Discovering influential text using convolutional neural networks

Anonymous

TL;DR: We design and interpret a model with convolutional structures to flexibly discover influential text treatments, and apply it to a corpus of censored Weibo posts as well as a collection of complaints to the Consumer Financial Protection Bureau.

Abstract: Experimental methods for estimating the impacts of text on human evaluation have been widely used in the social sciences. However, researchers in experimental settings are usually limited to testing a small numbers of pre-specified text treatments. While efforts to mine unstructured texts for features that causally affect outcomes have been ongoing in recent years, these models have primarily focused on the topics or specific words of text, which may not always be the mechanism of the effect. In this paper, we extend these efforts and present a flexible model utilizing convolutional neural networks for discovering clusters of similar phrases in text that are predictive of human reactions to those texts. When used in an experimental setting, this method can identify candidate text treatments and effects under certain assumptions. We apply our model to two data sets. The first concerns censorship of social media posts and enables direct validation of our model. The second investigates complaints to the Consumer Financial Protection Bureau, and demonstrates the model's ability to flexibly discover text treatments with varying textual structures.

Paper Type: long

Research Area: Computational Social Science and Cultural Analytics

Contribution Types: Model analysis & interpretability

Languages Studied: English, Chinese