Biomedical Data-to-Text Generation via Fine-Tuning Transformers

09/03/2021 ∙ by Ruslan Yermakov, et al. ∙ Bayer 0

Data-to-text (D2T) generation in the biomedical domain is a promising - yet mostly unexplored - field of research. Here, we apply neural models for D2T generation to a real-world dataset consisting of package leaflets of European medicines. We show that fine-tuned transformers are able to generate realistic, multisentence text from data in the biomedical domain, yet have important limitations. We also release a new dataset (BioLeaflets) for benchmarking D2T generation models in the biomedical domain.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Data-to-text (D2T) systems are attracting considerable interest due to their ability to automate the time-consuming writing of data-driven reports. There is a hitherto largely untapped potential for text generation in the biomedical domain. Potential applications of natural language generation of patient-friendly biomedical text include preparation of the first draft of package leaflets, patient education materials, or direct-to-consumer promotional materials in countries where this is permitted. Here we focus on a D2T task aiming to generate fluent and fact-based descriptions from biomedical data.

2 Related Work

Recently, neural D2T models have significantly improved the quality of short text generation (usually one sentence long) from input data compared to multi-stage pipelined or template-based approaches. Examples include biographies from Wikipedia fact tables  (Lebret et al., 2016), restaurant descriptions from meaning representations  (Novikova et al., 2017), and basketball game summaries from statistical tables  (Wiseman et al., 2017). Still, neural D2T approaches have major challenges, as outlined by  Wiseman et al. (2017) and  Parikh et al. (2020) which hinder their application to many real-world applications. These include hallucination effects (generated phrases not supported or contradictory to the source data), missing facts (generated text does not include input information), intersentence incoherence, and repetitiveness in the generated text. Following the success of leveraging pre-trained large-scale language models for a large variety of tasks, Kale and Rastogi (2020) fine-tuned T5 models  (Raffel et al., 2020) for D2T generation. This strategy achieved state-of-the-art performance on task-oriented dialogue (MultiWoz)  (Budzianowski et al., 2018), tables-to-text (ToTTo)  (Parikh et al., 2020) and graph-to-text (WebNLG)  (Gardent et al., 2017).

To the best of our knowledge, recent neural approaches and transfer learning strategies have not been applied to multi-sentence generation from input data, nor have they been applied in the biomedical domain. Our contribution is two-fold: we introduce a real-world biomedical dataset

BioLeaflets, and demonstrate that transformers can generate high-quality multi-sentence text from data in the biomedical domain. The BioLeaflets dataset, fine-tuned models, code, and generated samples are available at https://github.com/bayer-science-for-a-better-life/data2text-bioleaflets.

(a)
Original section content
novonorm is an oral antidiabetic medicine containing repaglinide which helps your pancreas produce more insulin and thereby lower your blood sugar (glucose). type 2 diabetes is a disease in which your pancreas does not make enough insulin to control the sugar in your blood or where your body does not respond normally to the insulin it produces. novonorm is used to control type 2 diabetes in adults as an add-on to diet and exercise: treatment is usually started if diet, exercise and weight reduction alone have not been able to control (or lower) your blood sugar. novonorm can also be given with metformin, another medicine for diabetes. novonorm has been shown to lower the blood sugar, which helps to prevent complications from your diabetes.
(b)
Input:
entities as a flat string
<PRODUCT_NAME> novonorm </PRODUCT_NAME> <TREATMENT>
an_oral_antidiabetic_medicine </TREATMENT> <GENERIC_NAME> repaglinide
</GENERIC_NAME> <SYSTEM_ORGAN_SITE> pancreas </SYSTEM_ORGAN_SITE>
<GENERIC_NAME> insulin </GENERIC_NAME> <TEST> your_blood_sugar </TEST>
<TEST> the_blood_sugar </TEST> <DX_NAME> complications </DX_NAME>
<PROBLEM> your_diabetes </PROBLEM>
(c)
T5 generated section
novonorm is an oral antidiabetic medicine containing repaglinide which helps your pancreas produce more insulin and thereby lower your blood sugar (glucose). type 2 diabetes is a disease in which your panccreases does not make enough insulin to control the sugar in your blood or where your body does not respond normally to the insulin it produces. novonourm is used to control type 2 diabetic patients who still have their diabetes and who are still unable to respond normally despite treatment, exercise and weight reduction, which have not helped to control your blood glucose. there is also a test that novonorgm can be given in combination with metformin, another medicine for diabetes. it is important that you continue to follow the advice on diet, exercising and diet that your nurse or doctor has given you.
(d)
BART generated section
novondorm is an oral antidiabetic medicine containing repaglinide which helps your pancreas produce more insulin and thereby lower your blood sugar (glucose). type 2 diabetes is a disease in which your pancrea does not make enough insulin to control the sugar in your blood or where your body does not respond normally to the insulin it produces. novonorm is used to treat type 2 diabetic in adults. treatment is usually started with diet and exercise and weight reduction. your blood glucose may be increased when you start to take novonor on its own or in combination with metformin. if you have diabetes, no
(e)
Content planner generated section
novonorm contains the active substance repaglinide which helps to lower your blood sugar (glucose). type 2 diabetes is a disease where your body does not make enough insulin to control the sugar in your blood or where your body does not respond normally to the insulin it produces. repaglinide krka is used to control type 2 diabetes in adults as type 2 diabetes. type 2 diabetes is also called non - insulin - dependent diabetes mellitus. type 2 diabetes is also a condition in which your body does not make enough insulin or the insulin that your body produces does not work as well as it should. your body can also make too much sugar. when this happens, sugar (glucose) builds up in the blood. this can lead to serious medical problems like heart disease, kidney disease, 2 and 2.
Table 1: Example of text generations. Entities are highlighted in bold, typos are underlined, and hallucinations are shown in red.

3 The BioLeaflets Dataset

We introduce a new biomedical dataset for D2T generation - BioLeaflets, a corpus of 1336 package leaflets of medicines authorised in Europe, which we obtain by scraping the European Medicines Agency (EMA) website. This dataset comprises the large majority ( 90%) of medicinal products authorised through the centralised procedure in Europe as of January 2021.

Package leaflets are published for medicinal products approved in the European Union (EU). They are included in the packaging of medicinal products and contain information to help patients use the product safely and appropriately, under the guidance of their healthcare professional. Package leaflets are required to be written in a way that is clear and understandable EU (2001). Each document contains six sections (see Table 2).The main challenges of this dataset for D2T generation are multi-sentence and multi-section target text, small sample size, specialized medical vocabulary and syntax.

3.1 Dataset Construction

The content of each section is not standardized, yet it is still well-structured. Thus, we identify sections via heuristics such as regular expressions and word overlap. The content of each section is lower-cased and tokenized by treating all special characters as separate tokens. Duplicates are also removed. We randomly split the dataset into training (80%), development (10%), and test (10%) set. Table

2 summarizes dataset statistics.

Section type No. samples Average length (characters) Average length (tokens) Average no. entities per section No. unique entities
1. What the product is and what it is used for
2. What you need to know before you take the product
3. How to take the product
4. Possible side effects
5. How to store the product
6. Content of the pack and other information
Table 2: BioLeaflets dataset statistics grouped by section type.

3.2 Dataset Annotations

We do not have annotations available for the package leaflet text. To create the required input for D2T generation, we augment each document by leveraging named entity recognition (NER).

Parikh et al. (2020) indicated it is important that target summaries contain information that can be inferred from the input data to avoid dataset-induced hallucinations. To this end, we combine two NER frameworks: Amazon Comprehend Medical (ACM)  (Bhatia et al., 2019) and Stanford Stanza  (Qi et al., 2020; Zhang et al., 2021). ACM and Stanza achieved entity micro-averaged test F1 of 85.5% and 88.13% respectively on the 2010 i2b2/VA clinical dataset  (Uzuner et al., 2011). We further leverage ACM to detect medical conditions from ICD-10 (WHO, 2004) and medications from RxNorm. Additionally, we treat all digits as entities, and add the medicine name as first entity. In case of overlapping entities from different sources, we favor longer entities over shorter ones. As a result of the NER process, we obtain 26 unique entity types. Examples are: problem: (’active chronic hepatitis’, ’migraine pain’), system-organ-site: (’blood vessel’, ’kidneys’, ’surrounding tissue’), treatment: (’routine dental care’, ’a vaccination’, ’a chemotherapy medicinal product’), or procedure: (’injections’, ’spinal or epidural anaesthesia’, ’surgical intervention’, ’bone marrow or stem cell transplant’).

BioLeaflets proposes a conditional generation task: given an ordered set of entities as source, the goal is to produce a multi-sentence section. Since only the entities are provided as input, the structured data is underspecified. A human without specialized knowledge would likely be unable to produce satisfactory text. However, we expect that a labeling expert with profound knowledge of package leaflets would be able to generate (with some difficulty) satisfactory text in the large majority of cases. Successful generation thus requires the model to learn specific syntax, terminology, and writing style from the corpus (e.g., via fine-tuning).

Model Word-overlap metrics Semantic equivalence metrics
SacreBLEU ROUGE-L BERTScore BLEURT MoverScore-2l
Content Planner 27.78 39.32 0.214 -0.072 0.591
BART-base 8.76 0.02 42.73 0.11 0.370 0.001 0.268 0.002 0.609 0.0004
BART-base + cond 8.73 0.02 42.60 0.12 0.369 0.001 0.268 0.003 0.608 0.0004
T5-base 18.68 0.07 47.22 0.17 0.363 0.001 0.255 0.008 0.620 0.0005
T5-base + cond 18.63 0.14 47.31 0.22 0.364 0.002 0.256 0.006 0.621 0.0008
Table 3:

Results on the BioLeaflets test set (averaged over all sections). T5 and BART models are fine-tuned with seven different random seeds: average and standard deviation are reported. BLEURT-large-128 is used.

Model Adequacy Hallucination presence Entity inclusion Fluency
Content Planner annotator 1 4.1 3.0 6.8 3.2 4.8 3.2 5.1 3.3
annotator 2 3.7 2.6 6.4 2.5 5.1 2.5 5.4 2.3
BART-base annotator 1 7.5 2.1 3.1 2.6 7.4 2.3 8.6 1.8
annotator 2 6.6 2.2 3.3 2.1 8.1 1.8 8.0 1.3
T5-base annotator 1 7.8 1.8 3.0 2.4 7.6 2.1 9.0 1.4
annotator 2 6.5 2.2 3.5 1.9 7.8 1.7 8.2 1.2
Table 4:

Human evaluation of test samples. Values on a scale from one to ten; average and standard deviation are reported. The higher the better for all quantities, expect for “Hallucination presence”. Adequacy estimates the overall generation quality, taking into consideration fluency, amount of hallucination, and entities included in the generated text.

4 Experiments

Following Kale and Rastogi (2020), we represent the structured data (i.e., detected entities) as a flat string (linearization). The entities are kept in their order of appearance (Table1b). The models are then trained to predict - starting from these entities - the corresponding published leaflet text.

We present baseline results on BioLeaflets dataset by employing the following state-of-the-art approaches:

  • Content Planner: two stages neural architecture (content selection and planning) based on LSTM  (Puduppully et al., 2019)

    . Since only relevant entities are provided as input to the model, we solely use the content planning stage (encoder-decoder architecture with an attention mechanism). We train one model for each section, and use the same hyperparameters reported by

    Puduppully et al. (2019).

  • T5: a text-to-text transfer transformer model  (Raffel et al., 2020). Kale and Rastogi (2020): showed that T5 outperforms alternatives like BERT  (Devlin et al., 2019) and GPT-2  (Radford et al., 2019). After hyperparameter search on the development dataset, the following parameters (yielding the best ROUGE-L score  (Lin, 2004)

    ) are selected: constant learning rate of 0.001, batch size of 32, 20 epochs, greedy search as a decoding method.

  • BART

    : denoising autoencoder for pretraining sequence-to-sequence models with transformers  

    (Lewis et al., 2020). For computational reasons, we use the same hyperparameters as per T5 fine-tuning.

  • BART and T5 with conditioning: we add the prefix “section_” () to the (linearized) input data. This explicitly gives the model information on the section number and thus enforces a conditioning on the section type for text generation.

BART and T5 fine-tuning are performed via HuggingFace  (Wolf et al., 2020).

5 Evaluation

Table 1 shows the generated text for one test sample as illustrative example. All generated text is made available111https://github.com/bayer-science-for-a-better-life/data2text-bioleaflets. After a thorough inspection of the samples, we conclude that generated text is generally fluent and coherent. Text produced by T5 and BART is more fluent, factually and grammatically correct than those by Content Planner. Table 3 illustrates the performance of state-of-the-art models quantified by automatic metrics.

Figure 1: Two-dimensional projections of T5 internal representations (average of the last encoder hidden-states) for pre-trained (not fine-tuned) (left) and fine-tuned T5 model on BioLeaflets dataset (right). T5 implicitly learns to condition on section type during fine-tuning.

Word-overlap metrics such as (Sacre)BLUE  (Post, 2018) and ROUGE (Lin, 2004) have been shown to perform poorly in evaluation of natural language generation (Novikova et al., 2017), and thus we report them here only for completeness. Conversely, contextual embedding based metrics BERTScore  (Zhang* et al., 2020), BLEURT  (Sellam et al., 2020), and MoverScore-2  (Zhao et al., 2019) correlate with human judgment on sentence-level and system-level evaluation. They adequately capture semantic equivalence between generated and target text as well as fluency and overall quality. T5 and BART outperform Content Planner, as measured by BERTscore, BLEURT, and MoverScore-2. T5 and BART show similar performance. These results show that transformer-based models and transfer learning strategies achieve state-of-the-art performance on data-to-text tasks, generalizing the findings in Kale and Rastogi (2020) to multi-sentence and multi-section generation, biomedical text, and low-data setting.

To confirm these findings, human evaluation is performed for Section 1 of the test set by two annotators. Results are shown in Table 4. Similarly to Manning et al. (2020), we design a survey which includes adequacy (estimate of overall quality), presence of hallucinations, entity inclusion, and fluency. T5 and BART have similar performance, and they produce more adequate text than Content Planner. T5 and BART performance is more stable across samples (lower standard deviation). These conclusions coincide with the ones drawn from Table 3, thus confirming the usefulness of semantic equivalence metrics for automatic evaluation of text generation.

Interestingly, specifying the section type in the input records (i.e., explicit conditioning) did not improve model performances (Table 3). To rationalize this result, we analyze T5 internal representations. Specifically, for each test sample, we extract the (average) last encoder hidden-state for both pre-trained (not fine-tuned) and fine-tuned T5 (fine-tuned on BioLeaflets

but without explicit conditioning). We then project these vectors into two-dimensions using the non-linear dimensionality reduction method UMAP  

(McInnes et al., 2020). The results are depicted in Fig. 1. In Fig. 1 (right), we can identify six well-separated clusters, which correspond to (the internal representations of samples belonging to) the six document sections in the BioLeaflets dataset. Thus, after fine-tuning, T5 maps input data belonging to different sections to different parts of the internal representation space. The cluster separation is much less pronounced for the pre-trained (not-fine-tuned) T5 model (Fig. 1, left). This shows that during the fine-tuning process, T5 implicitly learns to condition on section type, thus learning to generate different sections, even despite the small dataset. Since conditioning is learned automatically, explicitly passing the section type as input does not increase model performance.

6 Error Analysis and Limitations

After thorough qualitative evaluation of numerous generated samples, the following general issues appear:

  • Typos: Even though models largely utilize the input entities correctly, typos appear in generated text by T5 and BART for out-of-vocabulary words, e.g. Table 1 (c, d). Content Planner does not seem to have this problem.

  • Hallucinations

    are present for all models. Loss functions like maximum likelihood do not directly minimize hallucinations, thus hindering consistent fact-based text generation.

  • Repetitiveness: Content Planner produce repetitions (e.g. Table 1 (e)), whereas T5 and BART language models do not.

  • Difficulties in producing coherent long text: In the BioLeaflets dataset, models perform well in generating section 1, which is 962 characters long on average. However, the quality of section 4 ”Possible side effects” (3 453 characters long on average) generation is poor.

Possible improvements to our work are: analysis of the impact of shuffling of entities for the input data generation, introduction of loss functions that explicitly favor factual correctness, usage of specialized biomedical embeddings, inclusion of more source input data (e.g. part-of-speech, dependency tag), generation of longer text (beyond the 512 tokens generated here).

7 Conclusion

In this study, we introduce a new biomedical dataset (BioLeaflets), which could serve as a benchmark for biomedical text generation models. We demonstrate the feasibility of generating coherent multi-sentence biomedical text using patient-friendly language, based on input consisting of biomedical entities. These results show the potential of text generation for real-world biomedical applications. Nevertheless, human evaluation is still a required step to validate the generated samples. Application of the methodology and models used here to different sets of biomedical text (e.g., generation of selected sections of clinical study reports) could be an area for further research.

References

  • P. Bhatia, B. Celikkaya, M. Khalilia, and S. Senthivel (2019) Comprehend medical: a named entity recognition and relationship extraction web service. In

    2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA)

    ,
    Vol. , pp. 1844–1851. External Links: Document Cited by: §3.2.
  • P. Budzianowski, T. Wen, B. Tseng, I. Casanueva, S. Ultes, O. Ramadan, and M. Gašić (2018) MultiWOZ - a large-scale multi-domain Wizard-of-Oz dataset for task-oriented dialogue modelling. In

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing

    ,
    Brussels, Belgium, pp. 5016–5026. External Links: Link, Document Cited by: §2.
  • J. Devlin, M. Chang, K. Lee, and K. Toutanova (2019) BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Minneapolis, Minnesota, pp. 4171–4186. External Links: Link, Document Cited by: 2nd item.
  • E. U. EU (2001) Directive 2001/83/ec of the european parliament and of the council of 6 november 2001 on the community code relating to medicinal products for human use. Brussels, Belgium. External Links: Link Cited by: §3.
  • C. Gardent, A. Shimorina, S. Narayan, and L. Perez-Beltrachini (2017) The WebNLG challenge: generating text from RDF data. In Proceedings of the 10th International Conference on Natural Language Generation, Santiago de Compostela, Spain, pp. 124–133. External Links: Link, Document Cited by: §2.
  • M. Kale and A. Rastogi (2020) Text-to-text pre-training for data-to-text tasks. In Proceedings of the 13th International Conference on Natural Language Generation, Dublin, Ireland, pp. 97–102. External Links: Link Cited by: §2, 2nd item, §4, §5.
  • R. Lebret, D. Grangier, and M. Auli (2016) Neural text generation from structured data with application to the biography domain. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, Austin, Texas, pp. 1203–1213. External Links: Link, Document Cited by: §2.
  • M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V. Stoyanov, and L. Zettlemoyer (2020) BART: denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Online, pp. 7871–7880. External Links: Link, Document Cited by: 3rd item.
  • C. Lin (2004) ROUGE: a package for automatic evaluation of summaries. In Text Summarization Branches Out, Barcelona, Spain, pp. 74–81. External Links: Link Cited by: 2nd item, §5.
  • E. Manning, S. Wein, and N. Schneider (2020) A human evaluation of amr-to-english generation systems. In Proceedings of the 28th International Conference on Computational Linguistics, COLING 2020, Barcelona, Spain (Online), December 8-13, 2020, D. Scott, N. Bel, and C. Zong (Eds.), pp. 4773–4786. External Links: Link, Document Cited by: §5.
  • L. McInnes, J. Healy, and J. Melville (2020) UMAP: uniform manifold approximation and projection for dimension reduction. External Links: 1802.03426 Cited by: §5.
  • J. Novikova, O. Dušek, A. Cercas Curry, and V. Rieser (2017)

    Why we need new evaluation metrics for NLG

    .
    In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, Copenhagen, Denmark, pp. 2241–2252. External Links: Link, Document Cited by: §5.
  • J. Novikova, O. Dušek, and V. Rieser (2017) The E2E dataset: new challenges for end-to-end generation. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue, Saarbrücken, Germany, pp. 201–206. External Links: Link, Document Cited by: §2.
  • A. P. Parikh, X. Wang, S. Gehrmann, M. Faruqui, B. Dhingra, D. Yang, and D. Das (2020) ToTTo: a controlled table-to-text generation dataset. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pp. 1173–1186. External Links: Link, Document Cited by: §2, §3.2.
  • M. Post (2018) A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, Brussels, Belgium, pp. 186–191. External Links: Link, Document Cited by: §5.
  • R. Puduppully, L. Dong, and M. Lapata (2019) Data-to-text generation with content selection and planning.

    Proceedings of the AAAI Conference on Artificial Intelligence

    33 (01), pp. 6908–6915.
    External Links: Link, Document Cited by: 1st item.
  • P. Qi, Y. Zhang, Y. Zhang, J. Bolton, and C. D. Manning (2020) Stanza: a python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, Online, pp. 101–108. External Links: Link, Document Cited by: §3.2.
  • A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever (2019) Language models are unsupervised multitask learners. Cited by: 2nd item.
  • C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, and P. J. Liu (2020) Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research 21 (140), pp. 1–67. External Links: Link Cited by: §2, 2nd item.
  • T. Sellam, D. Das, and A. Parikh (2020) BLEURT: learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Online, pp. 7881–7892. External Links: Link, Document Cited by: §5.
  • Ö. Uzuner, B. R. South, S. Shen, and S. L. DuVall (2011) 2010 i2b2/VA challenge on concepts, assertions, and relations in clinical text. Journal of the American Medical Informatics Association 18 (5), pp. 552–556. External Links: ISSN 1067-5027, Document, Link, https://academic.oup.com/jamia/article-pdf/18/5/552/33015279/18-5-552.pdf Cited by: §3.2.
  • W. WHO (2004)

    ICD-10 : international statistical classification of diseases and related health problems : tenth revision

    .
    2nd ed edition, World Health Organization. Cited by: §3.2.
  • S. Wiseman, S. Shieber, and A. Rush (2017) Challenges in data-to-document generation. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, Copenhagen, Denmark, pp. 2253–2263. External Links: Link, Document Cited by: §2.
  • 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. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. Rush (2020) Transformers: state-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Online, pp. 38–45. External Links: Link, Document Cited by: §4.
  • Y. Zhang, Y. Zhang, P. Qi, C. D. Manning, and C. P. Langlotz (2021) Biomedical and clinical English model packages for the Stanza Python NLP library. Journal of the American Medical Informatics Association. External Links: ISSN 1527-974X Cited by: §3.2.
  • T. Zhang*, V. Kishore*, F. Wu*, K. Q. Weinberger, and Y. Artzi (2020) BERTScore: evaluating text generation with bert. In International Conference on Learning Representations, External Links: Link Cited by: §5.
  • W. Zhao, M. Peyrard, F. Liu, Y. Gao, C. M. Meyer, and S. Eger (2019) MoverScore: text generation evaluating with contextualized embeddings and earth mover distance. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, pp. 563–578. External Links: Link, Document Cited by: §5.