Improving social relationships in face-to-face human-agent interactions: when the agent wants to know user’s likes and dislikes

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1064–1073, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Improving social relationships in face-to-face human-agent interactions: when the agent wants to know user’s likes and dislikes Caroline Langlet Institut Mines-T´el´ecom T´el´ecom ParisTech CNRS LTCI caroline.langlet@ telecom-paristech.fr Chlo´e Clavel Institut Mines-T´el´ecom T´el´ecom ParisTech CNRS LTCI chloe.clavel @telecom-paristech.fr Abstract This paper tackles the issue of the detec- tion of user’s verbal expressions of likes and dislikes in a human-agent interaction. We present a system grounded on the theo- retical framework provided by (Martin and White, 2005) that integrates the interac- tion context by jointly processing agent’s and user’s utterances. It is designed as a rule-based and bottom-up process based on a symbolic representation of the struc- ture of the sentence. This article also describes the annotation campaign – car- ried out through Amazon Mechanical Turk – for the creation of the evaluation data- set. Finally, we present all measures for rating agreement between our system and the human reference and obtain agreement scores that are equal or higher than sub- stantial agreements. 1 Introduction In the research field of the embodied conversa- tional agents (ECA), detecting sentiment-related phenomena 1 appears as a key task to improve human-agent interactions and to build long-term social relationships (Pecune et al., 2013). Sev- eral models and applications have been proposed which mostly take into account non-verbal cues (acoustic features, facial or bodily expressions) to determine the user’s emotions (Schuller et al., 2011). The verbal content is more and more inte- grated but still partially exploited in human-agent interactions. The very infrequent works, integrat- ing the detection of user’s sentiments in ECAs 1The term sentiment-related phenomena is used in (Clavel et al., 2013) to regroup all the phenomena related to sentiment in the literature, from opinion to affect and emotion. based on linguistic cues, concern avatars and vi- sualisation issues rather than face-to-face interac- tion, (Zhang et al., 2008; Neviarouskaya et al., 2010b). We identify so far two studies that in- tegrate a sentiment detection module for human- agent interaction (Smith et al., 2011; Yildirim et al., 2011). However, the research field of sentiment anal- ysis and opinion mining provides a set of inter- esting works dealing with the subjective informa- tion conveyed by the verbal content. Three types of approaches are considered: machine-learning, rule-based approaches and hybrid approaches that are a combination of the first two types. Machine learning methods have proven their worth for the positive and negative classification of sentences or texts (Pang and Lee, 2008). Rule-based ap- proaches are grounded on syntactic and semantic analyses of the sentence and provide deeper anal- yses of sentiment-related phenomena. For exam- ple, (Neviarouskaya et al., 2010a) and (Moilanen and Pulman, 2007) provide linguistic rules deal- ing with the principle of compositionality in or- der to improve the detection of opinion targets and the resolution of polarity. Similarly, (Shaikh et al., 2009) …

Content truncated to 500 words