In this paper, we present our progress in pre-training monolingual Transformers for Czech and contribute to the research community by releasing our models for public. The need for such models emerged from our effort to employ Transformers in our language-specific tasks, but we found the performance of the published multilingual models to be very limited. Since the multilingual models are usually pre-trained from 100+ languages, most of low-resourced languages (including Czech) are under-represented in these models. At the same time, there is a huge amount of monolingual training data available in web archives like Common Crawl. We have pre-trained and publicly released two monolingual Czech Transformers and compared them with relevant public models, trained (at least partially) for Czech. The paper presents the Transformers pre-training procedure as well as a comparison of pre-trained models on text classification task from various domains.
In recent years, large pre-trained transformers have led to substantial gains in performance over traditional retrieval models and feedback approaches. However, these results are primarily based on the MS Marco/TREC Deep Learning Track setup, with its very particular setup, and our understanding of why and how these models work better is fragmented at best. We analyze effective BERT-based cross-encoders versus traditional BM25 ranking for the passage retrieval task where the largest gains have been observed, and investigate two main questions. On the one hand, what is similar? To what extent does the neural ranker already encompass the capacity of traditional rankers? Is the gain in performance due to a better ranking of the same documents (prioritizing precision)? On the other hand, what is different? Can it retrieve effectively documents missed by traditional systems (prioritizing recall)? We discover substantial differences in the notion of relevance identifying strengths and weaknesses of BERT that may inspire research for future improvement. Our results contribute to our understanding of (black-box) neural rankers relative to (well-understood) traditional rankers, help understand the particular experimental setting of MS-Marco-based test collections.
Researchers and scientists increasingly rely on specialized information retrieval (IR) or recommendation systems (RS) to support them in their daily research tasks. Paper recommender systems are one such tool scientists use to stay on top of the ever-increasing number of academic publications in their field. Improving research paper recommender systems is an active research field. However, less research has focused on how the interfaces of research paper recommender systems can be tailored to suit the needs of different research domains. For example, in the field of biomedicine and chemistry, researchers are not only interested in textual relevance but may also want to discover or compare the contained chemical entity information found in a paper's full text. Existing recommender systems for academic literature do not support the discovery of this non-textual, but semantically valuable, chemical entity data. We present the first implementation of a specialized chemistry paper recommender system capable of visualizing the contained chemical structures, chemical formulae, and synonyms for chemical compounds within the document's full text. We review existing tools and related research in this field before describing the implementation of our ChemVis system. With the help of chemists, we are expanding the functionality of ChemVis, and will perform an evaluation of recommendation performance and usability in future work.
Automatic Process Discovery aims at developing algorithmic methodologies for the extraction and elicitation of process models as described in data. While Process Discovery from event-log data is a well established area, that has already moved from research to concrete adoption in a mature manner, Process Discovery from text is still a research area at an early stage of development, which rarely scales to real world documents. In this paper we analyze, in a comparative manner, reference state-of-the-art literature, especially for what concerns the techniques used, the process elements extracted and the evaluations performed. As a result of the analysis we discuss important limitations that hamper the exploitation of recent Natural Language Processing techniques in this field and we discuss fundamental limitations and challenges for the future concerning the datasets, the techniques, the experimental evaluations, and the pipelines currently adopted and to be developed in the future.
This work aims to tackle the challenging heterogeneous graph encoding problem in the text-to-SQL task. Previous methods are typically node-centric and merely utilize different weight matrices to parameterize edge types, which 1) ignore the rich semantics embedded in the topological structure of edges, and 2) fail to distinguish local and non-local relations for each node. To this end, we propose a Line Graph Enhanced Text-to-SQL (LGESQL) model to mine the underlying relational features without constructing meta-paths. By virtue of the line graph, messages propagate more efficiently through not only connections between nodes, but also the topology of directed edges. Furthermore, both local and non-local relations are integrated distinctively during the graph iteration. We also design an auxiliary task called graph pruning to improve the discriminative capability of the encoder. Our framework achieves state-of-the-art results (62.8% with Glove, 72.0% with Electra) on the cross-domain text-to-SQL benchmark Spider at the time of writing.
Bayesian optimization (BO) has become a popular strategy for global optimization of many expensive real-world functions. Contrary to a common belief that BO is suited to optimizing black-box functions, it actually requires domain knowledge on characteristics of those functions to deploy BO successfully. Such domain knowledge often manifests in Gaussian process priors that specify initial beliefs on functions. However, even with expert knowledge, it is not an easy task to select a prior. This is especially true for hyperparameter tuning problems on complex machine learning models, where landscapes of tuning objectives are often difficult to comprehend. We seek an alternative practice for setting these functional priors. In particular, we consider the scenario where we have data from similar functions that allow us to pre-train a tighter distribution a priori. To verify our approach in realistic model training setups, we collected a large multi-task hyperparameter tuning dataset by training tens of thousands of configurations of near-state-of-the-art models on popular image and text datasets, as well as a protein sequence dataset. Our results show that on average, our method is able to locate good hyperparameters at least 3 times more efficiently than the best competing methods.
This work proposes a transformer architecture for user-level classification of gambling addiction and depression that is trainable end-to-end. As opposed to other methods that operate at the post level, we process a set of social media posts from a particular individual, to make use of the interactions between posts and eliminate label noise at the post level. We exploit the fact that, by not injecting positional encodings, multi-head attention is permutation invariant and we process randomly sampled sets of texts from a user after being encoded with a modern pretrained sentence encoder (RoBERTa / MiniLM). Moreover, our architecture is interpretable with modern feature attribution methods and allows for automatic dataset creation by identifying discriminating posts in a user's text-set. We perform ablation studies on hyper-parameters and evaluate our method for the eRisk 2022 Lab on early detection of signs of pathological gambling and early risk detection of depression. The method proposed by our team BLUE obtained the best ERDE5 score of 0.015, and the second-best ERDE50 score of 0.009 for pathological gambling detection. For the early detection of depression, we obtained the second-best ERDE50 of 0.027.
Arbitrary-shaped text detection is an important and challenging task in computer vision. Most existing methods require heavy data labeling efforts to produce polygon-level text region labels for supervised training. In order to reduce the cost in data labeling, we study weakly-supervised arbitrary-shaped text detection for combining various weak supervision forms (e.g., image-level tags, coarse, loose and tight bounding boxes), which are far easier for annotation. We propose an Expectation-Maximization (EM) based weakly-supervised learning framework to train an accurate arbitrary-shaped text detector using only a small amount of polygon-level annotated data combined with a large amount of weakly annotated data. Meanwhile, we propose a contour-based arbitrary-shaped text detector, which is suitable for incorporating weakly-supervised learning. Extensive experiments on three arbitrary-shaped text benchmarks (CTW1500, Total-Text and ICDAR-ArT) show that (1) using only 10% strongly annotated data and 90% weakly annotated data, our method yields comparable performance to state-of-the-art methods, (2) with 100% strongly annotated data, our method outperforms existing methods on all three benchmarks. We will make the weakly annotated datasets publicly available in the future.
We introduce a new framework for text detection named SA-Text meaning "Simple but Accurate," which utilizes heatmaps to detect text regions in natural scene images effectively. SA-Text detects text that occurs in various fonts, shapes, and orientations in natural scene images with complicated backgrounds. Experiments on three challenging and public scene-text-detection datasets, Total-Text, SCUT-CTW1500, and MSRA-TD500 show the effectiveness and generalization ability of SA-Text in detecting not only multi-lingual oriented straight but also curved text in scripts of multiple languages. To further show the practicality of SA-Text, we combine it with a powerful state-of-the-art text recognition model and thus propose a pipeline-based text spotting system called SAA ("text spotting" is used as the technical term for "detection and recognition of text"). Our experimental results of SAA on the Total-Text dataset show that SAA outperforms four state-of-the-art text spotting frameworks by at least 9 percent points in the F-measure, which means that SA-Text can be used as a complete text detection and recognition system in real applications.
Language Identification in textual documents is the process of automatically detecting the language contained in a document based on its content. The present Language Identification techniques presume that a document contains text in one of the fixed set of languages, however, this presumption is incorrect when dealing with multilingual document which includes content in more than one possible language. Due to the unavailability of large standard corpora for Hindi-English mixed lingual language processing tasks we propose the language lexicons, a novel kind of lexical database that supports several multilingual language processing tasks. These lexicons are built by learning classifiers over transliterated Hindi and English vocabulary. The designed lexicons possess richer quantitative characteristic than its primary source of collection which is revealed using the visualization techniques.