Acquiring Broad Commonsense Knowledge for Sentiment Analysis Using Human Computation
While artificial intelligence is successful in many applications that cover specific domains, for many commonsense problems there is still a large gap with human performance. Automated sentiment analysis is a typical example: while there are techniques that reasonably aggregate sentiments from texts in specific domains, such as online reviews of a particular product category, more general models have a poor performance. We argue that sentiment analysis can be covered more broadly by extending models with commonsense knowledge acquired at scale, using human computation. We study two sentiment analysis problems. We start with document-level sentiment classification, which aims to determine whether a text as a whole expresses a positive or a negative sentiment. We hypothesize that extending classifiers to include the polarities of sentiment words in context can help them scale to broad domains. We also study fine-grained opinion extraction, which aims to pinpoint individual opinions in a text, along with their targets. We hypothesize that extraction models can benefit from broad fine-grained annotations to boost their performance on unfamiliar domains. Selecting sentiment words in context and annotating texts with opinions and targets are tasks that require commonsense knowledge shared by all the speakers of a language. We show how these can be effectively solved through human computation. We illustrate how to define small tasks that can be solved by many independent workers so that results can form a single coherent knowledge base. We also show how to recruit, train, and engage workers, then how to perform effective quality control to obtain sufficiently high-quality knowledge. We show how the resulting knowledge can be effectively integrated into models that scale to broad domains and also perform well in unfamiliar domains. We engage workers through both enjoyment and payment, by designing our tasks as games played for money. We recruit them on a paid crowdsourcing platform where we can reach out to a large pool of active workers. This is an effective recipe for acquiring sentiment knowledge in English, a language that is known by the vast majority of workers on the platform. To acquire sentiment knowledge for other languages, which have received comparatively little attention, we argue that we need to design tasks that appeal to voluntary workers outside the crowdsourcing platform, based on enjoyment alone. However, recruiting and engaging volunteers has been more of an art than a problem that can be solved systematically. We show that combining online advertisement with games, an approach that has been recently proved to work well for acquiring expert knowledge, gives an effective recipe for luring and engaging volunteers to provide good quality sentiment knowledge for texts in French. Our solutions could point the way to how to use human computation to broaden the competence of artificial intelligence systems in other domains as well.
Programme doctoral Informatique et Communications
Faculté informatique et communications
Laboratoire d'intelligence artificielle
Jury: Dr Denis Gillet (président) ; Prof. Boi Faltings, Dr Pearl Pu Faltings (directeurs) ; Prof. Daniel Gatica-Perez, Prof. Bing Liu, Prof. Philippe Cudré-Mauroux (rapporteurs)
Public defense: 2016-10-7
Record created on 2016-09-21, modified on 2016-12-12