We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

This case study investigates the extent to which a language model (GPT-2) is able to capture native speakers’ intuitions about implicit causality in a sentence completion task. Study 1 reproduces earlier results (showing that the model’s surprisal values correlate with the implicit causality bias of the verb; Davis and van Schijndel 2021), and then examine the effects of gender and verb frequency on model performance. Study 2 examines the reasoning ability of GPT-2: Is the model able to produce more sensible motivations for why the subject VERBed the object if the verbs have stronger causality biases? For this study we took care to avoid human raters being biased by obscenities and disfluencies generated by the model.

This paper is part of the larger ReproHum project, where different teams of researchers aim to reproduce published experiments from the NLP literature. Specifically, ReproHum focuses on the reproducibility of human evaluation studies, where participants indicate the quality of different outputs of Natural Language Generation (NLG) systems. This is necessary because without reproduction studies, we do not know how reliable earlier results are. This paper aims to reproduce the second human evaluation study of Puduppully & Lapata (2021), while another lab is attempting to do the same. This experiment uses best-worst scaling to determine the relative performance of different NLG systems. We found that the worst performing system in the original study is now in fact the best performing system across the board. This means that we cannot fully reproduce the original results. We also carry out alternative analyses of the data, and discuss how our results may be combined with the other reproduction study that is carried out in parallel with this paper.

In this paper, we describe our reproduction ef- fort of the paper: Towards Best Experiment Design for Evaluating Dialogue System Output by Santhanam and Shaikh (2019) for the 2022 ReproGen shared task. We aim to produce the same results, using different human evaluators, and a different implementation of the automatic metrics used in the original paper. Although overall the study posed some challenges to re- produce (e.g. difficulties with reproduction of automatic metrics and statistics), in the end we did find that the results generally replicate the findings of Santhanam and Shaikh (2019) and seem to follow similar trends.

We observe a severe under-reporting of the different kinds of errors that Natural Language Generation systems make. This is a problem, because mistakes are an important indicator of where systems should still be improved. If authors only report overall performance metrics, the research community is left in the dark about the specific weaknesses that are exhibited by ‘state-of-the-art’ research. Next to quantifying the extent of error under-reporting, this position paper provides recommendations for error identification, analysis and reporting.

Preregistration refers to the practice of specifying what you are going to do, and what you expect to find in your study, before carrying out the study. This practice is increasingly common in medicine and psychology, but is rarely discussed in NLP. This paper discusses preregistration in more detail, explores how NLP researchers could preregister their work, and presents several preregistration questions for different kinds of studies. Finally, we argue in favour of registered reports, which could provide firmer grounds for slow science in NLP research. The goal of this paper is to elicit a discussion in the NLP community, which we hope to synthesise into a general NLP preregistration form in future research.

We introduce GEM, a living benchmark for natural language Generation (NLG), its Evaluation, and Metrics. Measuring progress in NLG relies on a constantly evolving ecosystem of automated metrics, datasets, and human evaluation standards. Due to this moving target, new models often still evaluate on divergent anglo-centric corpora with well-established, but flawed, metrics. This disconnect makes it challenging to identify the limitations of current models and opportunities for progress. Addressing this limitation, GEM provides an environment in which models can easily be applied to a wide set of tasks and in which evaluation strategies can be tested. Regular updates to the benchmark will help NLG research become more multilingual and evolve the challenge alongside models. This paper serves as the description of the data for the 2021 shared task at the associated GEM Workshop.

Evaluations of image description systems are typically domain-general: generated descriptions for the held-out test images are either compared to a set of reference descriptions (using automated metrics), or rated by human judges on one or more Likert scales (for fluency, overall quality, and other quality criteria). While useful, these evaluations do not tell us anything about the kinds of image descriptions that systems are able to produce. Or, phrased differently, these evaluations do not tell us anything about the cognitive capabilities of image description systems. This paper proposes a different kind of assessment, that is able to quantify the extent to which these systems are able to describe humans. This assessment is based on a manual characterisation (a context-free grammar) of English entity labels in the PEOPLE domain, to determine the range of possible outputs. We examined 9 systems to see what kinds of labels they actually use. We found that these systems only use a small subset of at most 13 different kinds of modifiers (e.g. tall and short modify HEIGHT, sad and happy modify MOOD), but 27 kinds of modifiers are never used. Future research could study these semantic dimensions in more detail.

NLG researchers often use uncontrolled corpora to train and evaluate their systems, using textual similarity metrics, such as BLEU. This position paper argues in favour of two alternative evaluation strategies, using grammars or rule-based systems. These strategies are particularly useful to identify the strengths and weaknesses of different systems. We contrast our proposals with the (extended) WebNLG dataset, which is revealed to have a skewed distribution of predicates. We predict that this distribution affects the quality of the predictions for systems trained on this data. However, this hypothesis can only be thoroughly tested (without any confounds) once we are able to systematically manipulate the skewness of the data, using a rule-based approach.

Human assessment remains the most trusted form of evaluation in NLG, but highly diverse approaches and a proliferation of different quality criteria used by researchers make it difficult to compare results and draw conclusions across papers, with adverse implications for meta-evaluation and reproducibility. In this paper, we present (i) our dataset of 165 NLG papers with human evaluations, (ii) the annotation scheme we developed to label the papers for different aspects of evaluations, (iii) quantitative analyses of the annotations, and (iv) a set of recommendations for improving standards in evaluation reporting. We use the annotations as a basis for examining information included in evaluation reports, and levels of consistency in approaches, experimental design and terminology, focusing in particular on the 200+ different terms that have been used for evaluated aspects of quality. We conclude that due to a pervasive lack of clarity in reports and extreme diversity in approaches, human evaluation in NLG presents as extremely confused in 2020, and that the field is in urgent need of standard methods and terminology.

Earlier research has shown that evaluation metrics based on textual similarity (e.g., BLEU, CIDEr, Meteor) do not correlate well with human evaluation scores for automatically generated text. We carried out an experiment with Chinese speakers, where we systematically manipulated image descriptions to contain different kinds of errors. Because our manipulated descriptions form minimal pairs with the reference descriptions, we are able to assess the impact of different kinds of errors on the perceived quality of the descriptions. Our results show that different kinds of errors elicit significantly different evaluation scores, even though all erroneous descriptions differ in only one character from the reference descriptions. Evaluation metrics based solely on textual similarity are unable to capture these differences, which (at least partially) explains their poor correlation with human judgments. Our work provides the foundations for future work, where we aim to understand why different errors are seen as more or less severe.

Traditionally, most data-to-text applications have been designed using a modular pipeline architecture, in which non-linguistic input data is converted into natural language through several intermediate transformations. By contrast, recent neural models for data-to-text generation have been proposed as end-to-end approaches, where the non-linguistic input is rendered in natural language with much less explicit intermediate representations in between. This study introduces a systematic comparison between neural pipeline and end-to-end data-to-text approaches for the generation of text from RDF triples. Both architectures were implemented making use of the encoder-decoder Gated-Recurrent Units (GRU) and Transformer, two state-of-the art deep learning methods. Automatic and human evaluations together with a qualitative analysis suggest that having explicit intermediate steps in the generation process results in better texts than the ones generated by end-to-end approaches. Moreover, the pipeline models generalize better to unseen inputs. Data and code are publicly available.

Currently, there is little agreement as to how Natural Language Generation (NLG) systems should be evaluated. While there is some agreement regarding automatic metrics, there is a high degree of variation in the way that human evaluation is carried out. This paper provides an overview of how human evaluation is currently conducted, and presents a set of best practices, grounded in the literature. With this paper, we hope to contribute to the quality and consistency of human evaluations in NLG.

Task effects in NLG corpus elicitation recently started to receive more attention, but are usually not modeled statistically. We present a controlled replication of the study by Van Miltenburg et al. (2018b), contrasting spoken with written descriptions. We collected additional written Dutch descriptions to supplement the spoken data from the DIDEC corpus, and analyzed the descriptions using mixed effects modeling to account for variation between participants and items. Our results show that the effects of modality largely disappear in a controlled setting.

Automatic image description systems are commonly trained and evaluated on written image descriptions. At the same time, these systems are often used to provide spoken descriptions (e.g. for visually impaired users) through apps like TapTapSee or Seeing AI. This is not a problem, as long as spoken and written descriptions are very similar. However, linguistic research suggests that spoken language often differs from written language. These differences are not regular, and vary from context to context. Therefore, this paper investigates whether there are differences between written and spoken image descriptions, even if they are elicited through similar tasks. We compare descriptions produced in two languages (English and Dutch), and in both languages observe substantial differences between spoken and written descriptions. Future research should see if users prefer the spoken over the written style and, if so, aim to emulate spoken descriptions.

Image description datasets, such as Flickr30K and MS COCO, show a high degree of variation in the ways that crowd-workers talk about the world. Although this gives us a rich and diverse collection of data to work with, it also introduces uncertainty about how the world should be described. This paper shows the extent of this uncertainty in the PEOPLE-domain. We present a taxonomy of different ways to talk about other people. This taxonomy serves as a reference point to think about how other people should be described, and can be used to classify and compute statistics about labels applied to people.

Automatic image description systems typically produce generic sentences that only make use of a small subset of the vocabulary available to them. In this paper, we consider the production of generic descriptions as a lack of diversity in the output, which we quantify using established metrics and two new metrics that frame image description as a word recall task. This framing allows us to evaluate system performance on the head of the vocabulary, as well as on the long tail, where system performance degrades. We use these metrics to examine the diversity of the sentences generated by nine state-of-the-art systems on the MS COCO data set. We find that the systems trained with maximum likelihood objectives produce less diverse output than those trained with additional adversarial objectives. However, the adversarially-trained models only produce more types from the head of the vocabulary and not the tail. Besides vocabulary-based methods, we also look at the compositional capacity of the systems, specifically their ability to create compound nouns and prepositional phrases of different lengths. We conclude that there is still much room for improvement, and offer a toolkit to measure progress towards the goal of generating more diverse image descriptions.

We present a corpus of spoken Dutch image descriptions, paired with two sets of eye-tracking data: Free viewing, where participants look at images without any particular purpose, and Description viewing, where we track eye movements while participants produce spoken descriptions of the images they are viewing. This paper describes the data collection procedure and the corpus itself, and provides an initial analysis of self-corrections in image descriptions. We also present two studies showing the potential of this data. Though these studies mainly serve as an example, we do find two interesting results: (1) the eye-tracking data for the description viewing task is more coherent than for the free-viewing task; (2) variation in image descriptions (also called ‘image specificity’; Jas and Parikh, 2015) is only moderately correlated across different languages. Our corpus can be used to gain a deeper understanding of the image description task, particularly how visual attention is correlated with the image description process.

Automatic image description systems are commonly trained and evaluated on large image description datasets. Recently, researchers have started to collect such datasets for languages other than English. An unexplored question is how different these datasets are from English and, if there are any differences, what causes them to differ. This paper provides a cross-linguistic comparison of Dutch, English, and German image descriptions. We find that these descriptions are similar in many respects, but the familiarity of crowd workers with the subjects of the images has a noticeable influence on the specificity of the descriptions.

This research proposal discusses pragmatic factors in image description, arguing that current automatic image description systems do not take these factors into account. I present a general model of the human image description process, and propose to study this process using corpus analysis, experiments, and computational modeling. This will lead to a better characterization of human image description behavior, providing a road map for future research in automatic image description, and the automatic description of perceptual stimuli in general.

This paper presents a collection of annotations (tags or keywords) for a set of 2,133 environmental sounds taken from the Freesound database (www.freesound.org). The annotations are acquired through an open-ended crowd-labeling task, in which participants were asked to provide keywords for each of three sounds. The main goal of this study is to find out (i) whether it is feasible to collect keywords for a large collection of sounds through crowdsourcing, and (ii) how people talk about sounds, and what information they can infer from hearing a sound in isolation. Our main finding is that it is not only feasible to perform crowd-labeling for a large collection of sounds, it is also very useful to highlight different aspects of the sounds that authors may fail to mention. Our data is freely available, and can be used to ground semantic models, improve search in audio databases, and to study the language of sound.

This paper discusses the need for a dictionary of affixal negations and regular antonyms to facilitate their automatic detection in text. Without such a dictionary, affixal negations are very difficult to detect. In addition, we show that the set of affixal negations is not homogeneous, and that different NLP tasks may require different subsets. A dictionary can store the subtypes of affixal negations, making it possible to select a certain subset or to make inferences on the basis of these subtypes. We take a first step towards creating a negation dictionary by annotating all direct antonym pairs inWordNet using an existing typology of affixal negations. By highlighting some of the issues that were encountered in this annotation experiment, we hope to provide some insights into the necessary steps of building a negation dictionary.

We describe Open Dutch WordNet, which has been derived from the Cornetto database, the Princeton WordNet and open source resources. We exploited existing equivalence relations between Cornetto synsets and WordNet synsets in order to move the open source content from Cornetto into WordNet synsets. Currently, Open Dutch Wordnet contains 117,914 synsets, of which 51,588 synsets contain at least one Dutch synonym, which leaves 66,326 synsets still to obtain a Dutch synonym. The average polysemy is 1.5. The resource is currently delivered in XML under the CC BY-SA 4.0 license1 and it has been linked to the Global Wordnet Grid. In order to use the resource, we refer to: https: //github.com/MartenPostma/OpenDutchWordnet.

Le and Fokkens (2015) recently showed that taxonomy-based approaches are more reliable than corpus-based approaches in estimating human similarity ratings. On the other hand, distributional models provide much better coverage. The lack of an established similarity metric for adjectives in WordNet is a case in point. I present initial work to establish such a metric, and propose ways to move forward by looking at extensions to WordNet. I show that the shortest path distance between derivationally related forms provides a reliable estimate of adjective similarity. Furthermore, I find that a hybrid method combining this measure with vector-based similarity estimations gives us the best of both worlds: more reliable similarity estimations than vectors alone, but with the same coverage as corpus-based methods.