BiodivBERT: a Pre-Trained Language Model for the Biodiversity Domain Biodiversity, Language Model, BERT, Pre-training, Fine-tuning , Felicitas Löffler Friedrich Schiller University Jena , Jena , Germany Michael Stifel Center Jena , Germany 2023

Information Extraction in the Life Sciences is getting increasing attention due to the constantly growing amount of data and text. The advancements of deep learning models further accelerate this development. However, applying these models to domain-specific data is crucial as applied domains often require different entity type extractions than general ones. This paper introduces BiodivBERT, the first pretrained language model for the biodiversity domain. We constructed two pre-training corpora (abstracts and abstracts + full text) based on a keyword search strategy from two leading publishers in the Life Sciences. In addition, we fine-tuned BiodivBERT on two downstream tasks, i.e., Named Entity Recognition (NER) and Relation Extraction (RE), using various state-of-the-art benchmarks. The results show that BiodivBERT outperforms the state-of-the-art approaches. Moreover, we discuss a potential application of BiodivBERT for ontology auto-population. We publicly release data and code for both pre-training and fine-tuning.

 -

1,2,3

 1. Introduction

Motivated by the predicted impending loss of biodiversity and the consequences of this loss for humanity [ 1 ], research in the biodiversity domain has recently witnessed accelerated growth. For instance, the Biodiversity Heritage Library (BHL)1 currently holds over 55 million digitized pages of legacy biology text from the 15ℎ − 21 centuries, representing a massive amount of textual content [ 2 ]. Moreover, Google Scholar returns more than 85,000 hits for a search using the term “biodiversity” from 2021 till the time of writing, November 2022. Thus, text mining tools are an open demand in the field to leverage this untapped wealth. Recent progress in the development of deep learning models for Natural Language Processing (NLP) promises to answer this demand. However, directly applying such NLP techniques to biodiversity texts is not promising. Modern word representation models such as Word2Vec [ 3 ], GloVe [ 4 ], ELMo [ 5 ], Sciences ∗Corresponding author. CEUR Workshop Proceedings htp:/ceur-ws.org ISN1613-073

CEUR Workshop Proceedings (CEUR-WS.org) and BERT[ 6 ] are trained and tested on general domain texts (e.g., Wikipedia). However, domainspecific, texts contain many instances of domain-specific entity types. E.g., Helianthus (genus of sunflower species), calcareous grassland, or growth rate. Thus, it is difficult to estimate the performance of general-purpose models on domain-specific datasets. Approaches to improve the performance of cutting-edge approaches like BERT on domain-specific benchmarks have first been developed for the (bio-)medical domain with BioBERT [ 7 ] and clinicalBERT [ 8 ]. BioBERT is initialized with BERT weights and pre-trained on biomedical corpora that are based on PubMed2 and PubMed Central (PMC)3. It showed a significant improvement on three downstream tasks, namely Named Entity Recognition (NER), Relation Extraction (RE), and Question Answering (QA). To the best of our knowledge, there is no language model for the biodiversity domain that supports the extraction of named entities and relations from textual data.

The contributions of our paper are as follows: i) BiodivBERT is the first biodiversity-specific BERT-based model pre-trained on biodiversity corpora. ii) We show that pre-training BERT on biodiversity corpora improves its performance on two downstream tasks, NER and RE. iii) We discuss potential applications for BiodivBERT, e.g., ontology population. v) We make our pretraining corpora [ 9 ], pre-processed datasets [ 10 ], the pre-trained weights [ 11 ] of BiodivBERT, and the source code for pre-training and fine-tuning publicly available4.

The rest of this paper is organized as follows: We give an overview of our approach in Section 2. We explain the pre-training and fine-tuning tasks in Section 3 and Section 4 respectively. We show our results in Section 5. We demonstrate a potential application of the current work in Section 6. We conclude in Section 7.

 2. Approach

In this paper, we introduce BiodivBERT, a pre-trained language representation model for the biodiversity domain. The overall process of pre-training and fine-tuning BiodivBERT is shown in Figure 1. First, we initialize BiodivBERT with weights from BERT [ 6 ], which was pre-trained on general domain corpora (English Wikipedia and BooksCorpus). Then, BiodivBERT is pre-trained on our collected corpora from the biodiversity domain. The first corpus is based on abstracts (+Abs), while the other contains both abstracts and full text (+Abs+Full). To demonstrate the effectiveness of BiodivBERT in biodiversity text mining, we have fine-tuned and evaluated it on two downstream tasks, NER and RE, using various task-specific datasets.

3. Pre-training BiodivBERT

In this section, we explain our pre-training data sources, selection strategy, and data statistics. In addition, we discuss the pre-training task for BiodivBERT.

3.1. Pre-training Data

In this section, we discuss the construction of our two pre-training corpora that are based on Abstracts (+Abs), and Full text (+Abs+Full). Thus, we explain our used keywords search strategy, workflow, and the resultant corpora statistics.

2https://pubmed.ncbi.nlm.nih.gov/ 3https://www.ncbi.nlm.nih.gov/pmc/ 4https://github.com/fusion-jena/BiodivBERT

Pre-training of BiodivBERT Pre-training Corpora BiodivBERT Pre-training

Fine-tuning of BiodivBERT

Task-specific Datasets BiodivBERT Fine-tuning Weights initialization T1 T2 . . TN Tm Tm Tm Tm Tm Tm E1 E2 … EN

BERT

T1 T2 ... TN Tm Tm Tm Tm

Tm

Tm E1 E2 … EN Pre-trained BiodivBERT with biodiversity domain corpora

Named Entity Recognition

(NER) Species-800, LINNAUS, COPIOUS, QEMP, BiodivNER Relation Extraction

(RE) BiodivRE, GAD, EU_ADR,

BioRelEx richness B-QUA of O fungi B-ORG in O T1 ahabTit2at BO.-.E.NV TN Tm Tm Tm

Corpora Construction Pipeline: To show the effect of the pre-training data on the model performance through the downstream tasks, we created two pre-training corpora. One is based on abstracts (+Abs) only, while the other contains abstracts and full text of publications (+Abs+Full). Under the access rights provided by the selected publishers, we used abstracts and full texts of open-access papers and abstracts only for other publications. To construct the +Abs corpus, we used the pre-selected keywords and year range as input for both Elsevier and Springer’s provided full-text search APIs. For each of them, we retrieved the corresponding DOIs for each keyword in the given year’s range. We then applied a deduplication method to the result set. We applied the same procedure for the second corpus, which is based on the full texts (+Abs+Full). Elsevier provided a straightforward API to obtain the parsed full text for a given article. However, for Springer’s full text, we downloaded the corresponding PDF file for each DOI, converted it to an XML format, and then extracted the text. We converted the downloaded PDFs to XML files using the GROBID service, and client [ 12 ]. We cleaned and merged the final text from both data sources. We applied both shallow and deep cleaning steps for the collected data. For instance, we filtered the sentences to include unique ones. In addition, we used regular expressions to remove URLs and DOIs.

5https://dev.elsevier.com/ 6https://dev.springernature.com/ 7https://github.com/fusion-jena/BiodivBERT/blob/main/keywords_search 8The starting year of this project.

rsse h li b u P Elsevier Full Text Downloaded Unique Available

Corpora Statistics: Table 1 gives an overview of our final corpora statistics. +Abs and +Abs+Full are around 1GB and 4GB in size, respectively. Our corpora include around 1M abstracts with 92K full publications. We faced data loss as shown in Figure 2 due to several reasons: 1) not found (404) errors for some articles because of either technical issues on the provider side or invalid DOIs. 2) Elsevier allows crawling only 6000 articles per keyword through its API. 3) GROBID failed to convert some PDF files from Springer. So, the shown numbers indicate the best we can use from both publishers under such circumstances. 5M 25M 134M 548M

Size 876 MB 3.81 GB

 3.2. Pre-training Task

We pre-trained BiodivBERT on our domain-specific corpora (+Abs) and (+Abs+Full). We initialized BiodivBERT with the B E R T _ b a s e _ c a s e d weights for computation efficiency and to leverage the general domain learned weights from the Wiki and books corpora by the original model. For tokenization, we used the same BERT WordPiece [ 13 ] tokenizer, which overcomes the out-of-vocab (OOV) issue. Similar to BioBERT [ 7 ], we used the cased vocabulary in our setting as it has higher performance on the downstream tasks, and we used the original vocabulary of B E R T _ b a s e _ c a s e d for the same reason and to be compatible with both BERT and BioBERT. We compare BiodivBERT to the SOTA BERT-based models. In addition, we tested different combinations of pre-training corpora. Thus, we compare such settings in Table 2. We pre-trained BiodivBERT+Abs, and BiodivBERT+Abs+Full using transformers by Hugging Face [ 14 ] library on a single V100 GPU (16 GB) for 3 and 5 days respectively. We used 512 for the maximum sequence length and 15% of the masked language model probability for hyperparameters. In addition, we used Adam’s optimizer with 1e-3 learning weight and default betas. Moreover, we set the batch size to 16 and enabled the gradient accumulation with four steps for faster training.

4. Fine-tuning BiodivBERT

With minimal architectural modification, BiodivBERT can be applied to various downstream text mining tasks. In this work, we fine-tuned it on both NER and RE using P100 (16 GB) by Colab Pro9 using various state-of-the-art datasets.

Named Entity Recognition is the task of identifying the domain-specific proper nouns inside a given text. We leveraged the original BERT structure for NER such that it uses a single output layer based on the representations from its last layer to compute only token-level probabilities. We used entity-level precision, recall, and F1 score as the evaluation metrics of NER. We selected various SOTA datasets to test the performance of BiodivBERT on NER. COPIOUS [ 2 ] is based on BHL documents and has six entity types, a.k.a. tags, including, e.g., habitat and taxon names. QEMP [ 15 ] is created from datasets metadata files and contains four entity types (tags), e.g., quality and material. BiodivNER [ 16 ] (dataset [ 17 ]) is constructed from metadata files and abstracts from PubMed and has five tags, e.g., organism and phenomena. Species800 [ 18 ] and Linnaeus [ 19 ] are designed for species names that are normalised to NCBI Taxonomy database10. We pre-processed all of them to follow the BIO11 format for token classification.

Table 3 Overview of the selected NER datasets.

Dataset BiodivNER

QEMP COPIOUS Species800 Linnaeus

Tags #Docs #Statements

#Annotations Relation Extraction is the task of classifying relations among named entities in a corpus. We utilized the sentence classifier of the original version of BERT, which uses a [CLS] token for the classification of relations. Relations in such a way use a single output layer based on a [CLS] token representation from BERT. To the best of our knowledge, BiodivRE [ 17 ] is the only available RE corpus for the biodiversity domain, so we included it in our fine-tuning setting. In addition, we included the BioRelEx [ 20 ], EU-ADR [ 21 ], and GAD [ 22 ] corpora from the biomedical domain. BiodivRE contains relations among the five entity types of BiodivNER, like occur_in and influence in a multi-class and a binary format. BioRelEx classifies the bindings 9https://colab.research.google.com/?utm_source=scs-index 10https://www.ncbi.nlm.nih.gov/taxonomy 11https://natural-language-understanding.fandom.com/wiki/Named_entity_recognition#BIO between genes and diseases into three categories: exists (1), not exists (-1), and unsure (0). EU-ADR and GAD include relations between genes and diseases. In this work, we used the binary format that BiodivRE provides. For BioRelEx, we constructed a binary relation corpus by excluding the unsure relations. Moreover, similar to BioBERT, we anonymized the target named entities in a sentence using their tags, e.g., @COMPLEXPROTEIN$ and @GENE$. For EU-ADR and GAD, we used the provided pre-processed version by BioBERT’s team since the original data are unavailable. Table 4 shows the selected RE datasets’ statistics.

5. Results & Discussion

To gain a first impression of the performance of our approach, we ran a mask-filling task on a typical biodiversity topic on BERT, BioBERT, and BiodivBERT using the following test case: “Diversification and [MASK] in brood pollination mutualisms.”. Table 5 shows that BiodivBERT has produced the most realistic results compared to the other two models. For example, BiodivBERT can generate both “diversity” and “evolution”. Such results demonstrate the effectiveness of the pre-training data.

Table 6 and Table 7 shows the scores of fine-tuning the BiodivBERT, BERT, and BioBERT models on the two downstream tasks NER, and RE respectively. In addition, we developed a single layer of the Bidirectional Long Short Term Memory (BiLSTM) with 10% dropout as a baseline approach. We micro-averaged the results per dataset to generate the scores of all systems. We fine-tuned these models on a single P100 GPU provided by Colab Pro. First, we found that BioBERTv1.1 obtained higher scores than BERTBASE on the downstream tasks, Second, BiodivBERT+Abs+Full and BiodivBERT+Abs gained the best results among all the others by achieving either first or second place on the given datasets. In detail, for NER, BiodivBERT+Abs+Full exceeds BioBERTv1.1 for all datasets except QEMP ,where BiodivBERT+Abs exceeds BioBERTv1.1, with 1% F1 score. In addition, we notice that the scores of the Speciesrelated datasets (Species-800 and LINNAEUS) gained higher scores than others; we argue that is due to such datasets being easier than those with fuzzy categories to identify. E.g., QEMP and BiodivNER have a class, “QUALITY” that groups data measures that cover vast and various attributes of the biodiversity domain and would be harder to detect. For RE, we have mixed results; for example, BiodivBERT+Abs+Full outperforms BioBERTv1.1 with a 2.5% F1 score for EU-ADR. However, BioBERTv1.1 overcomes BiodivBERT+Abs+Full with 3% F1 score for BiodivRE. We plan to apply different fine-tuning settings on those datasets to enhance the scores.

 6. Application

In this section, we discuss a possible application for BiodivBERT: A while ago, we started developing a biodiversity domain ontology, Biodivonto [ 23 ]. Such an ontology is needed for many applications both by us and for other researchers in the field. BiodivOnto consists of eight core concepts including, e.g., “Organism”, “Environment”, “Location”, and three core relations: have, in, and influence. So far, the population of the ontology with instances is incomplete. Adding instances manually is prohibitively expensive. Since the ontology is the basis of BiodivNERE [ 16 ] for both tasks, BiodivBERT could be used to auto-populate the ontology. For instance, Seabass would be classified as an “Organism”. In a final step, the identified instances should be linked to existing knowledge graphs. e.g., Seabass would be mapped to http://www.wikidata.org/entity/Q307102 from Wikidata. This could be done, e.g. using our approach described in [ 24 ]. With this, BiodivBERT has the potential to bring us closer to our ultimate goal, the creation of a comprehensive biodiversity knowledge graph out of textual data.

Dataset Spieces-800 LINNANUS COPIOUS

QEMP BiodivNER

P R F P R F P R F P R F P R F

 7. Conclusions & Future Work

In this paper, we introduced BiodivBERT as a pre-trained language model on two domainspecific corpora based on modern research data from the biodiversity domain. In addition, we fine-tuned it on two downstream tasks for text mining: named entity recognition (NER) and relation extraction (RE). BiodivBERT outperforms the state-of-the-art approaches on taskspecific datasets.

Future Work: We plan to pre-train and fine-tune a lightweight model, e.g., distilBERT [ 25 ]. We also plan to investigate the reasons behind the low scores on the RE task, especially with the BiodivRE and BioRelEx datasets. For example, we could try different settings for fine-tuning. In addition, we fine-tune BiodivBERT on more task-specific datasets whenever they are available. Finally, we deploy BiodivBERT for actual token and sequence prediction for the biodiversity literature to auto-populate the BiodivOnto.

Acknowledgments The authors thank the Carl Zeiss Foundation for the financial support of the project “A Virtual Werkstatt for Digitization in the Sciences (K3, P5)” within the scope of the program line “Breakthroughs: Exploring Intelligent Systems for Digitization” - explore the basics, use applications. Our sincere thanks to Björn Barz, Luise Modersohn, Jitendra Gaikwad, Anahita Kazem, Alsayed Algergawy, Leila Feddoul, Anirudh Ashok, and Andreas Ostrowski, University Jena for their recommendations, and help. Data Journal 7 (2019) e29626. URL: https://doi.org/10.3897/BDJ.7.e29626. doi:1 0 . 3 8 9 7 / B D J . 7 . e 2 9 6 2 6 .

 [1] E. S. Brondizio , J. Settele , S. Díaz , H. T. Ngo, Global assessment report on biodiversity and ecosystem services of the intergovernmental science-policy platform on biodiversity and ecosystem services ( 2019 ). URL: https://ipbes.net/global-assessment. [2] N. T.H.  Nguyen, R. S.   Gabud , S. Ananiadou , Copious: A gold standard corpus of named entities towards extracting species occurrence from biodiversity literature , Biodiversity [3] T. Mikolov , I. Sutskever, K. Chen , G. S. Corrado , J. Dean , Distributed representations of words and phrases and their compositionality , Advances in neural information processing systems 26 ( 2013 ). [4] J. Pennington , R. Socher , C. D. Manning , Glove: Global vectors for word representation , in: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) , 2014 , pp. 1532 - 1543 . [5] M. E. Peters , M. Neumann , M. Iyyer , M. Gardner , C. Clark , K. Lee , L. Zettlemoyer , Deep contextualized word representations , in: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , Volume 1 ( Long Papers) , Association for Computational Linguistics , New Orleans, Louisiana, 2018 , pp. 2227 - 2237 . URL: https://aclanthology.org/N18-1202. doi:1 0 . 1 8 6 5 3 / v 1 / N 1 8 - 1 2 0 2 . [6] J. Devlin , M. Chang , K. Lee , K. Toutanova , BERT: pre-training of deep bidirectional transformers for language understanding , in: J. Burstein , C. Doran , T. Solorio (Eds.), Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis , MN, USA, June 2-7, 2019 , Volume 1 (Long and Short Papers), Association for Computational Linguistics , 2019 , pp. 4171 - 4186 . doi: 1 0 . 1 8 6 5 3 / v 1 / n 1 9 - 1 4 2 3 . [7] J. Lee , W. Yoon , S. Kim , D. Kim , S. Kim , C. H. So , J. Kang, BioBERT: a pre-trained biomedical language representation model for biomedical text mining , Bioinformatics 36 ( 2020 ) 1234 - 1240 . [8] E. Alsentzer , J. Murphy , W. Boag , W.-H. Weng , D. Jin , T. Naumann , M. McDermott, Publicly available clinical BERT embeddings , in: Proceedings of the 2nd Clinical Natural Language Processing Workshop , Association for Computational Linguistics, Minneapolis, Minnesota, USA, 2019 , pp. 72 - 78 . doi:1 0 . 1 8 6 5 3 / v 1 / W 1 9 - 1 9 0 9 . [9] N. Abdelmageed , F. Löffler , B. König-Ries , Biodivbert: Pre-training corpora dois , 2022. doi:1 0 . 5 2 8 1 / z e n o d o . 6 5 5 5 6 9 0 . [10] N. Abdelmageed , F. Löffler , B. König-Ries , BiodivBERT: Pre-processed Datasets for NER and RE Downstream Tasks, 2022 . doi:1 0 . 5 2 8 1 / z e n o d o . 6 5 5 4 2 0 8 . [11] N. Abdelmageed , F. Löffler , B. König-Ries , BiodivBERT: Pre-trained weights, configuration , and training arguments, 2022. doi:1 0 . 5 2 8 1 / z e n o d o . 6 5 5 4 1 4 1 . [12] GROBID, https://github.com/kermitt2/grobid, 2008- 2021 . [13] Y. Wu , M. Schuster , Z. Chen , Q. V. Le , M. Norouzi , W. Macherey , M. Krikun , Y. Cao , Q. Gao , K. Macherey , et al., Google's neural machine translation system: Bridging the gap between human and machine translation , arXiv preprint arXiv:1609.08144 ( 2016 ). [14] T. Wolf , L. Debut , V. Sanh , J. Chaumond , C. Delangue , A. Moi , P. Cistac , T. Rault , R. Louf , M. Funtowicz , J. Davison , S. Shleifer , P. von Platen, C. Ma, Y. Jernite , J. Plu , C. Xu , T. L. Scao , S. Gugger , M. Drame , Q. Lhoest , A. M. Rush , Transformers: State-of-the-art natural language processing , in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , Association for Computational Linguistics , Online, 2020 , pp. 38 - 45 . URL: https://www.aclweb.org/anthology/2020.emnlp-demos. 6 . [15] F. Löffler , N. Abdelmageed , S. Babalou , P. Kaur , B. König-Ries , Tag me if you can! semantic annotation of biodiversity metadata with the QEMP corpus and the BiodivTagger , in: Proceedings of the 12th Language Resources and Evaluation Conference , European Language Resources Association, Marseille, France, 2020 , pp. 4557 - 4564 . URL: https://aclanthology.org/ 2020 .lrec- 1 . 560 . [16] N. Abdelmageed , F. Löffler , L. Feddoul , A. Algergawy , S. Samuel , J. Gaikwad , A. Kazem , B. König-Ries , Biodivnere: Gold standard corpora for named entity recognition and relation extraction in the biodiversity domain , Biodiversity Data Journal 10 ( 2022 ) e89481 . URL: https://doi.org/10.3897/BDJ.10.e89481. doi:1 0 . 3 8 9 7 / B D J . 1 0 . e 8 9 4 8 1 . [17] N. Abdelmageed , F. Löffler , L. Feddoul , A. Algergawy , S. Samuel , J. Gaikwad , A. Kazem , B. König-Ries , BiodivNERE: Gold Standard Corpora for Named Entity Recognition and Relation Extraction in Biodiversity Domain , 2022 . URL: https://doi.org/10.5281/zenodo. 6458503. doi:1 0 . 5 2 8 1 / z e n o d o . 6 4 5 8 5 0 3 . [18] E. Pafilis , S. P. Frankild , L. Fanini , S. Faulwetter , C. Pavloudi , A. Vasileiadou , C. Arvanitidis , L. J. Jensen , The species and organisms resources for fast and accurate identification of taxonomic names in text , PloS one 8 ( 2013 ) e65390 . [19] M. Gerner , G. Nenadic, C. M. Bergman , Linnaeus: a species name identification system for biomedical literature , BMC bioinformatics 11 ( 2010 ) 1 - 17 . [20] H. Khachatrian , L. Nersisyan , K. Hambardzumyan , T. Galstyan , A. Hakobyan , A. Arakelyan , A. Rzhetsky , A . Galstyan, BioRelEx 1 . 0: Biological relation extraction benchmark , in: Proceedings of the 18th BioNLP Workshop and Shared Task, Association for Computational Linguistics, Florence, Italy, 2019 , pp. 176 - 190 . URL: https://aclanthology.org/W19-5019. doi:1 0 . 1 8 6 5 3 / v 1 / W 1 9 - 5 0 1 9 . [21] E. M. Van Mulligen , A. Fourrier-Reglat , D. Gurwitz , M. Molokhia , A. Nieto , G. Trifiro, J. A. Kors , L. I. Furlong , The eu-adr corpus: annotated drugs, diseases, targets, and their relationships , Journal of biomedical informatics 45 ( 2012 ) 879 - 884 . URL: https: //www.sciencedirect.com/science/article/pii/S1532046412000573. [22] À. Bravo , J. Piñero , N. Queralt-Rosinach , M. Rautschka , L. I. Furlong , Extraction of relations between genes and diseases from text and large-scale data analysis: implications for translational research , BMC bioinformatics 16 ( 2015 ) 1 - 17 . URL: https://bmcbioinformatics. biomedcentral.com/track/pdf/10.1186/s12859-015-0472-9.pdf. [23] N. Abdelmageed , A. Algergawy , S. Samuel , B. König-Ries , Biodivonto: Towards a core ontology for biodiversity , in: The Semantic Web: ESWC 2021 Satellite Events - Virtual Event , June 6-10, 2021 , Revised Selected Papers, volume 12739 , Springer, 2021 , pp. 3 - 8 . doi:1 0 . 1 0 0 7 / 9 7 8 - 3 - 0 3 0 - 8 0 4 1 8 - 3 \ _ 1 . [24] N. Abdelmageed , S. Schindler , Jentab: Matching tabular data to knowledge graphs., in: Proceedings of the Semantic Web Challenge on Tabular Data to Knowledge Graph Matching (SemTab 2020) co-located with the 19th International Semantic Web Conference ISWC , 2020 , pp. 40 - 49 . URL: https://ceur-ws. org/ Vol- 2775 /paper4.pdf. [25] V. Sanh , L. Debut , J. Chaumond , T. Wolf, Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter , arXiv preprint arXiv: 1910 . 01108 ( 2019 ).