interlingua is applicable to other domains (and other languages) where a dedicated high-quality multilingual thesauri cannot be obtained. The second role of DBPedia is to provide additional structural and annotation information, such as
dbpedia-owl:abstract for term explanations, dcterms:subject for elds of study
where the term is coming from and dbpedia-owl:wikiPageWikiLink for di erent
terms, which are mentioned in the same context, etc. The enrichment process is
shown in Fig.3. More information about the enrichment algorithm is available
in [
Glossary have been (automatically) linked to its correspondent entities in DBpedia. The linking algorithm rst constructs a core set of terms from the glossary for which a matching DBpedia resource was unambiguously discovered. Then, it recursively uses textual content associated with recognised DBPedia entities as the context information to facilitate disambiguation of newly discovered candidate matches. The process repeats itself until no new terms can be match with su cient certainty. This only needs to be done once, before the enrichment of individual textbook models.
index term extracted from the textbook is compared against each term of the ISI Glossary, and when the similarity is over 90 %, the index term is linked to the ISI Glossary term.
a linked ISI Glossary term that has a DBpedia resource, semantic information is extracted to enrich the model of the textbook. The following information is obtained from the linked DBpedia resources: abstracts, links to the corresponding Wikipedia pages of the resources, DBpedia categories of the resources, and direct relations among the linked index terms using the links that exist in the Wikipedia pages of the resources. At the end of this approach, the textbook's model is updated with the gathered information. 3.3
The linking among the di erent parts of textbooks is done based on the ISI glossary as a reference model to build a Vector Space Model (VSM). The VSM consists of several documents (each section) per textbook in the rows, and each entry of the ISI glossary as the columns. The weight for each term of each document is calculated from the strength of the association between the term and the document. The VSM is improved pruning the terms in the glossary that are not found in the corpora. After the VSM has been created, it is split into matrices, one for each language, and a similarity matrix is created for each language pair by rst multiplying the two appropriate parts of the VSM together and then normalising the resulting matrix. These matrices contain the similarity value for any pair of sections of textbooks from di erent languages. Finally, each section of each textbook is linked to the ve most similar documents in each target textbook to create the links among the textbooks in di erent languages. 3.4
The assessment engine generates Multiple Choice Questions (MCQ) for the learners to examine their understanding of the terminologies related to the section that they are currently reading. The basic MCQ asks the learners to select the correct translation of one of the introduced concepts in the current section into their native language. Each question has only one right answer and several distractors. To generate the questions the assessment engine uses the ISI glossary to get the labels in di erent languages for the same concept. Distractors are chosen using the relations among concepts obtained from the enrichment of the model to select reasonably di cult distractors. In other words, the semantic distance between the correct answer and a distarctor in the model should not be very long and the lengths of both textual labels should be comparable.
When a student accesses the Interlingua client, the entry page shows a list of
available textbooks in the target language of study. A student can indicate her
language of study and the mother tongue. Currently, Intelingua uses the
following textbooks: [
A student selects one of the available textbooks to open the textbook navigation page (see Fig. 4). It consists of the outline panel on the left side and the content panel with multiple tabs on the right side of the window.
In the outline panel, a student can brows through sections and sub-sections of the textbook. She can click on a magnifying glass icon annotating each section title to load the respective section into the content panel. The arrow icon is used to display a list of related readings of the selected sub-section in the mother tongue of the student (see Fig. 5).
The student can browse through a subsection content in a similar way to any standard PDF viewer application. Index terms are recognised and highlighted in the content to indicate that additional interaction with them is possible. When a student clicks in a highlighted word, an action menu appears with three available options: translate, explain, and assess (see Fig. 6). The translate action opens a pop-up window that presents the translation of the term into the mother tongue of the student, and the de nition of the term extracted from DBpedia (if this index term has been found in DBpEdia). The explain action opens a pop-up window that gives access to the sections of the textbooks in the mother tongue of the user in which the term is explained. The assess action generates and displays in a pop-up window an question about the translation of the term into the mother tongue of the user. As mention before, in the navigation page a student also can open related readings for the current section in her mother tongue (see Fig. 4). A pop-up window suggests a list of related readings (see Fig. 5). if a student selects any of the suggested links, a corresponding sub-section will be loaded in a new tab. The student can switch between the reading tabs that now contain the related content in two di erent languages.
Finally, the user can click on the 'Assessment' tab in the content panel to generate an assessment composed of several MCQ related to the content of the currently browsed section (the questions are similar to the one at the bottom pop-up in Fig 6).
The paper has presented the motivation, the approach and the implementation of the Interlingua system that provides foreign students with on-demand access to related textbook material in di erent languages.
There might be several possible concerns both practical and pragmatic about the scale and applicability of this approach. One question that needs to be addressed is the availability of textbooks and the copyrights issues. In our experience, university libraries can supply enough PDF-based textbooks on a variety of subjects. From the point of copyright protection, if a system provides enhanced access to these books but only to the students of the university holding necessary subscriptions, then publishers do not have a reason to object. In the worst case scenario, many good-quality textbooks are freely available online nowadays in open repositories such as Openstax5/Connections6, Open Textbook Library7, OER-Commons8, etc.
Another concern is how well the automated textbook modelling technology will cope with di erent textbook formatting. Naturally, formats can be quite di erent between di erent textbooks, yet the authors and the publishers make every e ort to ensure they are consistent within textbooks (otherwise they would cause unnecessary confusion) and we believe, that the formatting patterns of textbooks do follow a set of logical rules. When implementing the rule-based component of the model extraction algorithm we try to capture these patterns. We have observed that the variability is quite manageable and as we have been processing more and more textbooks, we update our set of rules progressively less.
Finally, we realise that none of the extracted models is guaranteed to provide high-quality representation of a domain. These models can potentially su er from several drawbacks: (1) Subjectivity: they can contain terms are only marginally related to an objective picture of the domain semantics; (2) Coverage: they can miss important terms if a textbook does not cover (enough) a particular part of a domain; (3) Granularity: the terms in an index can be too detailed, too broad, or inconsistently alternate in their granularity; (4) Lack of semantics: most indices are at lists of terms with no relations between them. To combat these potential problems, glossaries extracted from individual textbooks can be integrated between each other and to external models: global as DBPedia, or domain-focused as ISI glossary. We plan to further investigate these and other issues. For example, an important direction for future work is to evaluate the e ectiveness of the system in a user study with real students from a target group.
From a broader perspective, Interlingua is an interesting example of a service that can be built on top of linked semantic models extracted from related textbooks. However, a corpus of semantically linked high-quality educational
content can be used to implement a range of di erent services including adaptive navigation through or recommendation of textbook content, enrichment of textbook content with external (interactive) educational resources, or extraction of di erent types of learning objects from the textbook themselves.
Finally, it is interesting to explore thew applicability of the the approach towards automated modelling and linking of textbook that underlies Interlingua in other, less formal domains (e.g., medicine) or domains with con icting viewpoints (e.g. history).