On cross-lingual retrieval with multilingual text encoders.

Pretrained multilingual text encoders based on neural transformer architectures, such as multilingual BERT (mBERT) and XLM, have recently become a default paradigm for cross-lingual transfer of natural language processing models, rendering cross-lingual word embedding spaces (CLWEs) effectively obsolete. In this work we present a systematic empirical study focused on the suitability of the state-of-the-art multilingual encoders for cross-lingual document and sentence retrieval tasks across a number of diverse language pairs. We first treat these models as multilingual text encoders and benchmark their performance in unsupervised ad-hoc sentence- and document-level CLIR. In contrast to supervised language understanding, our results indicate that for unsupervised document-level CLIR-a setup with no relevance judgments for IR-specific fine-tuning-pretrained multilingual encoders on average fail to significantly outperform earlier models based on CLWEs. For sentence-level retrieval, we do obtain state-of-the-art performance: the peak scores, however, are met by multilingual encoders that have been further specialized, in a supervised fashion, for sentence understanding tasks, rather than using their vanilla 'off-the-shelf' variants. Following these results, we introduce localized relevance matching for document-level CLIR, where we independently score a query against document sections. In the second part, we evaluate multilingual encoders fine-tuned in a supervised fashion (i.e., we learn to rank) on English relevance data in a series of zero-shot language and domain transfer CLIR experiments. Our results show that, despite the supervision, and due to the domain and language shift, supervised re-ranking rarely improves the performance of multilingual transformers as unsupervised base rankers. Finally, only with in-domain contrastive fine-tuning (i.e., same domain, only language transfer), we manage to improve the ranking quality. We uncover substantial empirical differences between cross-lingual retrieval results and results of (zero-shot) cross-lingual transfer for monolingual retrieval in target languages, which point to "monolingual overfitting" of retrieval models trained on monolingual (English) data, even if they are based on multilingual transformers.

Visual Summary Identification From Scientific Publications via Self-Supervised Learning.

The exponential growth of scientific literature yields the need to support users to both effectively and efficiently analyze and understand the some body of research work. This exploratory process can be facilitated by providing graphical abstracts-a visual summary of a scientific publication. Accordingly, previous work recently presented an initial study on automatic identification of a central figure in a scientific publication, to be used as the publication's visual summary. This study, however, have been limited only to a single (biomedical) domain. This is primarily because the current state-of-the-art relies on supervised machine learning, typically relying on the existence of large amounts of labeled data: the only existing annotated data set until now covered only the biomedical publications. In this work, we build a novel benchmark data set for visual summary identification from scientific publications, which consists of papers presented at conferences from several areas of computer science. We couple this contribution with a new self-supervised learning approach to learn a heuristic matching of in-text references to figures with figure captions. Our self-supervised pre-training, executed on a large unlabeled collection of publications, attenuates the need for large annotated data sets for visual summary identification and facilitates domain transfer for this task. We evaluate our self-supervised pretraining for visual summary identification on both the existing biomedical and our newly presented computer science data set. The experimental results suggest that the proposed method is able to outperform the previous state-of-the-art without any task-specific annotations.

Automatic Annotation of Narrative Radiology Reports.

Narrative texts in electronic health records can be efficiently utilized for building decision support systems in the clinic, only if they are correctly interpreted automatically in accordance with a specified standard. This paper tackles the problem of developing an automated method of labeling free-form radiology reports, as a precursor for building query-capable report databases in hospitals. The analyzed dataset consists of 1295 radiology reports concerning the condition of a knee, retrospectively gathered at the Clinical Hospital Centre Rijeka, Croatia. Reports were manually labeled with one or more labels from a set of 10 most commonly occurring clinical conditions. After primary preprocessing of the texts, two sets of text classification methods were compared: (1) traditional classification models-Naive Bayes (NB), Logistic Regression (LR), Support Vector Machine (SVM), and Random Forests (RF)-coupled with Bag-of-Words (BoW) features (i.e., symbolic text representation) and (2) Convolutional Neural Network (CNN) coupled with dense word vectors (i.e., word embeddings as a semantic text representation) as input features. We resorted to nested 10-fold cross-validation to evaluate the performance of competing methods using accuracy, precision, recall, and F 1 score. The CNN with semantic word representations as input yielded the overall best performance, having a micro-averaged F 1 score of 86 . 7 % . The CNN classifier yielded particularly encouraging results for the most represented conditions: degenerative disease ( 95 . 9 % ), arthrosis ( 93 . 3 % ), and injury ( 89 . 2 % ). As a data-hungry deep learning model, the CNN, however, performed notably worse than the competing models on underrepresented classes with fewer training instances such as multicausal disease or metabolic disease. LR, RF, and SVM performed comparably well, with the obtained micro-averaged F 1 scores of 84 . 6 % , 82 . 2 % , and 82 . 1 % , respectively.