Automatic text summarization reduces a long text to its most important parts and generates a summary. Usually, a learning-based summarization model is developed in two basic steps: model development and model evaluation
. Given a collection of documents accompanied by one or more human-written (reference) summaries, first a set of features representing the documents is manually created or automatically extracted through supervised learning. The resulting model is then used to generate one or more (candidate) summaries, which are analyzed manually and/or with evaluation measures for their similarity to the reference summaries. These steps are iterated, optimizing the model and its parameters using a validation set. The models that perform best in the validation are selected for evaluation on the test set. With standardized test sets for each document collection, comparisons with models created earlier are reported.
However, these steps are associated with comparatively tedious tasks: During model development, summaries of individual documents are often generated and immediately evaluated to identify deficiencies and improve the model, including comparisons to other models. The latter requires third-party models to be operational despite their heterogeneous software stacks. Such “on-the-fly evaluation” during development entails that candidate and reference summaries as well as source documents are analyzed manually or by automatic measures. This multi-text comparison is often not supported by visualization, although this leads to a better understanding of the content coverage and possible selection biases of a model Vig et al. (2021); Syed et al. (2021b). The analysis of evaluation results for model selection also benefits from visual support Tenney et al. (2020). Previous research in the field of automatic summarization has not yet resulted in a unified set of tools for these purposes which is the main goal of this paper.
With Summary Workbench, we introduce the first unified combination of application and visual evaluation environments for text summarization models. Currently, it integrates 15 well-known summarization models (26 variants in total) and 10 standard evaluation measures from the literature. With FeatureSum
, it also includes a new feature-based extractive summarization model that implements features from the literature predating the deep learning era. Underlying all of the above is a specification and interface that allows easy integration of new models and measures to facilitate large-scale experiments and their reproducibility.
In what follows, Section 2 reviews related work on tools to assist summarization research and development. Section 3 overviews the key design principles of the Summary Workbench, and Section 4 provides a complete overview of all the models and measures hosted to date. Included are general-purpose models, guided models that accept user prompts to guide summary generation, and models tailored to argumentative language and to news articles. A wide range of commonly employed evaluation measures are included, covering both lexical as well as semantic overlap measures.111 Source code is available at https://github.com/webis-de/summary-workbench.
2 Related Work
The development of tools for summarization research has gained momentum recently, and several tools have been presented for (sub)tasks of the two steps above: Tools such as HuggingFace Wolf et al. (2020), FairSeq Ott et al. (2019), SummerTime Ni et al. (2021), TorchMetrics Detlefsen et al. (2022), SacreROUGEDeutsch and Roth (2020)
, PyTorch Hub,222https://pytorch.org/hub/
and TensorFlow Hub333https://www.tensorflow.org/hub/ focus on hosting several state-of-the-art text summarization models and automatic evaluation measures. These tools have significantly improved accessibility to working models. However, only some provide a very minimal interface for inference of summaries and their online/offline comparative analyses. Many authors also choose to share their models independently, be they on GitHub or elsewhere, as standalone repositories instead of integrating with any tools. To lower the bar of (latter) integration as much as possible, Summary Workbench simplifies model and measure integration as plugins (using Docker). In this way, models under development or private ones can be locally compared to others and can be archived together with all their dependencies for reproducibility. Similar efforts have been made in the information retrieval community via the Docker-based toolkit such as Anserini Yang et al. (2018).
Tools such as LIT Tenney et al. (2020), SummVis Vig et al. (2021), and Summary Explorer Syed et al. (2021b) focus on qualitative model evaluation by providing static visual analyses of the relation between the summary and its source document. Summary Workbench adapts some of their visualizations next to new ones, and complements them with interactive visual analytics for quantitative evaluation according to multiple measures. Users can explore the distribution of scores, select data points of interest and inspect them in relation to the source document to better understand the dataset.
The success of past summarization research and development has relied a lot on in-depth manual error analyses. This being one of the most laborious tasks in every natural language generation evaluation, we believe that visually comparing summaries from multiple models for many different texts, and contextualizing manual review with multiple measures is crucial to both scale up error analysis, and to better understand the capabilities and limitations of the technology. As this still requires juggling many different, incompatible tools, the unified approach ofSummary Workbench aims at lowering the bar for scaling up interactive experimentation.
3 An Interactive Visual Summarization Model Development & Evaluation Tool
Summary Workbench implements two interactive views corresponding to the two basic summarization model development steps: a summarization view and an evaluation view.
3.1 Summarization View
Figure 1 shows the summarization view, where multiple extractive/abstractive summarization models can be used to summarize texts, web pages, or scientific documents on demand, controlling for summary length. For scientific documents, relevant sections from a given paper to be summarized can be chosen. Explicit guidance signals for focused summarization can be provided as input to corresponding models (reviewed in Section 4).
Generated summaries (candidates) can be visually inspected for their lexical overlaps (highlighted on demand) with their source document or with another summary. This provides a quick overview of the models’ effectiveness at capturing important content as well as any factual errors prevalent in abstractive summarization Maynez et al. (2020). Additional functionalities include uploading multiple documents to be summarized via a single file, and command line access to all models.
3.2 Evaluation View
Figure 2 shows the evaluation view, where candidate summaries are compared with reference summaries using multiple lexical/semantic content overlap measures. Candidate summaries from multiple models, either generated using the summarization view or uploaded as a file where each example is encoded as can be evaluated. Lexical/semantic overlap of candidate/reference summaries / with the source document can also be visualized. Computed scores can be neatly exported as CSV or LaTeX tables.
Scores from the chosen evaluation measures can be further explored through an interactive plotter
. Among other things, the plotter allows users to visually correlate different evaluation measures, identify outliers/challenging source documents or strongly abstractive summaries among the candidates. This facilitates a deeper understanding of the quantitative performance of the models as well as an understanding of the evaluation datasets. Two more use cases of the interactive plotter are explained in Section5.
3.3 Plugin Server
New summarization models and evaluation measures are integrated as container-based plugins. A model/measure plugin can either be a local directory or a remote Git repository containing specification of dependent software and data (checkpoints, embeddings, lexicons), and implementations of the interfaces SummarizerPlugin and MeasurePlugin. Model metadata such as name, type, version, source, citation, and other custom arguments are provided as YAML configurations. Each plugin runs inside a Docker container with its own server that handles API calls following the OpenAPI specification.444https://www.openapis.org This setup allows users to safely self-host the entire application. Developed plugins can be easily shared with the community via DockerHub images or Git repositories. Examples are found in our tool’s technical documentation.555https://webis.de/summary-workbench/
4 Models and Measures
Summary Workbench hosts 15 extractive/abstractive summarization models and 10 lexical/semantic evaluation measures for English text. Each of these is configured as a Docker-based plugin that can be customized and instantiated accordingly. For details on model checkpoints, see Appendix A.
4.1 Summarization Models
We provide a diverse set of models applicable to multiple text domains such as news, argumentative texts, web pages, and product reviews. Model types include extractive, abstractive, supervised, unsupervised, and guided summarization, the latter requiring additional user input.
Models that work in an unsupervised fashion or leverage external knowledge via contextual embeddings are supposed to be capable of summarizing any kind of text. We provide the following models suitable for general-purpose text summarization.
FeatureSum is our new extractive summarization model which scores a sentence in the text based on a combination of standard features to identify key sentences Luhn (1958); Nenkova and McKeown (2012): TF-IDF, content units (named entities, noun phrases, numbers), position in text, mean lexical connectivity (number of tokens shared with the remaining sentences), ratio of words that are not stop words, length (relative to the longest sentence in the text), and word overlap with the title. The final score of a sentence is the product of the individual feature values. Sentences are then ranked based on these scores to produce the final summary. Different combinations of these features can be chosen by simply toggling them in the interface. This also allows for dynamically reproducing existing models from the literature provided their specific feature sets are available.
TextRank Mihalcea and Tarau (2004) is a graph-based model which employs PageRank Brin and Page (1998) on the document graph consisting of sentences as nodes to compute the strength of their connections. Top-ranked sentences within a length budget are taken as the extractive summary. We also provide the two variants PositionRank and TopicRank, which consider the sentence position and its overlap with topic sentence (document’s title or its first sentence) to compute the ranking via PyTextRank Nathan (2016).
to extract key sentences in an unsupervised fashion by first clustering all sentence embeddings using k-meansHartigan and Wong (1979) and then retrieving those closest to the centroids as the summary.
PMISum Padmakumar and He (2021) is an unsupervised extractive model that includes measures to score the relevance and redundancy of the sentences of the source document. These measures are based on pointwise mutual information (PMI) computed by pre-trained language models. Summary sentences are selected via a greedy algorithm to maximize relevance and minimize redundancy.
LoBART Manakul and Gales (2021) addresses the input length limitations of transformers Vaswani et al. (2017) that restrict capturing long-span dependencies in long document summarization. Local self-attention and explicit content selection modules are introduced to effectively summarize long documents such as podcast transcripts and scientific documents.
The following models accept explicit inputs provided by users to guide the summarization process towards generating user-specific summaries.
Biased TextRank Kazemi et al. (2020) is an extension of the TextRank model which takes an explicit user input as the “focus”, represented via contextual embeddings to guide the ranking of the document sentences. Summary extraction is based on the semantic alignment between the document sentences and the provided focus signal.
GSum Dou et al. (2021) is a guidance-based abstractive model that takes different types of external guidance signals: text inputs, highlighted sentences, keywords, or extractive oracle summaries derived from the training data. These signals along with the source text are used to generate focused and faithful abstractive summaries.
Summarizing argumentative texts (opinions, product reviews) requires that the model be able to identify high-quality, informative, and argumentative sentences from the text. We provide three models specifically developed for this task.
ArgsRank Alshomary et al. (2020) is an extractive model for creating argument snippets. It augments TextRank with two new criteria: centrality in context and argumentativeness to help the model retrieve important and argumentative sentences.
ConcluGen Syed et al. (2021a) is a transformer model for generating informative conclusions of argumentative texts by balancing the trade-off between abstractiveness and informativeness of the output. It was finetuned on the Conclugen corpus comprised of pairs of argumentative text and a human-written conclusion.
COOP Iso et al. (2021)
is an unsupervised opinion summarization model that employs latent vector aggregation by searching for optimal input combinations of sentence embeddings to address the summary vector degeneration problem caused by simple averaging. Specifically, it finds convex combinations that maximize the word overlap between the source document and its summary.
A majority of the existing summarization models are trained on news datasets, since news have been and are readily available. These models have shown strong performance in creating fluent abstractive summaries Huang et al. (2020). We provide the following models for summarizing news.
BART Lewis et al. (2020)
is a transformer denoising autoencoder for pre-training sequence-to-sequence models. Its main objective is to reconstruct the source text corrupted by employing arbitrary noising functions (masking text spans, randomly shuffling sentences) which helps the model learn better representations of the source texts for text summarizationHuang et al. (2020).
T5 Raffel et al. (2020)
is a unified text-to-text transformer model that exploits the strengths of transfer learning on a variety of problems that can be modeled as text generation tasks. A task-specific prefix is added to each input sequence (e.g., “summarize:
Pegasus Zhang et al. (2020a) is a transformer model pre-trained with a self-supervised summarization-specific training objective called “gap-sentences generation”: important sentences are removed/masked from the source text and must be jointly generated as output from the remaining sentences, similar to an extractive summary.
CLIFF Cao and Wang (2021) leverages contrastive learning for generating abstractive summaries that are faithfully and factually consistent with the source texts. Reference summaries are used as positive examples while automatically generated erroneous summaries are used as the negative examples for training the model.
4.2 Evaluation Measures
Evaluation measures for summarization typically quantify the lexical/semantic overlap of a candidate summary with a reference summary. We provide the following measures covering both.
Measures based on lexical overlap return precision, recall, or F1 scores on varying granularities of text between the candidate summary and one or more reference summaries.
BLEU Papineni et al. (2002) is a standard measure for machine translation adapted for summarization. It includes a brevity penalty to account for length differences while computing n-gram overlap.
ROUGE Lin (2004) is the most common measure for summarization which computes precision, recall, and F1 scores based on n-gram overlap, where n-grams include unigrams, bigrams, and the longest common subsequence.
METEOR Banerjee and Lavie (2005)
aligns a candidate with a set of references by mapping each unigram of a candidate to 0/1 unigrams of the reference based on exact, stem, synonym, and paraphrase matches. It then computes precision, recall, and F9 scores (i.e., weighted harmonic mean, strongly emphasizing recall) based on that.
Measures based on semantic overlap compute the semantic alignment between candidates and references at the token/word/sentence level based on their static/contextual embeddings.
Greedy Matching Rus and Lintean (2012)
aligns a candidate and a reference by greedily matching each candidate word to a reference word based on their embeddings’ cosine similarity. Average similarity over all candidate words aligned to reference words and vice versa are computed whose average is the final score.
MoverScore Zhao et al. (2019) combines contextual embeddings from BERT using the word mover’s distance Kusner et al. (2015) to compare a candidate against a set of references by considering both the amount of shared content as well as the extent of deviation between them.
BERTScore Zhang et al. (2020b) computes a similarity score for each candidate token with each reference token using contextual embeddings from BERT. The measure is also robust to adversarial modifications of the generated text.
BLEURT Sellam et al. (2020) is a learned measure based on BERT that models human judgments with a few thousand biased training examples. The model is pre-trained using millions of synthetic examples created via scores from existing measures (BLEU, ROUGE, BERTScore), and textual entailment, for better generalization.
BARTScore Yuan et al. (2021)
uses the weighted log probability of generating one text given another to compute faithfulness (sourcecandidate), precision (reference candidate), recall (candidate reference), and the F1 score.
5 Interaction Use Cases
Figure 3 shows two use cases of the interactive plotter. First, users can analyze any correlation between two measures of choice for a summarization model. Here, we find that MoverScore and BERTScore have strong correlation as they both employ contextual embeddings from BERT to compute the overlap between candidate and reference summaries. Likewise, we find that the static token embeddings from Spacy have a broader distribution of scores in comparison.
As a second use case, the interactive plotter allows comparing two variants of the same model architecture using any measure. Here, we inspect the T5 model (its 3B and 11B variants) using BERTScore to find that the larger variant generates a summary very similar to the reference while the smaller variant creates a summary that is topically related but not accurate in comparison to the reference.
In this paper we present Summary Workbench, a tool that unifies the application and evaluation of text summarization models. The tool supports integrating summarization models and evaluation measures of all kinds via a Docker-based plugin system that can also be locally deployed. This allows safe inspection and comparison of models on existing benchmarks and easy sharing with the research community in a software stack-agnostic manner. We have curated an initial set of 15 models (26 including all variants) and 10 evaluation measures and welcome contributions from the text summarization community. An extension of the tool’s features to related text generation tasks such as paraphrasing and question answering is foreseen.
7 Ethical Statement and Limitations
Our tool builds on open source models and evaluation measures contributed by the corresponding authors. We expect all users of our tool to diligently cite the authors of all models and measures when they use them via our tool, instead of just citing our tool. The tool provides direct links to the relevant sources for each hosted summarization model and evaluation measure to facilitate this.
The models and measures may have intrinsic biases, which ideally, our tool may help to identify. However, our tool itself may have biases, especially with respect to its visualizations: Visualization in general is a difficult task, and visual analytics for data analysis in particular may lead to invalid conclusions if the underlying visualization itself is flawed. Although we did our best to avoid any non-standard visualizations and relied on widely used tools to plot them, we caution users of possible errors from either the dependent libraries or their integration in our tool. Validating a visual analytics tool poses non-trivial research tasks of its own right, which we leave for future work. We do hope that the community will diligently report any errors they may encounter.
To account for color blindness, we strived to provide multiple (soft) gradient-based color schemes and grayscale colors (Figure 4) for all our visualizations. Instant previews of each color scheme are available to help users customize the tool’s visuals.
- Extractive snippet generation for arguments. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July 25-30, 2020, J. Huang, Y. Chang, X. Cheng, J. Kamps, V. Murdock, J. Wen, and Y. Liu (Eds.), pp. 1969–1972. External Links: Cited by: §4.
- METEOR: an automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization@ACL 2005, Ann Arbor, Michigan, USA, June 29, 2005, J. Goldstein, A. Lavie, C. Lin, and C. R. Voss (Eds.), pp. 65–72. External Links: Cited by: §4.
- Longformer: the long-document transformer. CoRR abs/2004.05150. External Links: Cited by: §4.
- The anatomy of a large-scale hypertextual web search engine. Comput. Networks 30 (1-7), pp. 107–117. External Links: Cited by: §4.
CLIFF: contrastive learning for improving faithfulness and factuality in abstractive summarization.
Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, M. Moens, X. Huang, L. Specia, and S. W. Yih (Eds.), pp. 6633–6649. External Links: Cited by: §4.
- TorchMetrics - Measuring Reproducibility in PyTorch External Links: Cited by: §2.
SacreROUGE: an open-source library for using and developing summarization evaluation metrics. In Proceedings of Second Workshop for NLP Open Source Software (NLP-OSS), pp. 120–125. External Links: Cited by: §2.
- 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, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), J. Burstein, C. Doran, and T. Solorio (Eds.), pp. 4171–4186. External Links: Cited by: §4.
- GSum: A general framework for guided neural abstractive summarization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, Online, June 6-11, 2021, K. Toutanova, A. Rumshisky, L. Zettlemoyer, D. Hakkani-Tür, I. Beltagy, S. Bethard, R. Cotterell, T. Chakraborty, and Y. Zhou (Eds.), pp. 4830–4842. External Links: Cited by: §4.
- Algorithm as 136: a k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics) 28 (1), pp. 100–108. Cited by: §4.
- spaCy: Industrial-strength Natural Language Processing in Python. External Links: Cited by: §4.
- What have we achieved on text summarization?. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, B. Webber, T. Cohn, Y. He, and Y. Liu (Eds.), pp. 446–469. External Links: Cited by: §4, §4.
- Convex aggregation for opinion summarization. In Findings of the Association for Computational Linguistics: EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 16-20 November, 2021, M. Moens, X. Huang, L. Specia, and S. W. Yih (Eds.), pp. 3885–3903. External Links: Cited by: §4.
- Biased textrank: unsupervised graph-based content extraction. 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. 1642–1652. External Links: Cited by: §4.
From word embeddings to document distances.
Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, F. R. Bach and D. M. Blei (Eds.), JMLR Workshop and Conference Proceedings, Vol. 37, pp. 957–966. External Links: Cited by: §4.
- 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, ACL 2020, Online, July 5-10, 2020, D. Jurafsky, J. Chai, N. Schluter, and J. R. Tetreault (Eds.), pp. 7871–7880. External Links: Cited by: §4.
- ROUGE: a package for automatic evaluation of summaries. In Text Summarization Branches Out, Barcelona, Spain, pp. 74–81. External Links: Cited by: §4.
- RoBERTa: A robustly optimized BERT pretraining approach. CoRR abs/1907.11692. External Links: Cited by: §4.
- The automatic creation of literature abstracts. IBM Journal of Reseach and Devopment 2 (2), pp. 159–165. Cited by: §4.
- Long-span summarization via local attention and content selection. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, C. Zong, F. Xia, W. Li, and R. Navigli (Eds.), pp. 6026–6041. External Links: Cited by: §4.
- On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, D. Jurafsky, J. Chai, N. Schluter, and J. R. Tetreault (Eds.), pp. 1906–1919. External Links: Cited by: §3.1.
- TextRank: bringing order into text. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing , EMNLP 2004, A meeting of SIGDAT, a Special Interest Group of the ACL, held in conjunction with ACL 2004, 25-26 July 2004, Barcelona, Spain, pp. 404–411. External Links: Cited by: §4.
- Leveraging BERT for extractive text summarization on lectures. CoRR abs/1906.04165. External Links: Cited by: §4.
- PyTextRank, a Python implementation of TextRank for phrase extraction and summarization of text documents External Links: Cited by: §4.
- A survey of text summarization techniques. In Mining text data, pp. 43–76. Cited by: §4.
- SummerTime: text summarization toolkit for non-experts. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, EMNLP 2021, Online and Punta Cana, Dominican Republic, 7-11 November, 2021, H. Adel and S. Shi (Eds.), pp. 329–338. External Links: Cited by: §2.
- Fairseq: A fast, extensible toolkit for sequence modeling. In 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, Demonstrations, W. Ammar, A. Louis, and N. Mostafazadeh (Eds.), pp. 48–53. External Links: Cited by: §2.
- Unsupervised extractive summarization using pointwise mutual information. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021, P. Merlo, J. Tiedemann, and R. Tsarfaty (Eds.), pp. 2505–2512. External Links: Cited by: §4.
- Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, 2002, Philadelphia, PA, USA, pp. 311–318. External Links: Cited by: §4.
- Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. 21, pp. 140:1–140:67. External Links: Cited by: §4.
- Sentence-bert: sentence embeddings using siamese bert-networks. 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 2019, Hong Kong, China, November 3-7, 2019, K. Inui, J. Jiang, V. Ng, and X. Wan (Eds.), pp. 3980–3990. External Links: Cited by: §4.
- Leveraging pre-trained checkpoints for sequence generation tasks. Trans. Assoc. Comput. Linguistics 8, pp. 264–280. External Links: Cited by: §4.
- A comparison of greedy and optimal assessment of natural language student input using word-to-word similarity metrics. In Proceedings of the Seventh Workshop on Building Educational Applications Using NLP, BEA@NAACL-HLT 2012, June 7, 2012, Montréal, Canada, J. R. Tetreault, J. Burstein, and C. Leacock (Eds.), pp. 157–162. External Links: Cited by: §4.
- BLEURT: learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, D. Jurafsky, J. Chai, N. Schluter, and J. R. Tetreault (Eds.), pp. 7881–7892. External Links: Cited by: §4.
- Generating informative conclusions for argumentative texts. In Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, Online Event, August 1-6, 2021, C. Zong, F. Xia, W. Li, and R. Navigli (Eds.), Findings of ACL, Vol. ACL/IJCNLP 2021, pp. 3482–3493. External Links: Cited by: §4.
- Summary explorer: visualizing the state of the art in text summarization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, EMNLP 2021, Online and Punta Cana, Dominican Republic, 7-11 November, 2021, H. Adel and S. Shi (Eds.), pp. 185–194. External Links: Cited by: §1, §2.
- The language interpretability tool: extensible, interactive visualizations and analysis for NLP models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, EMNLP 2020 - Demos, Online, November 16-20, 2020, Q. Liu and D. Schlangen (Eds.), pp. 107–118. External Links: Cited by: §1, §2.
- Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, I. Guyon, U. von Luxburg, S. Bengio, H. M. Wallach, R. Fergus, S. V. N. Vishwanathan, and R. Garnett (Eds.), pp. 5998–6008. External Links: Cited by: §4.
- CIDEr: consensus-based image description evaluation. In , pp. 4566–4575. External Links: Cited by: §4.
- SummVis: interactive visual analysis of models, data, and evaluation for text summarization. CoRR abs/2104.07605. External Links: Cited by: §1, §2.
- 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: Cited by: §2.
- Anserini: reproducible ranking baselines using lucene. ACM J. Data Inf. Qual. 10 (4), pp. 16:1–16:20. External Links: Cited by: §2.
- BARTScore: evaluating generated text as text generation. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, M. Ranzato, A. Beygelzimer, Y. N. Dauphin, P. Liang, and J. W. Vaughan (Eds.), pp. 27263–27277. External Links: Cited by: §4.
- PEGASUS: pre-training with extracted gap-sentences for abstractive summarization. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, Proceedings of Machine Learning Research, Vol. 119, pp. 11328–11339. External Links: Cited by: §4.
- BERTScore: evaluating text generation with BERT. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020, External Links: Cited by: §4.
- 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 2019, Hong Kong, China, November 3-7, 2019, K. Inui, J. Jiang, V. Ng, and X. Wan (Eds.), pp. 563–578. External Links: Cited by: §4.
Appendix A Model Details