Grant: Discovery Grants Program - Individual
McGill University
$34,000.00
Robust conversational models for accessing the world's knowledge
FY 2023-24
Summary
Human language is an incredible tool for communication. My research goal is to enable communicating with computers in natural language for accessing all knowledge. I envision a future where specialized knowledge such as law, science and medicine is accessible to everyone in colloquial language irrespective of their education, where autonomous systems such as robots respond to language, and where mundane tasks of customer care agencies are automated by bots. In the last decade, we have seen a significant change in the way we access knowledge. Search engines have progressed from showing a list of URLs for keyword queries to directly answering well-formed questions. Smart assistants like Siri and Alexa have become a common trend where the main mode of interaction is language. While this progress is commendable, the limitations of their language capability are apparent within a few interactions. The main goal of the proposed research program is to build robust conversational models that allow accessing the world's knowledge (e.g., Wikipedia or Wikidata). We focus on the following objectives of robustness: Interpretability: Interpretability helps us to understand the strengths and weaknesses of current conversational models. It also allows users to decide whether to trust a model's prediction. We call a model interpretable if it exhibits general linguistic intelligence (e.g. resolving pronouns). We use latent-variable models like reinforcement learning to interpret existing conversational models. Transferability: A good conversational model requires various linguistic skills. Such skills are hard to learn from one single task. Here we build models that share knowledge across diverse multiple tasks or transfer knowledge from one conversational task to the other. We explore transfer learning techniques like pretraining, knowledge-distillation and multi-task learning to improve transferability. Expressivity: Current algorithms make assumptions that undermine or exaggerate the expressivity of models, leading to poor generalization or overfitting. We develop novel methods that incorporate inductive priors suitable for conversational modeling. Moreover, current interactions with AI assistants feel very robotic. We build models that actively learn from interactions to mimic-human like behavior. We are advancing into an era where language will be a key mode of interaction with machines. This research program is critical in making these interactions trustworthy, reliable and human-like. At a scientific level, this program combines cutting-edge machine learning techniques with linguistic insights. This combination could lead to novel machine learning models for language that are formally sound and practically robust. Moreover, this program will train highly qualified personnel in natural language processing, machine learning and linguistics. Such trainees are highly sought after in academia and industry.
Program
Discovery Grants Program - Individual
Recipient
Location
Quebec, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
RGPIN-2020-05885
Name nom
Reddy, Siva
Department département
Computer Science and Linguistics
Organization id
61
Institution établissement
McGill University
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2020
Award amount
34000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
1207
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Information, computer and communication technologies
Area of application fr
Technologie de l'information, des ordinateurs et des réseaux de communications
Research subject code
9999
Research subject group en
Not available
Research subject group fr
Not available
Research subject en
Non disponible
Research subject fr
Non disponible
Application title
Robust conversational models for accessing the world's knowledge
Keyword
Computational Linguistics; Conversational Systems; Deep Learning; Dialogue Systems; Machine Reading; Natural Language Processing; Question Answering; Representation Learning; Semantic Parsing; Transfer Learning
Application summary
Human language is an incredible tool for communication. My research goal is to enable communicating with computers in natural language for accessing all knowledge. I envision a future where specialized knowledge such as law, science and medicine is accessible to everyone in colloquial language irrespective of their education, where autonomous systems such as robots respond to language, and where mundane tasks of customer care agencies are automated by bots. In the last decade, we have seen a significant change in the way we access knowledge. Search engines have progressed from showing a list of URLs for keyword queries to directly answering well-formed questions. Smart assistants like Siri and Alexa have become a common trend where the main mode of interaction is language. While this progress is commendable, the limitations of their language capability are apparent within a few interactions. The main goal of the proposed research program is to build robust conversational models that allow accessing the world's knowledge (e.g., Wikipedia or Wikidata). We focus on the following objectives of robustness: Interpretability: Interpretability helps us to understand the strengths and weaknesses of current conversational models. It also allows users to decide whether to trust a model's prediction. We call a model interpretable if it exhibits general linguistic intelligence (e.g. resolving pronouns). We use latent-variable models like reinforcement learning to interpret existing conversational models. Transferability: A good conversational model requires various linguistic skills. Such skills are hard to learn from one single task. Here we build models that share knowledge across diverse multiple tasks or transfer knowledge from one conversational task to the other. We explore transfer learning techniques like pretraining, knowledge-distillation and multi-task learning to improve transferability. Expressivity: Current algorithms make assumptions that undermine or exaggerate the expressivity of models, leading to poor generalization or overfitting. We develop novel methods that incorporate inductive priors suitable for conversational modeling. Moreover, current interactions with AI assistants feel very robotic. We build models that actively learn from interactions to mimic-human like behavior. We are advancing into an era where language will be a key mode of interaction with machines. This research program is critical in making these interactions trustworthy, reliable and human-like. At a scientific level, this program combines cutting-edge machine learning techniques with linguistic insights. This combination could lead to novel machine learning models for language that are formally sound and practically robust. Moreover, this program will train highly qualified personnel in natural language processing, machine learning and linguistics. Such trainees are highly sought after in academia and industry.