Transformer-based methods are appealing for multilingual text classification, but common research benchmarks like XNLI (Conneau et al., 2018) do not reflect the data availability and task variety of industry applications. We present an empirical comparison of transformer-based text classification models in a variety of practical monolingual and multilingual pretraining and fine-tuning settings. We evaluate these methods on two distinct tasks in five different languages. Departing from prior work, our results show that multilingual language models can outperform monolingual ones in some downstream tasks and target languages. We additionally show that practical modifications such as task- and domain-adaptive pretraining and data augmentation can improve classification performance without the need for additional labeled data.

The ability to recognize harmful content within online communities has come into focus for researchers, engineers and policy makers seeking to protect users from abuse. While the number of datasets aiming to capture forms of abuse has grown in recent years, the community has not standardized around how various harmful behaviors are defined, creating challenges for reliable moderation, modeling and evaluation. As a step towards attaining shared understanding of how online abuse may be modeled, we synthesize the most common types of abuse described by industry, policy, community and health experts into a unified typology of harmful content, with detailed criteria and exceptions for each type of abuse.

A capable, automatic Question Answering (QA) system can provide more complete and accurate answers using a comprehensive knowledge base (KB). One important approach to constructing a comprehensive knowledge base is to extract information from Wikipedia infobox tables to populate an existing KB. Despite previous successes in the Infobox Extraction (IBE) problem (e.g., DBpedia), three major challenges remain: 1) Deterministic extraction patterns used in DBpedia are vulnerable to template changes; 2) Over-trusting Wikipedia anchor links can lead to entity disambiguation errors; 3) Heuristic-based extraction of unlinkable entities yields low precision, hurting both accuracy and completeness of the final KB. This paper presents a robust approach that tackles all three challenges. We build probabilistic models to predict relations between entity mentions directly from the infobox tables in HTML. The entity mentions are linked to identifiers in an existing KB if possible. The unlinkable ones are also parsed and preserved in the final output. Training data for both the relation extraction and the entity linking models are automatically generated using distant supervision. We demonstrate the empirical effectiveness of the proposed method in both precision and recall compared to a strong IBE baseline, DBpedia, with an absolute improvement of 41.3% in average F1. We also show that our extraction makes the final KB significantly more complete, improving the completeness score of list-value relation types by 61.4%.

We introduce the Scratchpad Mechanism, a novel addition to the sequence-to-sequence (seq2seq) neural network architecture and demonstrate its effectiveness in improving the overall fluency of seq2seq models for natural language generation tasks. By enabling the decoder at each time step to write to all of the encoder output layers, Scratchpad can employ the encoder as a “scratchpad” memory to keep track of what has been generated so far and thereby guide future generation. We evaluate Scratchpad in the context of three well-studied natural language generation tasks — Machine Translation, Question Generation, and Text Summarization — and obtain state-of-the-art or comparable performance on standard datasets for each task. Qualitative assessments in the form of human judgements (question generation), attention visualization (MT), and sample output (summarization) provide further evidence of the ability of Scratchpad to generate fluent and expressive output.