A limited amount of studies investigates the role of model-agnostic adversarial behavior in toxic content classification. As toxicity classifiers predominantly rely on lexical cues, (deliberately) creative and evolving language-use can be detrimental to the utility of current corpora and state-of-the-art models when they are deployed for content moderation. The less training data is available, the more vulnerable models might become. This study is, to our knowledge, the first to investigate the effect of adversarial behavior and augmentation for cyberbullying detection. We demonstrate that model-agnostic lexical substitutions significantly hurt classifier performance. Moreover, when these perturbed samples are used for augmentation, we show models become robust against word-level perturbations at a slight trade-off in overall task performance. Augmentations proposed in prior work on toxicity prove to be less effective. Our results underline the need for such evaluations in online harm areas with small corpora.
Recent years have seen an increasing need for gender-neutral and inclusive language. Within the field of NLP, there are various mono- and bilingual use cases where gender inclusive language is appropriate, if not preferred due to ambiguity or uncertainty in terms of the gender of referents. In this work, we present a rule-based and a neural approach to gender-neutral rewriting for English along with manually curated synthetic data (WinoBias+) and natural data (OpenSubtitles and Reddit) benchmarks. A detailed manual and automatic evaluation highlights how our NeuTral Rewriter, trained on data generated by the rule-based approach, obtains word error rates (WER) below 0.18% on synthetic, in-domain and out-domain test sets.
Written language contains stylistic cues that can be exploited to automatically infer a variety of potentially sensitive author information. Adversarial stylometry intends to attack such models by rewriting an author’s text. Our research proposes several components to facilitate deployment of these adversarial attacks in the wild, where neither data nor target models are accessible. We introduce a transformer-based extension of a lexical replacement attack, and show it achieves high transferability when trained on a weakly labeled corpus—decreasing target model performance below chance. While not completely inconspicuous, our more successful attacks also prove notably less detectable by humans. Our framework therefore provides a promising direction for future privacy-preserving adversarial attacks.
This paper describes the CACAPO dataset, built for training both neural pipeline and end-to-end data-to-text language generation systems. The dataset is multilingual (Dutch and English), and contains almost 10,000 sentences from human-written news texts in the sports, weather, stocks, and incidents domain, together with aligned attribute-value paired data. The dataset is unique in that the linguistic variation and indirect ways of expressing data in these texts reflect the challenges of real world NLG tasks.
The task of obfuscating writing style using sequence models has previously been investigated under the framework of obfuscation-by-transfer, where the input text is explicitly rewritten in another style. A side effect of this framework are the frequent major alterations to the semantic content of the input. In this work, we propose obfuscation-by-invariance, and investigate to what extent models trained to be explicitly style-invariant preserve semantics. We evaluate our architectures in parallel and non-parallel settings, and compare automatic and human evaluations on the obfuscated sentences. Our experiments show that the performance of a style classifier can be reduced to chance level, while the output is evaluated to be of equal quality to models applying style-transfer. Additionally, human evaluation indicates a trade-off between the level of obfuscation and the observed quality of the output in terms of meaning preservation and grammaticality.
The majority of research on extracting missing user attributes from social media profiles use costly hand-annotated labels for supervised learning. Distantly supervised methods exist, although these generally rely on knowledge gathered using external sources. This paper demonstrates the effectiveness of gathering distant labels for self-reported gender on Twitter using simple queries. We confirm the reliability of this query heuristic by comparing with manual annotation. Moreover, using these labels for distant supervision, we demonstrate competitive model performance on the same data as models trained on manual annotations. As such, we offer a cheap, extensible, and fast alternative that can be employed beyond the task of gender classification.
Word embeddings have recently seen a strong increase in interest as a result of strong performance gains on a variety of tasks. However, most of this research also underlined the importance of benchmark datasets, and the difficulty of constructing these for a variety of language-specific tasks. Still, many of the datasets used in these tasks could prove to be fruitful linguistic resources, allowing for unique observations into language use and variability. In this paper we demonstrate the performance of multiple types of embeddings, created with both count and prediction-based architectures on a variety of corpora, in two language-specific tasks: relation evaluation, and dialect identification. For the latter, we compare unsupervised methods with a traditional, hand-crafted dictionary. With this research, we provide the embeddings themselves, the relation evaluation task benchmark for use in further research, and demonstrate how the benchmarked embeddings prove a useful unsupervised linguistic resource, effectively used in a downstream task.
In most languages, new words can be created through the process of compounding, which combines two or more words into a new lexical unit. Whereas in languages such as English the components that make up a compound are separated by a space, in languages such as Finnish, German, Afrikaans and Dutch these components are concatenated into one word. Compounding is very productive and leads to practical problems in developing machine translators and spelling checkers, as newly formed compounds cannot be found in existing lexicons. The Automatic Compound Processing (AuCoPro) project deals with the analysis of compounds in two closely-related languages, Afrikaans and Dutch. In this paper, we present the development and evaluation of two datasets, one for each language, that contain compound words with annotated compound boundaries. Such datasets can be used to train classifiers to identify the compound components in novel compounds. We describe the process of annotation and provide an overview of the annotation guidelines as well as global properties of the datasets. The inter-rater agreements between the annotators are considered highly reliable. Furthermore, we show the usability of these datasets by building an initial automatic compound boundary detection system, which assigns compound boundaries with approximately 90% accuracy.