The GEM Benchmark: Natural Language Generation, its Evaluation and Metrics

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. However, 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 corpora and 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 initial release for which we are organizing a shared task at our ACL 2021 Workshop and to which we invite the entire NLG community to participate.

On the use of human reference data for evaluating automatic image descriptions

Automatic image description systems are commonly trained and evaluated using crowdsourced, human-generated image descriptions. The best-performing system is then determined using some measure of similarity to the reference data (BLEU, Meteor, CIDER, etc). Thus, both the quality of the systems as well as the quality of the evaluation depends on the quality of the descriptions. As Section 2 will show, the quality of current image description datasets is insufficient. I argue that there is a need for more detailed guidelines that take into account the needs of visually impaired users, but also the feasibility of generating suitable descriptions. With high-quality data, evaluation of image description systems could use reference descriptions, but we should also look for alternatives.

Neural data-to-text generation: A comparison between pipeline and end-to-end architectures

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. In 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 state-of-the art deep learning methods as the encoder-decoder Gated-Recurrent Units (GRU) and Transformer. 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.

Cross-linguistic differences and similarities in image descriptions

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 description specificity.

Room for improvement in automatic image description: an error analysis

In recent years we have seen rapid and significant progress in automatic image description but what are the open problems in this area? Most work has been evaluated using text-based similarity metrics, which only indicate that there have been improvements, without explaining what has improved. In this paper, we present a detailed error analysis of the descriptions generated by a state-of-the-art attention-based model. Our analysis operates on two levels: first we check the descriptions for accuracy, and then we categorize the types of errors we observe in the inaccurate descriptions. We find only 20% of the descriptions are free from errors, and surprisingly that 26% are unrelated to the image. Finally, we manually correct the most frequently occurring error types (e.g. gender identification) to estimate the performance reward for addressing these errors, observing gains of 0.2--1 BLEU point per type.

Pragmatic factors in image description: the case of negations

We provide a qualitative analysis of the descriptions containing negations (no, not, n't, nobody, etc) in the Flickr30K corpus, and a categorization of negation uses. Based on this analysis, we provide a set of requirements that an image description system should have in order to generate negation sentences. As a pilot experiment, we used our categorization to manually annotate sentences containing negations in the Flickr30K corpus, with an agreement score of K=0.67. With this paper, we hope to open up a broader discussion of subjective language in image descriptions.

Stereotyping and Bias in the Flickr30K Dataset

An untested assumption behind the crowdsourced descriptions of the images in the Flickr30K dataset (Young et al., 2014) is that they "focus only on the information that can be obtained from the image alone" (Hodosh et al., 2013, p. 859). This paper presents some evidence against this assumption, and provides a list of biases and unwarranted inferences that can be found in the Flickr30K dataset. Finally, it considers methods to find examples of these, and discusses how we should deal with stereotype-driven descriptions in future applications.

Detecting and ordering adjectival scalemates

This paper presents a pattern-based method that can be used to infer adjectival scales, such as <lukewarm, warm, hot>, from a corpus. Specifically, the proposed method uses lexical patterns to automatically identify and order pairs of scalemates, followed by a filtering phase in which unrelated pairs are discarded. For the filtering phase, several different similarity measures are implemented and compared. The model presented in this paper is evaluated using the current standard, along with a novel evaluation set, and shown to be at least as good as the current state-of-the-art.