Safety for E2E Conversational AI

Overview

Recent progress in end-to-end (E2E) open-domain dialogue agents has shown a dramatic improvement in the ability of these models to learn patterns and mimic conversational behaviors from the training data. Unfortunately these models still exhibit many concerning problems. First, they may learn undesirable features present in the training data such as biased, toxic, or otherwise harmful language. Second, they do not respond reliably in safety critical situations, such as the user asking for medical advice, indicating intention of self-harm, or during public health crises such as COVID-19.

These problems are difficult, as what is deemed as “offensive” or even “sensitive” is both contextually and culturally dependent, and picking up on more subtle examples of unsafe language often requires a level of language understanding that is well beyond current capabilities.

A first workshop was held to discuss these topics on October 2020. Following on from that, this special session will take place at SIGDial 2021.

Location, format, and schedule

SafeConvAI Special Session @ SIGDial 2021 will be an all-virtual event.

The special session will consist of a keynote, a panel, lightning talks, oral and poster paper presentations.

Schedule

July 29th

14:00 UTC: Welcome by the organisers (Verena Rieser)

14:10 UTC: Keynote: Laurence Devillers (session chair: Saadia Gabriel)

14:50 UTC: Coffee break

15:00 UTC: Paper presentations (session chair: Gavin Abercrombie)

Assessing Political Prudence of Open-domain Chatbots

Authors: Yejin Bang, Nayeon Lee, Etsuko Ishii, Andrea Madotto and Pascale Fung

Large-Scale Quantitative Evaluation of Dialogue Agents' Response Strategies against Offensive Users

Authors: Haojun Li, Dilara Soylu and Christopher Manning

15:30 UTC: panel discussion (session chair: Emily Dinan)

16:15 UTC end

Panel

• Pascale Fung is a professor in the Department of Electronic & Computer Engineering and the Department of Computer Science & Engineering at the Hong Kong University of Science & Technology.

• Pilar Manchon is a Senior Director of Research Strategy at Google AI.

• Ehud Reiter is a Professor of Computing Science, University of Aberdeen and Chief Scientist at Arria NLG.

• Michelle Zhou is a Co-founder and CEO of Juji, Inc., a California-based company that powers Cognitive Artificial Intelligence (AI) Assistants in the form of chatbots.

Objectives

Our objectives for this special session are:

1. To showcase recent and ongoing work related to building safer conversational AI systems

2. To bridge the gap between dialogue and neighboring disciplines, such as safety security, computational social science, policy, ethics, applications such as health, etc.

3. To identify paths forward for more responsible conversational AI research

Call for papers

We invite regular and work-in-progress papers that report:

• Current research on identifying and/or preventing dialogue models from generating messages containing: hate speech, pornographic or sexual content, offensive or profane language, medical or self-harm advice, privacy leaks, or otherwise harmful behaviour.

• Work on understanding how social biases are embedded in conversational AI and conversational agents, and approaches to mitigate the harms of such biases,

• Qualitative or quantitative investigations on strategies for how models should best respond to various types of unsafe content,

• Relevant work in related fields, including HRI, virtual agents, etc.

• New methods and metrics for evaluating the relative safety of conversational AI models,

• New (dynamic) benchmarks and datasets to investigate these issues,

• Dataset releases that address current biases or imbalances in data used to train/evaluate dialogue systems,

• Conversational abuse detection and mitigation across multiple languages, dialects, and sociolects; spoken and typed interactions etc., this includes abusive, derogatory, or otherwise harmful speech, including but not limited to: bullying, microaggressive behavior, toxic language, hate speech and offensive language.

• Reliably and robustly detecting safety critical situations,

• Related issues including privacy and privacy leaks, persona-design, fairness, accessibility and inclusion for Conversational AI systems,

• Best practices for researching and releasing E2E conversational AI,

• Position papers to reflect on the current state of the art in this topic, take stock of where we have been, where we are, where we are going and where we should go.

Submission Format

• Long papers must describe substantial, original, completed, and unpublished work. Wherever appropriate, concrete evaluation and analysis should be included. These papers will undergo the same peer-review process by the SIGDial program committee as papers submitted to the main SIGdial track. These papers will appear in the main SIGdial proceedings and are presented with the main track. Long papers must be no longer than 8 pages, including title, text, figures, and tables. An unlimited number of pages is allowed for references. Two additional pages are allowed for appendices containing sample discourses/dialogues and algorithms, and an extra page is allowed in the final version to address reviewers’ comments.

• Short papers must describe original and unpublished work. These papers would go through the same peer-review process by the SIGDial program committee as papers submitted to the main SIGdial track. These papers will appear in the main SIGdial proceedings and are presented with the main track. Please note that a short paper is not a shortened long paper. Instead, short papers should have a point that can be made in a few pages, such as a small, focused contribution; a negative result; or an interesting application nugget. It should be no longer than 4 pages, including title, text, figures, and tables. An unlimited number of pages is allowed for references. One additional page is allowed for sample discourses/dialogues and algorithms, and an extra page is allowed in the final version to address reviewers’ comments. An unlimited number of pages are allowed for references.

• Late-breaking and Work-in-progress abstracts will showcase ongoing work and focused relevant contributions. Submissions need not present original work. Late-breaking and work-in-progress abstracts should be no longer than 1 page. The submissions will be reviewed by the SafeConvAI program committee and posted on the special session website. Abstracts will be presented as lightning talks and/or posters during the session. Authors will retain the copyright to their work so that they may submit to other venues as their work matures.

Important Dates

• Regular Papers (Long and Short) Deadline (SIGDial): April 10, 2021 (23:59 GMT-08)

• Notification for Regular Papers: TBA (as per SIGDial 2021)

• Camera-ready due (Regular Papers): TBA (as per SIGDial 2021)

• Late-breaking (LBR) and Work-In-Progress (WiP) abstracts Deadline: May 31, 2021

• Notification for LBR/WiP abstracts: July 5, 2021

• Event date: July 29

Submission Instructions

• Regular Papers (Long and Short): please follow the SIGDial 2021 submission instructions.

• Late-breaking and Work-In-Progress abstracts: please submit to safeconvai@macs.hw.ac.uk

• Please follow SIGDial 2021 submission instructions and formatting guidelines.

Organizing committee

• Gavin Abercrombie, Heriot-Watt University

• Y-Lan Boureau, Facebook AI

• Emily Dinan, Facebook AI

• Saadia Gabriel, University of Washington

• Alborz Geramifard, Facebook AI

• Verena Rieser, Heriot-Watt University

• Yulia Tsvetkov, Carnegie Mellon University

• Zeerak Waseem, University of Sheffield

Senior advisory board

• Barbara Grosz, Harvard University

• Pilar Manchón, Google

• Ehud Reiter, University of Aberdeen

• Jason Williams, Apple

Contact

Contact the organisers at safety4convai@googlegroups.com