ExplainaBoard: An Explainable Leaderboard for NLP

Introduction

Natural language processing (NLP) research has been and is making astounding strides forward. This is true both for classical tasks such as machine translation Sutskever et al. (2014); Wu et al. (2016), as well as for new tasks Lu et al. (2016); Rajpurkar et al. (2016), domains Beltagy et al. (2019), and languages Conneau and Lample (2019). One way this progress is quantified is through leaderboards, which report and update performance numbers of state-of-the-art systems on one or more tasks. Some prototypical leaderboards include GLUE and SuperGLUE for natural language understanding Wang et al. (2018, 2019), XTREME and XGLUE Hu et al. (2020); Liang et al. (2020) for multilingual understanding, the WMT shared tasks Barrault et al. (2020) for machine translation, and GEM and GENIE for natural language generation Gehrmann et al. (2021); Khashabi et al. (2021), among many others.

These leaderboards serve an important purpose: they provide a standardized evaluation setup, often on multiple tasks, that eases reproducible model comparison across organizations. However, at the same time, due to the prestige imbued by a top, or high, place on a leaderboard they also can result in a singular focus on raising evaluation numbers at the cost of deeper scientific understanding of model properties (Ethayarajh and Jurafsky, 2020). In particular, we argue that, among others, the following are three major limitations of the existing leaderboard paradigm:

Interpretability: Most existing leaderboards commonly use a single number to summarize system performance holistically. This is conducive to system ranking but at the same time, the results are opaque, making the strengths and weaknesses of systems less interpretable.

Interactivity: Existing leaderboards are static and non-interactive, which limits the ability of users to dig deeper into the results. Thus, (1) they usually do not flexibly support more complex evaluation settings (e.g. multi-dataset, multi-metric, multi-language) (2) users may miss opportunities to understand the relationships between different systems. For example, where does model A outperform model B? Would the performance be further improved if we combine the Top-$3$ models?

Reliability: Given the increasing role that leaderboards have taken in guiding NLP research, it is important that information expressed in them is reliable, especially on datasets with small sample sizes, but most current leaderboards do not give an idea of the reliability of system rankings.

In this paper, we describe ExplainaBoard (see Fig.1), a software package and hosted leaderboard that satisfies all of the above desiderata. It also serves as a prototype implementation of some desirable features that may be included in future leaderboards, even independent of the provided software itself. We have deployed ExplainaBoard for 9 different tasks and 41 different datasets, and demonstrate how it can be easily adapted to new tasks of interest.

We expect that ExplainaBoard will benefit different steps of the research process:

(i) System Developement: ExplainaBoard provides more detailed information regarding the submitted systems (e.g. fine-grained results, confidence intervals), allowing system developers to diagnose successes and failures of their own systems, or compare their systems with baselines and understand where improvements of their proposed methods come from. This better understanding can lead to more efficient and effective system improvements. Additionally, ExplainaBoard can help system developers uncover their systems’ advantages over others, even when these systems have not achieved state-of-the-art performance holistically. (ii) Leaderboard Organization: The ExplainaBoard software both provides a ready-made platform for easy leaderboard development over different NLP tasks, and helps upgrade traditional leaderboards to allow for more fine-grained analysis. For example, we have already established an ExplainaBoard http://explainaboard.nlpedia.ai/leaderboard/xtreme/ for the existing XTREME benchmark.https://sites.research.google/xtreme/ (iii) Broad Analysis and Understanding: Because ExplainaBoard encourages system developers to provide their system outputs in an easy-to-analyze format, these will also help researchers, particularly those just starting in a particular NLP sub-field, get a broad sense of what current state-of-the-art models can and cannot do. This not only helps them quickly track the progress of different areas, but also can allow them to understand the relative advantages of diverse systems, suggesting insightful ideas for what’s left and what’s next.Since the first release of ExplainaBoard, we have received invitations from multiple companies, startups, and researchers to collaborate, and we are working together to make it better for the community.

ExplainaBoard

As stated above, ExplainaBoard extends existing leaderboards, improving their interpretability, interactivity, and reliability. It does so by providing a number of functionalities that are applicable to a wide variety of NLP tasks (as illustrated in Tab. 1). Many of these functionalities are grounded in existing research on evaluation and fine-grained diagnostics.

Interpretable evaluation Popović and Ney (2011); Stymne (2011); Neubig et al. (2019); Fu et al. (2020a), is a research area that considers methods that break down the holistic performance of each system into different interpretable groups. For example, in a Named Entity Recognition (NER) task, we may examine the accuracy along different dimensions of a concerned entity (such as “entity frequency,” telling us how well the model does on entities that appear in the training data a certain number of times) or sentences (such as “sentence length,” telling us how well the model does on entities that appear in longer or shorter sentences) Fu et al. (2020a). This makes it possible to understand where models do well and poorly, leading to further insights beyond those that can be gleaned by holistic evaluation numbers. Applying this to a new task involves the following steps: (i) Attribute definition: define attributes by which we can partition the test set into different groups. (ii) Bucketing: partition into different buckets based on defined attributes and calculate performance w.r.t each bucket.

Generally, previous work on interpretable evaluation has been performed over single tasks, while ExplainaBoard allows for comprehensive evaluation of different types of tasks in a single software package. We concretely show several ways interpretable evaluation can be defined within ExplainaBoard below:

For an individual system as input, generate a performance histogram that highlights the buckets where it performs well or poorly. For example, in Tab. 1 we demonstrate an example from NER where the input system does worse in dealing with longer entities ($\text{eLen}\geq 4$).

F2: Pairwise Analysis: Where is one system better (worse) than another?

Given a pair of systems, interpret where the performance gap occurs. Researchers could flexibly choose two systems they are interested in (e.g. selecting two rows from the leaderboard), and ExplainaBoard will output a performance gap histogram to describe how the performance differences change over different buckets of different attributes. Tab. 1 demonstrates how we can see one system is better than the other at longer or shorter entities.

F3: Data Bias Analysis: What are the characteristics of different evaluated datasets?

The defined attributes do not only help us interpret system performance, but also make it possible for users to take a closer look at characteristics of diverse datasets. For example, from Fig. 1 shows an example of analyzing differences in average entity length across several datasets.

2 Interactivity

ExplainaBoard also allows users to dig deeper, interacting with the results in more complex ways.

ExplainaBoard provides flexible fine-grained error analyses based on the above-described performance evaluation:

Users can choose multiple systems and see their common error cases, which can be useful to identify challenging samples or even annotation errors.

In single-system analysis, users can choose particular buckets in the performance histogramEach bin of the performance histogram is clickable, returning an error case table. and see corresponding error samples in that bucket (e.g. which long entities does the current system mispredict?).

In pairwise analysis, users can select a bucket, and the unique errors (e.g. system A succeeds while B fails and vice versa) of two models will be displayed.

F5: System Combination: Is there potential complementarity between different systems?

System combination Ting and Witten (1997); González-Rubio et al. (2011); Duh et al. (2011) is a technique to improve performance by combining the output from multiple existing systems. In ExplainaBoard, users can choose multiple systems and obtain combined results calculated by voting over multiple base systems.With the system combination button of Explainaboard, we observed the-state-of-the art performance of some tasks (e.g., NER, Chunking) can be further improved. In practice, for NER task, we use the recently proposed SpanNer Fu et al. (2021) as a combiner, and for text summarization we employed Refactor, a state-of-the-art ensemble approach Liu et al. (2021). Regarding the other tasks, we adopt the majority voting method for system combination.

3 Reliability

The experimental conclusions obtained from the evaluation metrics are not necessarily statistically reliable, especially when the experimental results can be affected by many factors. ExplainaBoard also makes a step towards more reliable interpretable evaluation.

ExplainaBoard can perform confidence analysis over both holistic and fine-grained performance metrics. As shown in Tab. 1, for each bucket, there is an error bar whose width reflects how reliable the performance value is. We claim this is an important feature for fine-grained analysis since the numbers of test samples in each bucket are imbalanced, and with the confidence interval, one could know how much uncertainty there is. In practice, we use bootstrapping method Efron (1992); Ye et al. (2021) to calculate the confidence interval.

F7: Calibration Analysis: How well is the confidence of prediction calibrated with its correctness?

One commonly-cited issue with modern neural predictors is that their probability estimates are not accurate (i.e. they are poorly calibrated), often being over-confident in the correctness of their predictions Guo et al. (2017). We also incorporate this feature into ExplainaBoard, allowing users to evaluate how well-calibrated their systems of interest are.

Tasks, Datasets and Systems

We have already added to ExplainaBoard 12 NLP tasks, 50 datasets, and 400 models,265 of these models are implemented by us, as unfortunately it is currently not standard in NLP to release the system outputs that ExplainaBoard needs. which cover many or most of top-scoring systems on these tasks.We briefly describe them below, and show high-level statistics in Tab. 2.

Prediction of one or multiple pre-defined label(s) for a given input text. The current interface includes datasets for sentiment classification Pang et al. (2002), topic identification Wang and Manning (2012), and intention detection Chen et al. (2013).

Text-Span Classification

Prediction of a pre-defined class from the input of a text and a span, such as aspect-based sentiment classification task Pappas and Popescu-Belis (2014). We collect top-perform system outputs from Dai et al. (2021).

Text Pair Classification

Prediction of a class given two texts, such as the natural language inference task Bowman et al. (2015).

Sequence Labeling

Prediction of a label for each token in a sequence. The ExplainaBoard currently includes four concrete tasks: named entity recognition Tjong Kim Sang and De Meulder (2003), part-of-speech tagging Toutanova et al. (2003), text chunking Ando and Zhang (2005), and Chinese word segmentation Chen et al. (2015).

Structure Prediction

Prediction of a syntactic or semantic structure from text, where ExplainaBoard currently covers semantic parsing tasks Berant et al. (2013); Yu et al. (2018).

Text Generation

ExplainaBoard also considers text generation tasks, and currently mainly focuses on conditional text generation, for example, text summarization Rush et al. (2015); Liu and Lapata (2019) and machine translation . System outputs on text summarization are expanded based on the previous work’s collection Bhandari et al. (2020) as well as recently state-of-the-art systems Liu and Liu (2021) while outputs from machine translation are collected from the WMT20.http://www.statmt.org/wmt20/metrics-task.html

Case Study

Here, we briefly showcase the actual ExplainaBoard interface through a case study on analyzing state-of-the-art NER systems.

We define attributes following Fu et al. (2020a) and three of them are used below: entity length, sentence length and label of entity.

Collection of Systems Outputs

Currently, we collect system outputs by either implementing them by ourselves or collecting from other researchers Fu et al. (2020b); Schweter and Akbik (2020); Yamada et al. (2020). Using these methods, we have gathered 74 models on six NER datasets with system output information.

2 Analysis using ExplainaBoard

Fig. 2 illustrates different types of results driven by four functionality buttonsAs it is relatively challenging to define calibration in structure prediction tasks, this feature is currently only provided for classification tasks. We will explore more in the future. over the top-3 NER systems: LUKE Yamada et al. (2020), FLERT Schweter and Akbik (2020) and FLAIR Akbik et al. (2019).

Box A breaks down the performance of the top-1 system over different attributes.Due to the page limitation, we only show three. We can intuitively observe that even the state-of-the-art system does worse on longer entities. Users can further print error cases in the longer entity bucket by clicking the corresponding bin.

Box B shows the 1st system’s (LUKE) performance minus the 2nd system’s (FLERT) performance. We can see that although LUKE surpasses FLERT holistically, it performs worse when dealing with PERSON entities.

Box C identifies samples that all systems mispredict. Further analysis of these samples uncovers challenging patterns or annotation errors.

Box D examines potential complementarity among these top-3 systems. The result shows that, by a simple voting ensemble strategy, a new state-of-the-art (94.65 F1) has been achieved on the CoNLL-2003 dataset.

Usage

To show the practical utility of ExplainaBoard, we first present examples of how it has been used as an analysis tool in existing published research papers. Fu et al. (2020b) (Tab.4) utilize single-system analysis with the attribute of label consistency for NER task while Zhong et al. (2019) (Tab.4-5) use it for text summarization with attributes of density and compression. Fig.4 and Tab.3 in Fu et al. (2020a) leverage the data bias analysis and pairwise system diagnostics to interpret top-performing NER systems while Tab.4 in Fu et al. (2020c) use single and pairwise system analysis to investigate what’s next for the Chinese Word Segmentation task. Liu et al. (2021) use system combiner functionality to make ensemble analysis of summarization systems and Fig.1 in Ye et al. (2021) use reliability analysis functionality to observe how confidence intervals change in different buckets of a performance histogram.

Using ExplainaBoard

Researchers can use ExplainaBoard in different ways: (i) We maintain a website where each task-specific ExplainaBoard allows researchers to interact with it, interpreting systems and datasets that they are interested in from different perspectives. (ii) We also release our back-end code for different NLP tasks so that researchers could flexibly use them to process their own system outputs, which can assist their research projects.

Contributing to ExplainaBoard

The community can contribute to ExplainaBoard in several ways: (i) Submit system outputs of their implemented models. (ii) Add more informative attributes for different NLP tasks. (iii) Add new datasets or benchmarks for existing or new tasks.

Implications and Roadmap

ExplainaBoard presents a new paradigm in leaderboard development for NLP. This is just the beginning of its development, and there are many future directions.

Due to the ability to analyze, contrast, or combine results from many systems ExplainaBoard incentivizes researchers to submit their results to explainaboard to better understand them and showcase their systems’ strengths. At the same time, ExplainaBoard will serve as a central repository for system outputs across many tasks, allowing for future avenues of research into cross-system analysis or system combination.

Enriching ExplainaBoard with Glass-box Analysis

ExplainaBoard currently performs black-box analysis, solely analyzing system outputs without accessing model internals. On the other hand, there are many other glass-box interpretability tools that look at model internals, such as the AllenNLP Interpret Wallace et al. (2019) and Language Interpretability Tool Tenney et al. (2020). Expanding leaderboards to glass-box analysis methods (see Lipton (2018); Belinkov and Glass (2019) for a survey) is an interesting future work.

In the future, we aim to improve the applicability and usefulness by following action items: (1) Collaborate with more leaderboard organizers of diverse tasks and set up corresponding ExplainaBoards for them. (2) Cover more tasks, datasets, models, as well as functionalities.

Acknowledgements

We thanks all reviewers for their valuable comments and authors who share their system outputs with us: Ikuya Yamada, Stefan Schweter, Colin Raffel, Yang Liu, Li Dong. We also thank Vijay Viswanathan, Yiran Chen, Hiroaki Hayashi for useful discussion and feedback about ExplainaBoard.

The work was supported in part by the Air Force Research Laboratory under agreement number FA8750-19-2-0200. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the Air Force Research Laboratory or the U.S. Government.

References