Teaching Machines to Classify from Natural Language Interactions

2019-01-16T19:32:30Z (GMT) by Shashank Srivastava
Humans routinely learn new concepts using natural language communications,<br>even in scenarios with limited or no labeled examples. For example, a<br>human can learn the concept of a phishing email from natural language explanations<br>such as ‘phishing emails often request your bank account number’.<br>On the other hand, purely inductive learning systems typically require a large<br>collection of labeled data for learning such a concept. We believe that advances<br>in Computational Linguistics and the growing ubiquity of computing devices<br>together can enable people to teach computers classification tasks using natural<br>language interactions.<br>Learning from language presents some key challenges. A preliminary challenge<br>lies in the basic problem of learning to interpret language, which refers<br>to an agent’s ability to map natural language explanations in pedagogical contexts<br>to formal semantic representations that computers can process and reason<br>over. A second challenge is that of learning from interpretations, which refers<br>to the mechanisms through which interpretations of language statements can<br>be used by computers to solve learning tasks in the environment. We address<br>aspects of both these problems, and provide an interface for guiding concept<br>learning methods using language.<br>For learning from interpretation, we focus on concept learning (binary classification)<br>tasks. We demonstrate that language can define rich and expressive<br>features for learning tasks, and show that machine learning can benefit substantially<br>from this ability. We also investigate assimilation of linguistic cues<br>in everyday language that implicitly provide constraints for classification models<br>(e.g., ‘Most emails are not phishing emails’). In particular, we focus on<br>conditional statements and linguistic quantifiers (such as usually, never, etc.),<br>and show that such advice can be used to train classifiers even with few or no<br>labeled examples of a concept.<br>For learning to interpret, we develop new algorithms for semantic parsing<br>that incorporate pragmatic cues, including conversational history and sensory<br>observation, to improve automatic language interpretation. We show that environmental<br>context can enrich semantic parsing methods by not only providing<br>discriminative features, but also reducing the need for expensive labeled data<br>used for training them.<br>A separate but immensely valuable attribute of human language is that<br>it is inherently conversational and interactive. We also briefly explore the<br>possibility of agents that can learn to interact with a human teacher in a mixedinitiative<br>setting, where the learner can also proactively engage the teacher by<br>asking questions, rather than only passively listen. We develop a reinforce<br>learning framework for learning effective question asking strategies in context<br>of conversational concept learning.