Commonsense for Machine Intelligence

Audience:

  • Beginners are introduced to commonsense knowledge mining.
  • Developers and designers can learn how to apply commonsense in various applications.
  • Researchers and practitioners can understand the state-of-the-art techniques in multi-modal commonsense mining and open problems in this increasingly relevant area.

The tutorial will be half a day long and covers the following topics.

Tutorial outline:

Part 1: Acquiring commonsense knowledge

Introduction to commonsense knowledge mining

Types of Knowledge

We introduce encyclopedic, lexical, and commonsense knowledge and compare them. We present recent advances in knowledge bases, discussing to what extend these provide commonsense.

Broad overview of commonsense acquisition models

We discuss the spectrum of models, including curated, distantly supervised and unsupervised models for mining commonsense, such as WordNet, Cyc, ConceptNet, WebChild, UWN.

Techniques

Representation learning

We discuss frame-based, graph-based, and, continuous representation based representations of commonsense knowledge.

Commonsense fact acquisition

We discuss models to obtain more advanced commonsense facts, both from text and from video, and other multimodal Web data.

Entailment graph mining

In addition to individual facts, we discuss models to mine entailments for commonsense reasoning.

Part 2: Detecting and correcting odd collocations in text

Introduction to collocation error correction

Collocations and odd collocations

We introduce collocated expressions and odd collocations that are collocation errors where expressions are found that are not typically used in correct communication, e.g., mighty tea instead of strong tea.

Broad overview of techniques to fix collocation errors

We discuss interesting tasks where collocation error correction is useful and common techniques to fix these .

Treatment of Collocation Errors

Linguistic classification

We discuss current research in the area of classifying collocation errors which leads to their correction, mainly from a linguistic classification perspective.

Detection and correction

We discuss frequency-based, semantic similarity, ranking and ensemble learning based techniques.

Part 3: Applications and open issues

Applications

Smart Cities

We describe the primary directions in smart cities research and highlight how commonsense has been useful currently, and envision potential application of commonsense in smart cities.

Intelligent Personal Assistants and robots

We then describe current applications of commonsense to robotics, such as in the Stanford’s Robo Brain project and discuss how commonsense computing is coming to consumers with Amazon Echo and JIBO robot. We discuss how research in collocations would positively impact machine translation and enable smarter communication.

Caption Generation

We discuss the applications of commonsense knowledge in computer vision.

Open research areas

Open problems in commonsense knowledge

We discuss some open problems in commonsense knowledge acquisition and deploying

Open problems in collocations

We discuss some open areas in research on collocations in text, e.g., domain knowledge, literary allusions and sparse data.