We present LaMDA: Language Models for Dialog Applications. LaMDA is a family of Transformerbased neural language models specialized for dialog, which have up to 137B parameters and are pre-trained on 1.56T words of public dialog data and web text. While model scaling alone can improve quality, it shows less improvements on safety and factual grounding. We demonstrate that fine-tuning with annotated data and enabling the model to consult external knowledge sources can lead to significant improvements towards the two key challenges of safety and factual grounding. The first challenge, safety, involves ensuring that the model's responses are consistent with a set of human values, such as preventing harmful suggestions and unfair bias. We quantify safety using a metric based on an illustrative set of human values, and we find that filtering candidate responses using a LaMDA classifier fine-tuned with a small amount of crowdworker-annotated data offers a promising approach to improving model safety. The second challenge, factual grounding, involves enabling the model to consult external knowledge sources, such as an information retrieval system, a language translator, and a calculator. We quantify factuality using a groundedness metric, and we find that our approach enables the model to generate responses grounded in known sources, rather than responses that merely sound plausible. Finally, we explore the use of LaMDA in the domains of education and content recommendations, and analyze their helpfulness and role consistency. * Work done while at Google.
The increasing availability of (digital) cultural heritage artefacts offers great potential for increased access to art content, but also necessitates tools to help users deal with such abundance of information. User-adaptive art recommender systems aim to present their users with art content tailored to their interests. These systems try to adapt to the user based on feedback from the user on which artworks he or she finds interesting. Users need to be able to depend on the system to competently adapt to their feedback and find the artworks that are most interesting to them. This paper investigates the influence of transparency on user trust in and acceptance of content-based recommender systems. A between-subject experiment (N = 60) evaluated interaction with three versions of a content-based art recommender in the cultural heritage domain. This recommender system provides users with artworks that are of interest to them, based on their ratings of other artworks. Version 1 was not transparent, version 2 explained to the user why a recommendation had been made and version 3 showed a rating of how certain the system was that a recommendation would be of interest to the user. Results show that explaining to the user why a recommendation was made increased acceptance of the recommendations. Trust in the system itself was not improved by transparency. Showing how certain the system was of a recommendation did not influence trust and acceptance. A number of guidelines for design of recommender systems in the cultural heritage domain have been derived from the study's results.
Galleries, Libraries, Archives and Museums (short: GLAMs) around the globe are beginning to explore the potential of crowdsourcing, i.e. outsourcing specific activities to a community though an open call. In this paper, we propose a typology of these activities, based on an empirical study of a substantial amount of projects initiated by relevant cultural heritage institutions. We use the Digital Content Life Cycle model to study the relation between the different types of crowdsourcing and the core activities of heritage organizations. Finally, we focus on two critical challenges that will define the success of these collaborations between amateurs and professionals: (1) finding sufficient knowledgeable, and loyal users; (2) maintaining a reasonable level of quality. We thus show the path towards a more open, connected and smart cultural heritage: open (the data is open, shared and accessible), connected (the use of linked data allows for interoperable infrastructures, with users and providers getting more and more connected), and smart (the use of knowledge and web technologies allows us to provide interesting data to the right users, in the right context, anytime, anywhere -both with involved users/consumers and providers). It leads to a future cultural heritage that is open, has intelligent infrastructures and has involved users, consumers and providers.
Big data is having a disruptive impact across the sciences. Human annotation of semantic interpretation tasks is a critical part of big data semantics, but it is based on an antiquated ideal of a single correct truth that needs to be similarly disrupted. We expose seven myths about human annotation, most of which derive from that antiquated ideal of truth, and dispell these myths with examples from our research. We propose a new theory of truth, crowd truth, that is based on the intuition that human interpretation is subjective, and that measuring annotations on the same objects of interpretation (in our examples, sentences) across a crowd will provide a useful representation of their subjectivity and the range of reasonable interpretations.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.