A5103159534- Hate Speech and Cyberbullying Detection
- Topic Modeling
- Logic, programming, and type systems
- Wikis in Education and Collaboration
- Natural Language Processing Techniques
- Software Engineering Research
- Cancer-related gene regulation
- Logic, Reasoning, and Knowledge
- Advanced Database Systems and Queries
- Recommender Systems and Techniques
- Formal Methods in Verification
- Computability, Logic, AI Algorithms
- Explainable Artificial Intelligence (XAI)
- Multimodal Machine Learning Applications
- Parallel Computing and Optimization Techniques
- Adversarial Robustness in Machine Learning
- Software Testing and Debugging Techniques
- Advanced Malware Detection Techniques
- Mathematics, Computing, and Information Processing
- Sentiment Analysis and Opinion Mining
- Artificial Intelligence in Games
- Text Readability and Simplification
- Ethics and Social Impacts of AI
- Spam and Phishing Detection
- Distributed and Parallel Computing Systems
Google (United States)
2010-2024
University of Bologna
2023
Athens University of Economics and Business
2019-2023
Télécom Paris
2021
Stockholm University
2021
University of Oxford
2020
University of South Carolina
2020
Cornell University
2018
University of Edinburgh
2003-2018
Wikimedia Foundation
2018
We describe novel computational techniques for constructing induction rules deductive synthesis proofs. Deductive holds out the promise of automated construction correct computer programs from specifications their desired behaviour. Synthesis with iteration or recursion requires inductive proof, but standard appropriate are restricted to recycling recursive structure specifications. What is needed rule that can introduce structures. show a combination rippling and use meta-variables as...
We introduce and illustrate a new approach to measuring mitigating unintended bias in machine learning models. Our definition of is parameterized by test set subset input features. how this can be used evaluate text classifiers using synthetic public corpus comments annotated for toxicity from Wikipedia Talk pages. also demonstrate imbalances training data lead the resulting models, therefore potentially unfair applications. use common demographic identity terms as features on which we...
The damage personal attacks cause to online discourse motivates many platforms try curb the phenomenon. However, understanding prevalence and impact of in at scale remains surprisingly difficult. contribution this paper is develop illustrate a method that combines crowdsourcing machine learning analyze scale. We show an evaluation for classifier terms aggregated number crowd-workers it can approximate. apply our methodology English Wikipedia, generating corpus over 100k high quality...
Unintended bias in Machine Learning can manifest as systemic differences performance for different demographic groups, potentially compounding existing challenges to fairness society at large. In this paper, we introduce a suite of threshold-agnostic metrics that provide nuanced view unintended bias, by considering the various ways classifier's score distribution vary across designated groups. We also large new test set online comments with crowd-sourced annotations identity references. use...
The damage personal attacks cause to online discourse motivates many platforms try curb the phenomenon. However, understanding prevalence and impact of in at scale remains surprisingly difficult. contribution this paper is develop illustrate a method that combines crowdsourcing machine learning analyze scale. We show an evaluation for classifier terms aggregated number crowd-workers it can approximate. apply our methodology English Wikipedia, generating corpus over 100k high quality...
Justine Zhang, Jonathan Chang, Cristian Danescu-Niculescu-Mizil, Lucas Dixon, Yiqing Hua, Dario Taraborelli, Nithum Thain. Proceedings of the 56th Annual Meeting Association for Computational Linguistics (Volume 1: Long Papers). 2018.
This work introduces Gemma, a family of lightweight, state-of-the art open models built from the research and technology used to create Gemini models. Gemma demonstrate strong performance across academic benchmarks for language understanding, reasoning, safety. We release two sizes (2 billion 7 parameters), provide both pretrained fine-tuned checkpoints. outperforms similarly sized on 11 out 18 text-based tasks, we present comprehensive evaluations safety responsibility aspects models,...
Traditional recommender systems leverage users' item preference history to recommend novel content that users may like. However, modern dialog interfaces allow express language-based preferences offer a fundamentally different modality for input. Inspired by recent successes of prompting paradigms large language models (LLMs), we study their use making recommendations from both item-based and in comparison state-of-the-art collaborative filtering (CF) methods. To support this investigation,...
Scaling language models with more data, compute and parameters has driven significant progress in natural processing. For example, thanks to scaling, GPT-3 was able achieve strong results on in-context learning tasks. However, training these large dense requires amounts of computing resources. In this paper, we propose develop a family named GLaM (Generalist Language Model), which uses sparsely activated mixture-of-experts architecture scale the model capacity while also incurring...
Moderation is crucial to promoting healthy online discussions. Although several 'toxicity' detection datasets and models have been published, most of them ignore the context posts, implicitly assuming that comments may be judged independently. We investigate this assumption by focusing on two questions: (a) does affect human judgement, (b) conditioning improve performance toxicity systems? experiment with Wikipedia conversations, limiting notion previous post in thread discussion title. find...
Discussing things you care about can be difficult, especially via online platforms, where sharing your opinion leaves open to the real and immediate threats of abuse harassment. Due these threats, people stop expressing themselves give up on seeking different opinions. Recent research efforts focus examining strengths weaknesses (e.g. potential unintended biases) using machine learning as a support tool facilitate safe space for discussions; example, through detecting various types negative...
This paper presents the application of two strong baseline systems for toxicity detection and evaluates their performance in identifying categorizing offensive language social media. PERSPECTIVE is an API, that serves multiple machine learning models improvement conversations online, as well a system, trained on wide variety comments from platforms across Internet. BERT recently popular representation model, fine tuned per task achieving state art NLP tasks. performed better than detecting...
One of the main challenges online social systems face is prevalence antisocial behavior, such as harassment and personal attacks. In this work, we introduce task predicting from very start a conversation whether it will get out hand. As opposed to detecting undesirable behavior after fact, aims enable early, actionable prediction at time when might still be salvaged. To end, develop framework for capturing pragmatic devices---such politeness strategies rhetorical prompts---used conversation,...
Platforms that support online commentary, from social networks to news sites, are increasingly leveraging machine learning assist their moderation efforts. But this process does not typically provide feedback the author would help them contribute according community guidelines. This is prohibitively time-consuming for human moderators do, and computational approaches still nascent. work focuses on models can suggest rephrasings of toxic comments in a more civil manner. Inspired by recent...
We introduce the Constructive Comments Corpus (C3), comprised of 12,000 annotated news comments, intended to help build new tools for online communities improve quality their discussions. define constructive comments as high-quality that make a contribution conversation. explain crowd worker annotation scheme and de ne taxonomy subcharacteristics constructiveness. The resulting dataset is evaluated using measurements inter-annotator agreement, expert assessment sample, by constructiveness...
String diagrams are a powerful tool for reasoning about physical processes, logic circuits, tensor networks and many other compositional structures. The distinguishing feature of these is that edges need not be connected to vertices at both ends, unconnected ends can interpreted as the inputs outputs diagram. In this paper, we give concrete construction string using special kind typed graph called an open-graph. While category open-graphs itself adhesive, introduce notion selective adhesive...
Yiqing Hua, Cristian Danescu-Niculescu-Mizil, Dario Taraborelli, Nithum Thain, Jeffery Sorensen, Lucas Dixon. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.
Automatic side-by-side evaluation has emerged as a promising approach to evaluating the quality of responses from large language models (LLMs). However, analyzing results this raises scalability and interpretability challenges. In paper, we present LLM Comparator, novel visual analytics tool for interactively automatic evaluation. The supports interactive workflows users understand when why model performs better or worse than baseline model, how two are qualitatively different. We...
Ilan Price, Jordan Gifford-Moore, Jory Flemming, Saul Musker, Maayan Roichman, Guillaume Sylvain, Nithum Thain, Lucas Dixon, Jeffrey Sorensen. Proceedings of the Fourth Workshop on Online Abuse and Harms. 2020.