Search

The use of automatic methods for the study of lexical semantic change (LSC) has led to the creation of evaluation benchmarks. Benchmark datasets, however, are intimately tied to the corpus used for their creation questioning their reliability as well as the robustness of automatic methods. This contribution investigates these aspects showing the impact of unforeseen social and cultural dimensions. We also identify a set of additional issues (OCR quality, named entities) that impact the performance of the automatic methods, especially when used to discover LSC.

Automatic subject prediction is a desirable feature for modern digital library systems, as manual indexing can no longer cope with the rapid growth of digital collections. This is an “extreme multi-label classification” problem, where the objective is to assign a small subset of the most relevant subjects from an extremely large label set. Data sparsity and model scalability are the major challenges we need to address to solve it automatically. In this paper, we describe an efficient and effective embedding method that embeds terms, subjects and documents into the same semantic space, where similarity can be computed easily. We then propose a novel Non-Parametric Subject Prediction (NPSP) method and show how effectively it predicts even very specialised subjects, which are associated with few documents in the training set and are not predicted by state-of-the-art classifiers.

In this paper we discuss the role of the Nanopublication (nanopub) model for scholarly publications with particular focus on the citation of nanopubs. To this end, we contribute to the state-of-the-art in data citation by proposing: the nanocitation framework that defines the main steps to create a text snippet and a machine-readable citation given a single nanopub; an ad-hoc metadata schema for encoding nanopub citations; and, an open-source and publicly available citation system.

Named Entity Recognition (NER) for rare long-tail entities as e.g., often found in domain-specific scientific publications is a challenging task, as typically the extensive training data and test data for fine-tuning NER algorithms is lacking. Recent approaches presented promising solutions relying on training NER algorithms in an iterative weakly-supervised fashion, thus limiting human interaction to only providing a small set of seed terms. Such approaches heavily rely on heuristics in order to cope with the limited training data size. As these heuristics are prone to failure, the overall achievable performance is limited. In this paper, we therefore introduce a collaborative approach which incrementally incorporates human feedback on the relevance of extracted entities into the training cycle of such iterative NER algorithms. This approach, called Coner, allows to still train new domain specific rare long-tail NER extractors with low costs, but with ever increasing performance while the algorithm is actively used in an application.

International audience; In this paper, we focus on epidemic event extraction in multilingual and low-resource settings. The task of extracting epidemic events is defined as the detection of disease names and locations in a document. We experiment with a multilingual dataset comprising news articles from the medical domain with diverse morphological structures (Chinese, English, French, Greek, Polish, and Russian). We investigate various Transformer-based models, also adopting a two-stage strategy, first finding the documents that contain events and then performing event extraction. Our results show that error propagation to the downstream task was higher than expected. We also perform an in-depth analysis of the results, concluding that different entity characteristics can influence the performance. Moreover, we perform several preliminary experiments for the low-resourced languages present in the dataset using the mean teacher semi-supervised technique. Our findings show the potential of pre-trained language models benefiting from the incorporation of unannotated data in the training process.

This paper describes utilization of article detection and extraction on the Finnish Digi (https://digi.kansalliskirjasto.fi/etusivu?set_language=en) newspaper material of the National Library of Finland (NLF) using data of one newspaper, Uusi Suometar 1869–1918. We use PIVAJ software [1] for detection and marking of articles in our collection. Out of the separated articles we can produce automatic clippings for the user. The user can collect clippings for own use both as images and as OCRed text. Together these functionalities improve usability of the digitized journalistic collection by providing a structured access to the contents of a page.

Word-clouds are a useful tool for providing overviews over texts, visualising relevant words. Multiple word-clouds can also be used to visualise changes over time in a text. This requires that the words in the individual word-clouds have stable positions, as otherwise it is very difficult so see what changed between two consecutive word-clouds. Existing approaches have used coordinated positioning algorithms, which do not allow for their use in an online, dynamic context. In this paper we present a fast word-cloud algorithm that uses word orthogonality to determine which words can share the same space in the word-clouds combined with a simple, but fast spiral-based layout algorithm. The evaluation shows that the algorithm achieves its goal of creating series of word-clouds fast enough to enable use in an online, dynamic context.

In this study, we have normalized and lemmatized an Old Literary Finnish corpus using a lemmatization model trained on texts from Agricola. We analyse the error types that occur and appear in different decades, and use word error rate (WER) and different error types as a proxy for measuring linguistic innovation and change. We show that the proposed approach works, and the errors are connected to accumulating changes and innovations, which also results in a continuous decrease in the accuracy of the model. The described error types also guide further work in improving these models, and document the currently observed issues. We also have trained word embeddings for four centuries of lemmatized Old Literary Finnish, which are available on Zenodo.