Computer Science

We study estimating inherent human disagreement (annotation label distribution) in natural language inference task. Post-hoc smoothing of the predicted label distribution to match the expected label entropy is very effective. Such simple manipulation can reduce KL divergence by almost half, yet will not improve majority label prediction accuracy or learn label distributions. To this end, we introduce a small amount of examples with multiple references into training. We depart from the standard practice of collecting a single reference per each training example, and find that collecting multiple references can achieve better accuracy under the fixed annotation budget. Lastly, we provide rich analyses comparing these two methods for improving label distribution estimation.

Biased associations have been a challenge in the development of classifiers for detecting toxic language, hindering both fairness and accuracy. As potential solutions, we investigate recently introduced debiasing methods for text classification datasets and models, as applied to toxic language detection. Our focus is on lexical (e.g., swear words, slurs, identity mentions) and dialectal markers (specifically African American English). Our comprehensive experiments establish that existing methods are limited in their ability to prevent biased behavior in current toxicity detectors. We then propose an automatic, dialect-aware data correction method, as a proof-of-concept. Despite the use of synthetic labels, this method reduces dialectal associations with toxicity. Overall, our findings show that debiasing a model trained on biased toxic language data is not as effective as simply relabeling the data to remove existing biases.

Much previous work characterizing language variation across Internet social groups has focused on the types of words used by these groups. We extend this type of study by employing BERT to characterize variation in the senses of words as well, analyzing two months of English comments in 474 Reddit communities. The specificity of different sense clusters to a community, combined with the specificity of a community's unique word types, is used to identify cases where a social group's language deviates from the norm. We validate our metrics using user-created glossaries and draw on sociolinguistic theories to connect language variation with trends in community behavior. We find that communities with highly distinctive language are medium-sized, and their loyal and highly engaged users interact in dense networks.

How do adults understand children's speech? Children's productions over the course of language development often bear little resemblance to typical adult pronunciations, yet caregivers nonetheless reliably recover meaning from them. Here, we employ a suite of Bayesian models of spoken word recognition to understand how adults overcome the noisiness of child language, showing that communicative success between children and adults relies heavily on adult inferential processes. By evaluating competing models on phonetically-annotated corpora, we show that adults' recovered meanings are best predicted by prior expectations fitted specifically to the child language environment, rather than to typical adult-adult language. After quantifying the contribution of this "child-directed listening" over developmental time, we discuss the consequences for theories of language acquisition, as well as the implications for commonly-used methods for assessing children's linguistic proficiency.

Platforms that support online commentary, from social networks to news sites, are increasingly leveraging machine learning to assist their moderation efforts. But this process does not typically provide feedback to the author that would help them contribute according to the community guidelines. This is prohibitively time-consuming for human moderators to do, and computational approaches are still nascent. This work focuses on models that can help suggest rephrasings of toxic comments in a more civil manner. Inspired by recent progress in unpaired sequence-to-sequence tasks, a self-supervised learning model is introduced, called CAE-T5. CAE-T5 employs a pre-trained text-to-text transformer, which is fine tuned with a denoising and cyclic auto-encoder loss. Experimenting with the largest toxicity detection dataset to date (Civil Comments) our model generates sentences that are more fluent and better at preserving the initial content compared to earlier text style transfer systems which we compare with using several scoring systems and human evaluation.

Click counts are related to the amount of money that online advertisers paid to news sites. Such business models forced some news sites to employ a dirty trick of click-baiting, i.e., using a hyperbolic and interesting words, sometimes unfinished sentence in a headline to purposefully tease the readers. Some Indonesian online news sites also joined the party of clickbait, which indirectly degrade other established news sites' credibility. A neural network with a pre-trained language model M-BERT that acted as a embedding layer is then combined with a 100 nodes hidden layer and topped with a sigmoid classifier was trained to detect clickbait headlines. With a total of 6632 headlines as a training dataset, the classifier performed remarkably well. Evaluated with 5-fold cross validation, it has an accuracy score of 0.914, f1-score of 0.914, precision score of 0.916, and ROC-AUC of 0.92. The usage of multilingual BERT in Indonesian text classification task was tested and is possible to be enhanced further. Future possibilities, societal impact, and limitations of the clickbait detection are discussed.

Outcome prediction from clinical text can prevent doctors from overlooking possible risks and help hospitals to plan capacities. We simulate patients at admission time, when decision support can be especially valuable, and contribute a novel admission to discharge task with four common outcome prediction targets: Diagnoses at discharge, procedures performed, in-hospital mortality and length-of-stay prediction. The ideal system should infer outcomes based on symptoms, pre-conditions and risk factors of a patient. We evaluate the effectiveness of language models to handle this scenario and propose clinical outcome pre-training to integrate knowledge about patient outcomes from multiple public sources. We further present a simple method to incorporate ICD code hierarchy into the models. We show that our approach improves performance on the outcome tasks against several baselines. A detailed analysis reveals further strengths of the model, including transferability, but also weaknesses such as handling of vital values and inconsistencies in the underlying data.

Urdu is a widely spoken language in South Asia. Though immoderate literature exists for the Urdu language still the data isn't enough to naturally process the language by NLP techniques. Very efficient language models exist for the English language, a high resource language, but Urdu and other under-resourced languages have been neglected for a long time. To create efficient language models for these languages we must have good word embedding models. For Urdu, we can only find word embeddings trained and developed using the skip-gram model. In this paper, we have built a corpus for Urdu by scraping and integrating data from various sources and compiled a vocabulary for the Urdu language. We also modify fasttext embeddings and N-Grams models to enable training them on our built corpus. We have used these trained embeddings for a word similarity task and compared the results with existing techniques.

Automatic essay grading (AEG) is a process in which machines assign a grade to an essay written in response to a topic, called the prompt. Zero-shot AEG is when we train a system to grade essays written to a new prompt which was not present in our training data. In this paper, we describe a solution to the problem of zero-shot automatic essay grading, using cognitive information, in the form of gaze behaviour. Our experiments show that using gaze behaviour helps in improving the performance of AEG systems, especially when we provide a new essay written in response to a new prompt for scoring, by an average of almost 5 percentage points of QWK.

In contrast to their word- or sentence-level counterparts, character embeddings are still poorly understood. We aim at closing this gap with an in-depth study of English character embeddings. For this, we use resources from research on grapheme-color synesthesia -- a neuropsychological phenomenon where letters are associated with colors, which give us insight into which characters are similar for synesthetes and how characters are organized in color space. Comparing 10 different character embeddings, we ask: How similar are character embeddings to a synesthete's perception of characters? And how similar are character embeddings extracted from different models? We find that LSTMs agree with humans more than transformers. Comparing across tasks, grapheme-to-phoneme conversion results in the most human-like character embeddings. Finally, ELMo embeddings differ from both humans and other models.