In this Tour de CLARIN blog post, we present an in-depth interview with Kaja Dobrovoljc, a Slovenian corpus linguist who works at the Centre for Language Resources and Technologies at the University of Ljubljana and regularly collaborates with CLARIN.SI and uses its infrastructure. The interview was conducted via e-mail.
1. Could you please introduce yourself – your research background, your national and international research networks and current projects?
I am a linguist with an undergraduate degree in translation studies and a doctoral degree in Slovene linguistics, awarded in 2018. As a researcher at the Centre for language resources and technologies of University of Ljubljana, my main research interests lie in the design, annotation and evaluation of machine-readable language resources, and their use in descriptive language research. I am currently also involved in two nationally-funded projects aimed at setting up the methodological foundations for a corpus-based grammar of Slovene (in collaboration with the Jožef Stefan Institute) and an interactive online portal for Slovene language learning (in collaboration with the University of Maribor).
2. How did you get involved with CLARIN.SI? How has CLARIN.SI supported your research? How have the results of your collaboration contributed to your research community?
Most of the projects I have collaborated on so far have been dedicated to publishing their results under open licenses to be freely available to anyone interested in their use and further modification. The establishment of the CLARIN.SI consortium in 2013 and the creation of the CLARIN.SI repository that followed soon after was therefore a very welcome addition to the Slovenian language infrastructure in general. On the one hand, it has enabled me and my colleagues to publish and disseminate fundamental language resources, such as the Sloleks morphological lexicon, the ssj500k training corpus or the Thesaurus of modern Slovene in a stable online repository with long-term technical support and assistance. On the other hand, I have also benefited from the ease of access to resources developed by others, such as the GOS corpus of spoken Slovene and the JANES corpus of computer-mediated Slovene, the key language resources in my PhD research on the usage of speech-specific discourse markers in online communication.
In addition to the repository, CLARIN.SI also provides several online services, such as the noSketchEngine web concordancer and the WebAnno annotation tool. I find these particularly useful in my everyday linguistic research, and was therefore happy to join CLARIN.SI’s initiative to organize hands-on training sessions for other researchers within the community as well. As the secretary of the Slovenian Language Technologies Society, I am also very grateful and proud of CLARIN.SI’s continuing support of JOTA, a monthly series of talks held by Slovenian and foreign researchers on topics related to languages technologies, which are also accessible online.
3. Despite being an early-career researcher, you’re one of the most prolific contributors of resources to the CLARIN.SI repository. Among others, you’ve created several sets of n-grams from various Slovene corpora. Could you discuss the importance of these resources for your own research as well as for the research community?
Although the lists of frequently recurring sequences of words in a language (also known as word n-grams) have traditionally been associated with the domain of natural language processing where they are used in language modelling and other computational tasks, these sequences are gaining increasing importance in linguistics as well. In addition to the most commonly studied groups of expressions, such as idioms and collocations, the lists of n-grams with outstanding frequency of usage (also known as formulaic sequences or lexical bundles) reveal an abundance of other multi-word expressions that are not necessarily fixed and idiomatic in the traditional phraseological sense, such as the expressions te dni ‘these days’, v sodelovanju z ‘in collaboration with’, po drugi strain pa ‘but on the other hand’ in written, or ali pa nekaj takega ‘or something like that’, gremo naprej ‘let’s move on’, veš kaj ‘you know what’ in spoken Slovenian. Phrases like these often seem uninteresting and self-evident to native speakers of a language, but they have been shown to have a special cognitive status in our brain nevertheless, and are also one of the key indicators of native-like fluency in language learners.
In my PhD work, I was mostly interested in formulaic sequences that contribute to discourse organization in spoken Slovenian. However, I applied the same extraction tool to several other reference corpora, such as written, computer-mediated and historical Slovene, producing the lists of most frequently recurring words, lemmas, POS tags and other feature combinations with two kinds of frequency counts. These open the way to numerous interesting explorations of the nature and use of formulaic expressions in the future in various linguistic disciplines, from language teaching and lexicography to psycholinguistics and diachronic language studies.
4. You’ve also been part of the team that created the manually annotated ssj500k corpus. Could you describe your role in the creation and annotation of the corpus? Why is this corpus important for Slovenian linguistics?
In a way, this corpus has been pivotal to my career as a researcher, as I first came into contact with language resources and technologies as a student annotator, checking for tokenization, lemmatization and tagging mistakes performed by the automatic morphosyntactic tagger. In subsequent projects, I continued working on this dataset by manual annotation of surface syntax with the JOS dependency labels and their subsequent conversion to the complementary Universal Dependencies scheme. In addition to these layers of linguistic annotation, ssj500k has also been annotated for named entities, semantic role labels and multi-word expressions. With more than 500,000 tokens or 27,000 sentences in total, ssj500k is the largest and most extensively manually annotated corpus of Slovenian and thus an invaluable resource for the development of fundamental language technologies, such as tokenizers, lemmatizers, taggers and parsers, which build their knowledge of the Slovenian language by observing its behaviour in such datasets. At the same time, this resource has had an important impact on Slovene linguistics as well, since many of the traditional linguistic categorizations of language phenomena in Slovenian had to be re-evaluated and improved in the annotation process, not only to meet the specific needs of machine-based applications, but also to enable systematic application to large amounts of authentic, real-world language data.
5. Together with Joakim Nivre you have worked on annotating the Treebank of Spoken Slovenian following the Universal Dependencies framework. What are the benefits of the Universal Dependencies framework and why is it important for Slovene to be part of the initiative? What are the challenges of creating a treebank of spoken language data? Why is it important for Slovene linguists and the society at large to have access to a treebank of the spoken language?
Universal Dependencies is an international initiative aimed at a cross-lingually consistent annotation scheme for morphological and syntactic annotation, which has already been applied to over 100 treebanks in more than 80 languages, including the written and spoken treebanks of Slovenian. Harmonizing the annotation of linguistic phenomena that are similar across languages has many important advantages for language technologies, since it enables the development of multilingual tools, such as taggers and parsers, and promotes consistent cross-lingual language technology research and evaluation in general. Many of these benefits are already visible, as several state-of-the-art tools have emerged based on this dataset and are directly applicable to all participating languages. This is especially important for small language communities that cannot necessarily afford a continuous development of high-performing language technology tools, especially in the era of fast-paced computational progress. At the same time, the large number of treebanks annotated in a unified way offers exciting opportunities for contrastive linguistic research, such as quantitative investigations into typological differences and similarities between different languages or language groups.
This comparative aspect was also the motivation behind the construction of the spoken Slovenian UD treebank, which, in contrast to its automatically converted written counterpart, has been manually annotated from scratch, using the CLARIN.SI WebAnno installation. In the process, many speech-specific phenomena had to be addressed, such as repairs, restarts, hesitations and other types of disfluencies. Interestingly, a comparison of the annotated written and spoken treebanks of Slovenian revealed that it is not just these obvious structural particularities that distinguish speech from writing, but that the two modes also differ in terms of sentence- and phrase-structure in general. For example, spoken data consists of shorter and more elliptic sentences, fewer and simpler nominal phrases, and more relations marking interaction, deixis and modality. Just like the written ssj500k treebank, the Spoken Slovenian Treebank thus represents an important language resource for future explorations in spoken language research and spoken language technologies alike, especially given the fact that it is the spoken language that is the primary and prevalent form of human communication.
6. How can research infrastructures such as CLARIN best serve early-stage researchers and how can they best contribute to the research infrastructure?
Undoubtedly, research infrastructures such as CLARIN represent an invaluable source of easily accessible resources, services and support for early-stage researchers, who are usually restricted to very limited funding and need help navigating the complex landscape of digital language resources. This is certainly the case with CLARIN.SI, where Tomaž Erjavec and his team provide continuing support with language data management, such as help with annotation tools, format conversions and validations, untrivial tasks for researchers in humanities and social sciences with little computational background. At the same time, online repositories, such as the one maintained by CLARIN.SI, offer early-stage researchers a unique chance to publish and disseminate our own research results in a stable online environment, which not only contributes to increased visibility, but also creates opportunities for future collaborations.
Click here to read more about Tour de CLARIN