Final Report Summary - ESSENCE (Evolution of Shared Semantics in Computational Environments)
Despite impressive recent advances in Artificial Intelligence (AI), we are still far away from developing intelligent technologies that are capable of combining the intelligence of different AI systems when these exhibit the kind of diversity that we observe among humans in terms of knowledge, skills, viewpoints, and motivations. And yet, we know that it is precisely this diversity that has enabled human societies to collectively accomplish their most impressive achievements, most of which require a level of intelligence that goes far beyond that of any single individual.
This raises a question that is of fundamental importance to the future advancement of AI: How can we make artificially intelligent systems diversity-aware, i.e. capable of establishing the differences and similarities between themselves and other systems (or humans), and enable them to take these differences and similarities into account when interacting with each other? Over the past four years, the ambitious vision of ESSENCE has been to bring together key disciplines that investigate the fundamental research problems that underpin this question, and to develop an integrated research training programme that will allow future generations of researchers to break new ground in developing truly diversity-aware AI.
Operating in a previously highly fragmented research landscape, the fifteen individual projects carried out within ESSENCE by pre- and post-doctoral researchers have made significant progress on various problems related to this broader vision, such as: the development of domain-independent dialogue planning algorithms; new theories for how agreement over meaning of concepts is established in human communication; simulation models that explain how compositional linguistic structures emerge through repeated interactions; methods for reconciling incompatible conceptualisations of a domain that different communicating agents may hold; and frameworks for crowdsourcing mappings between different interpretations of linguistically expressed concepts from human users.
While all of these efforts were driven by the aim to replicate the capabilities humans exhibit when negotiating meaning in AI systems, each of them applied methods from very different research areas, such as theoretical linguistics, multiagent systems, argumentation, semantic technologies, dialogue systems, and language evolution. Through extensive collaboration between different groups working in these areas, we identified diversity-awareness as the unifying characteristic of the methods we seek to bring together. A key result has been the insight that a holistic conceptual framework for diversity-awareness has the potential to bring about a step change in AI by enabling the re-use and re-combination of single-purpose systems in complex ecosystems of heterogeneous, collaborating AI systems and people. Though realising this ambitious vision is not something that can be fully accomplished by a single research project, the contribution of ESSENCE has been to break new ground in terms of understanding the relationships between different methodological approaches to diversity-awareness and taking significant steps toward their integration.
An important achievement of ESSENCE research has been to capture the full lifecycle of reasoning about diversity across different dimensions: Along the human-to-machine dimension, this covers a range of different settings, from autonomous agents creating new concepts based on objects and relations they encounter in their environments, to humans refining the conceptual differences between similar words in different languages. Along the generation-to-analysis dimension, it ranges from generating the right utterances to achieve a given communicative goal to measuring the diversity among different definitions of concepts by analysing how these concepts are used in practice. Along the semantics-to-pragmatics dimension, it can involve a range of approaches from dynamically learning which strategies of concept use are able to achieve meaningfully aligned communication to explicitly exchanging arguments among heterogeneous agents trying to agree on a shared definition of a concept. While the approaches developed by ESSENCE researchers do not aim to cover the full space of all possible solutions in this multi-dimensional space, each of them provides a piece in the greater puzzle, while at the same time making important scientific contributions to one or several relevant disciplines.
This scientific impact of ESSENCE is evidenced by 70 publications authored by the researchers trained in the network, collections of 15 software components and 9 datasets that have been made available online, seven international research schools and workshops, and a final conference that we have organised and which, taken together, have attracted almost 200 external participants. Through these activities, over 150 talks and presentations given by members of the network, and our presence as exhibitors in three of the largest international AI conferences that took place over the last four years, we have managed to put diversity-awareness “on the map” in the European AI landscape, and fostered the emergence of a new community in the area. Crucial to this has also been the development of a challenge scenario around commonsense reasoning in human-machine dialogue, which we showcased through an open competition at the largest international AI conference at the end of the project, and promoted through a publicly available mobile app we developed.
A major impact of ESSENCE on research training has been its innovative approach to interdisciplinary doctoral and post-doctoral training in an emerging thematic area for which no previous training programmes existed. This approach has been based on combining a broad range of training capabilities and resources across different countries, sectors, and disciplines, and enabling early-career researchers to make use of this ecosystem in ways that are tailored to the needs of their individual research project, both. Collaboration among researchers and development of a shared vision were fostered through several types of joint activities, including workshops, informal research meetings, secondments, training courses, research visits, and “coding camps”, which provided ample opportunities for ESSENCE members to spend time working closely with each other. The integration projects that involved gathering software components and benchmark datasets as well as developing a joint challenge scenario served as further points of collaboration to ensure overall coherence. In terms of impact on gender equality in AI research, ESSENCE has pursued a successful equal opportunities strategy, recruiting 33% female Fellows (a very high percentage for the subject area), appointing a dedicated equality officer who supported female Fellows throughout the project, and inviting many female speakers at our events. At the end of the project, we have developed a deep understanding of how a flexible and lightweight, yet coherent and sustainable training programme can be co-produced through collaboration between experienced scientists and early-career researchers, and have designed a curriculum, training materials, and a joint PhD degree programme structure that can be used to continue training in the thematic area of ESSENCE in the future.
Given the fundamental nature of the research conducted within the ESSENCE network, we expect the long-term socio-economic impact of the network to emerge from future follow-up projects where the researchers we have trained will employ cross-disciplinary foundations gained from their experience in the network to develop next-generation diversity-aware AI systems. First directly exploitable results are already being fielded for future commercial use at various partner institutions, such as a crowdsourcing-based multilingual lexical resource platform, a knowledge integration support system for organisations using heterogeneous ontologies, and a linked data framework for human-machine collaboration in web application development.
More information about ESSENCE is available from the network’s official website, www.essence-network.com. ESSENCE can be contacted directly by e-mail to essence-info@inf.ed.ac.uk.
This raises a question that is of fundamental importance to the future advancement of AI: How can we make artificially intelligent systems diversity-aware, i.e. capable of establishing the differences and similarities between themselves and other systems (or humans), and enable them to take these differences and similarities into account when interacting with each other? Over the past four years, the ambitious vision of ESSENCE has been to bring together key disciplines that investigate the fundamental research problems that underpin this question, and to develop an integrated research training programme that will allow future generations of researchers to break new ground in developing truly diversity-aware AI.
Operating in a previously highly fragmented research landscape, the fifteen individual projects carried out within ESSENCE by pre- and post-doctoral researchers have made significant progress on various problems related to this broader vision, such as: the development of domain-independent dialogue planning algorithms; new theories for how agreement over meaning of concepts is established in human communication; simulation models that explain how compositional linguistic structures emerge through repeated interactions; methods for reconciling incompatible conceptualisations of a domain that different communicating agents may hold; and frameworks for crowdsourcing mappings between different interpretations of linguistically expressed concepts from human users.
While all of these efforts were driven by the aim to replicate the capabilities humans exhibit when negotiating meaning in AI systems, each of them applied methods from very different research areas, such as theoretical linguistics, multiagent systems, argumentation, semantic technologies, dialogue systems, and language evolution. Through extensive collaboration between different groups working in these areas, we identified diversity-awareness as the unifying characteristic of the methods we seek to bring together. A key result has been the insight that a holistic conceptual framework for diversity-awareness has the potential to bring about a step change in AI by enabling the re-use and re-combination of single-purpose systems in complex ecosystems of heterogeneous, collaborating AI systems and people. Though realising this ambitious vision is not something that can be fully accomplished by a single research project, the contribution of ESSENCE has been to break new ground in terms of understanding the relationships between different methodological approaches to diversity-awareness and taking significant steps toward their integration.
An important achievement of ESSENCE research has been to capture the full lifecycle of reasoning about diversity across different dimensions: Along the human-to-machine dimension, this covers a range of different settings, from autonomous agents creating new concepts based on objects and relations they encounter in their environments, to humans refining the conceptual differences between similar words in different languages. Along the generation-to-analysis dimension, it ranges from generating the right utterances to achieve a given communicative goal to measuring the diversity among different definitions of concepts by analysing how these concepts are used in practice. Along the semantics-to-pragmatics dimension, it can involve a range of approaches from dynamically learning which strategies of concept use are able to achieve meaningfully aligned communication to explicitly exchanging arguments among heterogeneous agents trying to agree on a shared definition of a concept. While the approaches developed by ESSENCE researchers do not aim to cover the full space of all possible solutions in this multi-dimensional space, each of them provides a piece in the greater puzzle, while at the same time making important scientific contributions to one or several relevant disciplines.
This scientific impact of ESSENCE is evidenced by 70 publications authored by the researchers trained in the network, collections of 15 software components and 9 datasets that have been made available online, seven international research schools and workshops, and a final conference that we have organised and which, taken together, have attracted almost 200 external participants. Through these activities, over 150 talks and presentations given by members of the network, and our presence as exhibitors in three of the largest international AI conferences that took place over the last four years, we have managed to put diversity-awareness “on the map” in the European AI landscape, and fostered the emergence of a new community in the area. Crucial to this has also been the development of a challenge scenario around commonsense reasoning in human-machine dialogue, which we showcased through an open competition at the largest international AI conference at the end of the project, and promoted through a publicly available mobile app we developed.
A major impact of ESSENCE on research training has been its innovative approach to interdisciplinary doctoral and post-doctoral training in an emerging thematic area for which no previous training programmes existed. This approach has been based on combining a broad range of training capabilities and resources across different countries, sectors, and disciplines, and enabling early-career researchers to make use of this ecosystem in ways that are tailored to the needs of their individual research project, both. Collaboration among researchers and development of a shared vision were fostered through several types of joint activities, including workshops, informal research meetings, secondments, training courses, research visits, and “coding camps”, which provided ample opportunities for ESSENCE members to spend time working closely with each other. The integration projects that involved gathering software components and benchmark datasets as well as developing a joint challenge scenario served as further points of collaboration to ensure overall coherence. In terms of impact on gender equality in AI research, ESSENCE has pursued a successful equal opportunities strategy, recruiting 33% female Fellows (a very high percentage for the subject area), appointing a dedicated equality officer who supported female Fellows throughout the project, and inviting many female speakers at our events. At the end of the project, we have developed a deep understanding of how a flexible and lightweight, yet coherent and sustainable training programme can be co-produced through collaboration between experienced scientists and early-career researchers, and have designed a curriculum, training materials, and a joint PhD degree programme structure that can be used to continue training in the thematic area of ESSENCE in the future.
Given the fundamental nature of the research conducted within the ESSENCE network, we expect the long-term socio-economic impact of the network to emerge from future follow-up projects where the researchers we have trained will employ cross-disciplinary foundations gained from their experience in the network to develop next-generation diversity-aware AI systems. First directly exploitable results are already being fielded for future commercial use at various partner institutions, such as a crowdsourcing-based multilingual lexical resource platform, a knowledge integration support system for organisations using heterogeneous ontologies, and a linked data framework for human-machine collaboration in web application development.
More information about ESSENCE is available from the network’s official website, www.essence-network.com. ESSENCE can be contacted directly by e-mail to essence-info@inf.ed.ac.uk.