Teaching Machines to Classify from Natural Language Interactions

Humans routinely learn new concepts using natural language communications,even in scenarios with limited or no labeled examples. For example, ahuman can learn the concept of a phishing email from natural language explanationssuch as ‘phishing emails often request your bank account number’.On the other hand, purely inductive learning systems typically require a largecollection of labeled data for learning such a concept. We believe that advancesin Computational Linguistics and the growing ubiquity of computing devicestogether can enable people to teach computers classification tasks using naturallanguage interactions.Learning from language presents some key challenges. A preliminary challengelies in the basic problem of learning to interpret language, which refersto an agent’s ability to map natural language explanations in pedagogical contextsto formal semantic representations that computers can process and reasonover. A second challenge is that of learning from interpretations, which refersto the mechanisms through which interpretations of language statements canbe used by computers to solve learning tasks in the environment. We addressaspects of both these problems, and provide an interface for guiding conceptlearning methods using language.For learning from interpretation, we focus on concept learning (binary classification)tasks. We demonstrate that language can define rich and expressivefeatures for learning tasks, and show that machine learning can benefit substantiallyfrom this ability. We also investigate assimilation of linguistic cuesin everyday language that implicitly provide constraints for classification models(e.g., ‘Most emails are not phishing emails’). In particular, we focus onconditional statements and linguistic quantifiers (such as usually, never, etc.),and show that such advice can be used to train classifiers even with few or nolabeled examples of a concept.For learning to interpret, we develop new algorithms for semantic parsingthat incorporate pragmatic cues, including conversational history and sensoryobservation, to improve automatic language interpretation. We show that environmentalcontext can enrich semantic parsing methods by not only providingdiscriminative features, but also reducing the need for expensive labeled dataused for training them.A separate but immensely valuable attribute of human language is thatit is inherently conversational and interactive. We also briefly explore thepossibility of agents that can learn to interact with a human teacher in a mixedinitiativesetting, where the learner can also proactively engage the teacher byasking questions, rather than only passively listen. We develop a reinforcelearning framework for learning effective question asking strategies in contextof conversational concept learning