Learning the Peculiar Value of Actions

Proceedings of the Third Joint Conference on Lexical and Computational Semantics (*SEM 2014), pages 63–68, Dublin, Ireland, August 23-24 2014. Learning the Peculiar Value of Actions Daniel Dahlmeier Research & Innovation, SAP Asia, Singapore d.dahlmeier@sap.com Abstract We consider the task of automatically es- timating the value of human actions. We cast the problem as a supervised learning- to-rank problem between pairs of action descriptions. We present a large, novel data set for this task which consists of challenges from the I Will If You Will Earth Hour challenge. We show that an SVM ranking model with simple linguistic features can accurately predict the relative value of actions. 1 Introduction The question on how humans conceptualize value is of great interest to researchers in various fields, including linguistics (Jackendoff, 2006). The link between value and language arises from the fact that we cannot directly observe value due to its ab- stract nature and instead often study language ex- pressions that describe actions which have some value attached to them. This creates an interesting link between the semantics of the words that de- scribe the actions and the underlying moral value of the actions. Jackendoff (2006) describes value as an “inter- nal accounting system” for ethical decision pro- cesses that exhibits both valence (good or bad) and magnitude (better or worse). Most interest- ingly, value is governed by a “peculiar logic” that provides constraints on which actions are deemed morally acceptable and which are not. In par- ticular, the principal of reciprocity states that the valence and magnitude of reciprocal actions (ac- tions that are done “in return” for something else) should match, i.e., positive valued actions should This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http: //creativecommons.org/licenses/by/4.0/ match with positive valued reciprocal actions (re- actions) of similar magnitude, and conversely neg- atively valued actions should match with nega- tive valued reciprocal actions (reactions) of similar magnitude. In this paper, we consider the task of automati- cally estimating the value of actions. We present a simple and effective method for learning the value of actions from ranked pairs of textual action de- scriptions based on a statistical learning-to-rank approach. Our experiments are based on a novel data set that we create from challenges submit- ted to the I Will if You Will Earth Hour challenge where participants pledge to do something daring or challenging if other people commit to sustain- able actions for the planet. Our method achieves a surprisingly high accuracy of up to 94.72% in a 10-fold cross-validation experiment. The results show that the value of actions can accurately be estimated by machine learning methods based on lexical descriptions of the actions. The main contribution of this paper is that we show how the semantics of value in language can accurately be learned from empirical data using a learning-to-rank approach. Our work shows an in- teresting link between empirical research on se- mantics in …

Content truncated to 500 words