Translated texts or utterances bear several hallmarks distinct from texts originating in the language. This phenomenon, known as translationese, is well-documented, and when found in training or test sets can affect model performance. Still, work to mitigate the effect of translationese in human translated text is understudied. We hypothesize that Abstract Meaning Representation (AMR), a semantic representation which abstracts away from the surface form, can be used as an interlingua to reduce the amount of translationese in translated texts. By parsing English translations into an AMR graph and then generating text from that AMR, we obtain texts that more closely resemble non-translationese by macro-level measures. We show that across four metrics, and qualitatively, using AMR as an interlingua enables the reduction of translationese and we compare our results to two additional approaches: one based on round-trip machine translation and one based on syntactically controlled generation.
The evaluation of Handwritten Text Recognition (HTR) systems has traditionally used metrics based on the edit distance between HTR and ground truth (GT) transcripts, at both the character and word levels. This is very adequate when the experimental protocol assumes that both GT and HTR text lines are the same, which allows edit distances to be independently computed to each given line. Driven by recent advances in pattern recognition, HTR systems increasingly face the end-to-end page-level transcription of a document, where the precision of locating the different text lines and their corresponding reading order (RO) play a key role. In such a case, the standard metrics do not take into account the inconsistencies that might appear. In this paper, the problem of evaluating HTR systems at the page level is introduced in detail. We analyze the convenience of using a two-fold evaluation, where the transcription accuracy and the RO goodness are considered separately. Different alternatives are proposed, analyzed and empirically compared both through partially simulated and through real, full end-to-end experiments. Results support the validity of the proposed two-fold evaluation approach. An important conclusion is that such an evaluation can be adequately achieved by just two simple and well-known metrics: the Word Error Rate, that takes transcription sequentiality into account, and the here re-formulated Bag of Words Word Error Rate, that ignores order. While the latter directly and very accurately assess intrinsic word recognition errors, the difference between both metrics gracefully correlates with the Spearman's Foot Rule Distance, a metric which explicitly measures RO errors associated with layout analysis flaws.
Auditing unwanted social bias in language models (LMs) is inherently hard due to the multidisciplinary nature of the work. In addition, the rapid evolution of LMs can make benchmarks irrelevant in no time. Bias auditing is further complicated by LM brittleness: when a presumably biased outcome is observed, is it due to model bias or model brittleness? We propose enlisting the models themselves to help construct bias auditing datasets that remain challenging, and introduce bias measures that distinguish between types of model errors. First, we extend an existing bias benchmark for NLI (BBNLI) using a combination of LM-generated lexical variations, adversarial filtering, and human validation. We demonstrate that the newly created dataset (BBNLInext) is more challenging than BBNLI: on average, BBNLI-next reduces the accuracy of state-of-the-art NLI models from 95.3%, as observed by BBNLI, to 58.6%. Second, we employ BBNLI-next to showcase the interplay between robustness and bias, and the subtlety in differentiating between the two. Third, we point out shortcomings in current bias scores used in the literature and propose bias measures that take into account pro-/anti-stereotype bias and model brittleness. We will publicly release the BBNLI-next dataset to inspire research on rapidly expanding benchmarks to keep up with model evolution, along with research on the robustness-bias interplay in bias auditing. Note: This paper contains offensive text examples.
Entity names play an effective role in relation extraction (RE) and often influence model performance. As a result, the entity names in the benchmarks' test sets significantly influence the evaluation of RE models. In this work, we find that the standard RE benchmarks' datasets have a large portion of incorrect entity annotations, low entity name diversity, and are prone to have shortcuts from entity names to ground-truth relations. These issues make the standard benchmarks far from reflecting the real-world scenarios. Hence, in this work, we present EntRED, a challenging RE benchmark with reduced shortcuts and higher diversity of entities. To build EntRED, we propose an end-to-end entity replacement pipeline based on causal inference (CI): ERIC. ERIC performs type-constrained replacements on entities to reduce the shortcuts from entity bias to ground-truth relations. ERIC applies CI in two aspects: 1) targeting the instances that need entity replacements, and 2) determining the candidate entities for replacements. We apply ERIC on TACRED to produce EntRED. Our EntRED evaluates whether the RE model can correctly extract the relations from the text instead of relying on entity bias. Empirical results reveal that even the strong RE model has a significant performance drop on EntRED, which memorizes entity name patterns instead of reasoning from the textual context. We release ERIC's source code and the EntRED benchmark at https://github.com/wangywUST/ENTRED.
Information Extraction from visually rich documents is a challenging task that has gained a lot of attention in recent years due to its importance in several document-control based applications and its widespread commercial value. The majority of the research work conducted on this topic to date follow a two-step pipeline. First, they read the text using an off-the-shelf Optical Character Recognition (OCR) engine, then, they extract the fields of interest from the obtained text. The main drawback of these approaches is their dependence on an external OCR system, which can negatively impact both performance and computational speed. Recent OCR-free methods were proposed to address the previous issues. Inspired by their promising results, we propose in this paper an OCR-free end-to-end information extraction model named DocParser. It differs from prior end-to-end approaches by its ability to better extract discriminative character features. DocParser achieves state-of-the-art results on various datasets, while still being faster than previous works.
Recent advances in conditional generative image models have enabled impressive results. On the one hand, text-based conditional models have achieved remarkable generation quality, by leveraging large-scale datasets of image-text pairs. To enable fine-grained controllability, however, text-based models require long prompts, whose details may be ignored by the model. On the other hand, layout-based conditional models have also witnessed significant advances. These models rely on bounding boxes or segmentation maps for precise spatial conditioning in combination with coarse semantic labels. The semantic labels, however, cannot be used to express detailed appearance characteristics. In this paper, we approach fine-grained scene controllability through image collages which allow a rich visual description of the desired scene as well as the appearance and location of the objects therein, without the need of class nor attribute labels. We introduce "mixing and matching scenes" (M&Ms), an approach that consists of an adversarially trained generative image model which is conditioned on appearance features and spatial positions of the different elements in a collage, and integrates these into a coherent image. We train our model on the OpenImages (OI) dataset and evaluate it on collages derived from OI and MS-COCO datasets. Our experiments on the OI dataset show that M&Ms outperforms baselines in terms of fine-grained scene controllability while being very competitive in terms of image quality and sample diversity. On the MS-COCO dataset, we highlight the generalization ability of our model by outperforming DALL-E in terms of the zero-shot FID metric, despite using two magnitudes fewer parameters and data. Collage based generative models have the potential to advance content creation in an efficient and effective way as they are intuitive to use and yield high quality generations.
Generating consistent and high-quality images from given texts is essential for visual-language understanding. Although impressive results have been achieved in generating high-quality images, text-image consistency is still a major concern in existing GAN-based methods. Particularly, the most popular metric $R$-precision may not accurately reflect the text-image consistency, often resulting in very misleading semantics in the generated images. Albeit its significance, how to design a better text-image consistency metric surprisingly remains under-explored in the community. In this paper, we make a further step forward to develop a novel CLIP-based metric termed as Semantic Similarity Distance (SSD), which is both theoretically founded from a distributional viewpoint and empirically verified on benchmark datasets. Benefiting from the proposed metric, we further design the Parallel Deep Fusion Generative Adversarial Networks (PDF-GAN), which can fuse semantic information at different granularities and capture accurate semantics. Equipped with two novel plug-and-play components: Hard-Negative Sentence Constructor and Semantic Projection, the proposed PDF-GAN can mitigate inconsistent semantics and bridge the text-image semantic gap. A series of experiments show that, as opposed to current state-of-the-art methods, our PDF-GAN can lead to significantly better text-image consistency while maintaining decent image quality on the CUB and COCO datasets.
Masked Language Models (MLMs) have been successful in many natural language processing tasks. However, real-world stereotype biases are likely to be reflected in MLMs due to their learning from large text corpora. Most of the evaluation metrics proposed in the past adopt different masking strategies, designed with the log-likelihood of MLMs. They lack holistic considerations such as variance for stereotype bias and anti-stereotype bias samples. In this paper, the log-likelihoods of stereotype bias and anti-stereotype bias samples output by MLMs are considered Gaussian distributions. Two evaluation metrics, Kullback Leibler Divergence Score (KLDivS) and Jensen Shannon Divergence Score (JSDivS) are proposed to evaluate social biases in MLMs The experimental results on the public datasets StereoSet and CrowS-Pairs demonstrate that KLDivS and JSDivS are more stable and interpretable compared to the metrics proposed in the past.
Text-to-image generation (TTIG) models, a recent addition to creative AI, can generate images based on a text description. These models have begun to rival the work of professional creatives, and sparked discussions on the future of creative work, loss of jobs, and copyright issues, amongst other important implications. To support the sustainable adoption of TTIG, we must provide rich, reliable and transparent insights into how professionals perceive, adopt and use TTIG. Crucially though, the public debate is shallow, narrow and lacking transparency, while academic work has focused on studying the use of TTIG in a general artist population, but not on the perceptions and attitudes of professionals in a specific industry. In this paper, we contribute a qualitative, exploratory interview study on TTIG in the Finnish videogame industry. Through a Template Analysis on semi-structured interviews with 14 game professionals, we reveal 12 overarching themes, structured into 49 sub-themes on professionals' perception, adoption and use of TTIG systems in games industry practice. Experiencing (yet another) change of roles and creative processes, our participants' reflections can inform discussions within the industry, be used by policymakers to inform urgently needed legislation, and support researchers in games, HCI and AI to support the sustainable, professional use of TTIG to benefit people and games as cultural artefacts.
Training learnable metrics using modern language models has recently emerged as a promising method for the automatic evaluation of machine translation. However, existing human evaluation datasets in text simplification are limited by a lack of annotations, unitary simplification types, and outdated models, making them unsuitable for this approach. To address these issues, we introduce the SIMPEVAL corpus that contains: SIMPEVAL_ASSET, comprising 12K human ratings on 2.4K simplifications of 24 systems, and SIMPEVAL_2022, a challenging simplification benchmark consisting of over 1K human ratings of 360 simplifications including generations from GPT-3.5. Training on SIMPEVAL_ASSET, we present LENS, a Learnable Evaluation Metric for Text Simplification. Extensive empirical results show that LENS correlates better with human judgment than existing metrics, paving the way for future progress in the evaluation of text simplification. To create the SIMPEVAL datasets, we introduce RANK & RATE, a human evaluation framework that rates simplifications from several models in a list-wise manner by leveraging an interactive interface, which ensures both consistency and accuracy in the evaluation process. Our metric, dataset, and annotation toolkit are available at https://github.com/Yao-Dou/LENS.