Agents Learning Lying And Negotiation
Nearly every business transaction and every diplomatic agreement from haggling over a turnip at a farmers market to a multi-trillion dollar trade deal is a negotiation. These diplomatic arts are not just a series of cold calculations to optimize an objective function. They are built on cultural norms and emotional ties learned from human teachers and a history of past negotiations. If we want computers to help human negotiators and to hold their own against humans, the game of Diplomacy is an ideal testbed for teaching computer agents how to learn to negotiate. We will build agents that negotiate and cooperate in Diplomacy, a game of politics and war that relies on negotiated agreements between players that must ultimately be fluid, with lies and betrayal used to manipulate other players into temporarily behaving in helpful ways. To learn to play this game, which is notable for being beloved among political operatives like John F. Kennedy and Henry Kissinger, we will incorporate deep reinforcement learning, sophisticated strategy and tactics engines, and a dialogue model fine tuned to opportunistically lie. This work is a collaboration between the University of Maryland (Prime), Princeton University, and the University of Sydney.
Cross-Lingual Event and Argument Retrieval
The automatic extraction of events from text has empowered tasks as varied as the prediction of political stability forecasting or the automatic creation of in-depth biomedical information resources. However, most training data for event extraction models is available only in English. Under IARPA’s BETTER program, ISI researchers are developing an innovative end-to-end, cross-lingual system which will provide personalized, multi-lingual semantic extraction and retrieval from text in foreign languages, using only English training data. Our collaborator on this effort is UMass Amherst. This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via 2019-19051600007.
Combined Representations for Adept Learning
Modern applications of machine learning (ML) constitute transformative technologies, whose adoption is spreading rapidly and impacting myriad sectors of the economy. However, ML’s revolutionary potential is limited by its cost—both the cost of creating corpora of labeled training data and the difficulty of adapting successful models to perform well in related domains. The goals of CORAL are to create revolutionary machine learning (ML) technology that will achieve state-of-the-art (SOTA) performance on a broad range of tasks, but require a factor of 106 less training data than the SOTA to automatically create a new capability for a new domain and task, and require 106 less training data to automatically adapt to a related domain. We additionally will develop and test information-theoretic techniques that characterize the limits of ML and adaptation. Our collaborators are the California Institute of Technology, Carnegie Mellon University, Columbia University, the NVIDIA Corporation, and the University of Illinois, Urbana-Champaign. This work is based in part on research sponsored by Air Force Research Laboratory (AFRL) under agreement number FA8750-19-1-1000.
Dialogue Agent for Reducing Malicious Acts
Social media has transformed how people share information and exchange ideas; however, unfettered communication has also unleashed toxic, anti-social behaviors, such as harrassment, disagreement, personal attacks, malicious rhetoric and other toxic communication acts. Without mediation, these anti-social behaviors create discord within a community and impede its ability to collaborate, share information, create consensus and build trust. We will build a multilingual, adaptable bots that mediate online dialogue so as to limit communication breakdown due to toxic behaviors and facilitate civil discourse. DARMA bots will monitor online conversations looking for indicators of toxic behaviors such as hate speech, trolling, and polarization. Our bots will determine the appropriate time to intervene, and will do so using principles of world-building and shared mental state learned from improvisational theater corpora and conflict resolution dialogues. DARMA will be able to operate in up to 500 different world languages (which cover more than 90% of primary languages of current world population) via our massively multilingual machine translation systems, which can be adapted to new domains using only monolingual adaptation data. This work is carried out with funding from DARPA (HR0011-22-9-0025).
Exploiting Language Information for Situational Awareness
Today's automatic parsers, translators, extractors, and dictionaries cover a tiny fraction of the world's languages. Can we use general knowledge of how language works to extend the reach of natural language tools? In this project, we develop technology for rapidly constructing information extraction (IE), machine translation (MT), and topic and sentiment processing capabilities for new languages. Our collaborators are ICSI, Brno University of Technology, University of Pennsylvania, University of Notre Dame, Rensselaer Polytechnic Institute, and Next Century, Inc. This work is carried out with funding from DARPA (HR0011-15-C-0115).
Evidence Extraction Systems for the Molecular Interaction Literature
Biomedical databases describe the claims made by scientists in detail, but rarely provide descriptions of any supporting evidence that a consulting scientist could use to understand why the claims are being made. Currently, the process of curating evidence into databases is manual, time-consuming and expensive; thus, the evidence is recorded in papers but not generally captured in database systems. Although experimental evidence is complex, it conforms to certain principles of experimental design. Exploiting these principles has permitted us to devise a preliminary, robust, general-purpose representation for experimental evidence. A major goal of our project is that we will develop methods to extract this evidence from scientiﬁc papers automatically (1) by using natural language processing to read information from the text used by scientists to describe their results and (2) by using image processing on a speciﬁc subtype of ﬁgure that is common in molecular biology papers. We will develop these tools and package them so that they may then also be used for evidence pertaining to other areas of research in biomedicine. Funding statement: This work is carried out with funding from an NIH R01 (LM012592).
Learning Event Schema Temporally and Transmodally
Understanding a complex event requires organizing its typical sequence of actions and the pattern of duration and participation in these actions. By organizing events in this manner, downstream analytics (or analysts) are able to identify the early warning signs of a process and recognize important but missing information. While handcrafted models provide this capability for well-understood processes, a system that learns schemas from data offers the opportunity to extend event-focused analytics to new domains. We are developing the LESTAT system, which semi-automatically discovers schemas for complex events (CE). These automatically discovered schemas will be general, composable, and specializable to support domain-specific contexts. LESTAT’s schemas will be the basis for systems developed elsewhere to track and anticipate the activities of entities of interest based on detecting individually weak event signals in multi-modal, multilingual input. Our collaborators are Arizona State University and the University of Central Florida. This research is based upon work supported by DARPA's KAIROS program, Contract FA8750-19-2-0500.
Machine Intelligence from Common Sense
Our goal is to automatically learn a common sense repository (CSR) of knowledge that can be applied to diverse problems. In this project we learn common sense from textual and visual sources, as well as from rich pre-existing common sense knowledge bases (e.g. ConceptNet, YAGO) and observations of relations between objects and concepts as they appear in visual and linguistic sources (i.e. images, videos, documents). The resulting CSR will be a dense, low-dimensional space that efficiently and accurately encapsulates hundreds of millions of observations. To enable more robust reasoning and question answering over this space than the state of the art, we will develop novel inference techniques for both transitive and multi-hop reasoning over the space (e.g. bats have wings, wings enable flying, thus, bats can fly). We apply our system to shared tasks and via an interactive demonstration system which supports direct questioning (e.g. Can the Grand Canyon fly to New York?). Our collaborators are Columbia University, UCLA, and UMass-Amherst. This research is based upon work supported by DARPA's MCS program, Contract N660011924032.
Stylometric Authorship Discernment & Interpretation for Realistic Inputs
The goal of SADIRI is to develop effective, explainable, and multi-lingual technology for authorship attribution and privacy protection. We are developing techniques to generate a linguistically-informed feature space that is robust to topic and genre shift, allowing us to generate an “authorial fingerprint” vector for each document (or set of documents) that the system reads. This fingerprint is a neural representation of the author’s unique “style”, distinct from any content present in the original documents. We use these fingerprints directly to perform authorship attribution tasks (“who wrote this tweet?”) and probe them automatically to generate meaningful, linguistic explanations for system decisions. Working on the flip side of the attribution problem, we are also developing techniques to “translate” documents from one style to another (replacing the author’s original “fingerprint”), allowing us to protect author privacy while preserving meaning and fluency. Our collaborators are the University of Michigan, the University of Maryland, and the University of Birmingham. This research is based upon work supported by IARPA's HIATUS program, contract #2022-22072200006.
Summarization and Domain-Adaptive Retrieval Across Languages
How can a monolingual English speaker access and understand text and speech material in low-resource foreign languages? In this project, we develop cross-lingual retrieval and summarization techniques that will work for any language in the world, given minimal resources to work with. Our collaborators are University of Notre Dame, Rensselaer Polytechnic Institute, Massachusetts Institute of Technology, Northeastern University, University of Massachusetts, and Idiap. This work is supported by the IARPA MATERIAL program. (Acknowledgement: This research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via AFRL Contract FA8650-17-C-9116.