DeTEXT: Programming by Example

Data cleaning and manipulation is an expensive and tedious process that consumes valuable time and resources. If an end-user needs to manipulate a large amount of data, he has three primary options: record a macro, write a transformation program, or perform manual manipulations on the data. Recording a macro with a fixed series of actions can be simple, but does not generalize well and is not sufficiently expressive for many tasks. Writing a transformation program requires expert knowledge unavailable to many end-users. And depending on the quantity of data, it may not even be possible to perform manual manipulations. Programming by demonstration (PbD) offers an intuitive alternative to these approaches. In a PbD setting, an end-user demonstrates the actions required to transform input examples into output examples. A machine learning algorithm observes these transformation sequences, called traces, with the goal of inducing a program that transforms all future examples in the desired manner. A line of successful research on PbD uses a formalism called version space algebra [7] to represent the space of all hypotheses (programs) that are consistent with the input-output examples. The version space algebra defines useful operators like union and join that enable construction of complex version spaces from simple ones. SmartEDIT [5] is a PbD system for learning text transformations. However, it is not always possible to observe traces. Observing traces requires tight integration with text editors, which come in many forms and often do not expose their source code for PbD integration. Furthermore, input-output examples may have been pre-recorded, as in a partially completed spreadsheet. Spreadsheet users, who often need to clean or manipulate data, number over 500 million people worldwide; there is a huge opportunity for time savings and increased productivity. The problem of inductively learning programs without traces is called programming by example (PbE). PbE is a strictly harder problem than PbD; the learning algorithm observes less information and thus must consider a much larger space of possible hypotheses. Our research has focused on lifting the PbD methods used by SmartEDIT to the PbE context. This work was conducted independently of [1], which also aims to solve this problem. In this work, we present DeTEXT, a PbE system that is able to save users time in a number of repetitive text-editing scenarios. While the rich language of text transformations we originally envisioned is beyond the scope of this project, DeTEXT learns simple editing operations from three or fewer examples. We make the following contributions: