Definite Description Lexical Choice: taking Speaker’s Personality into account

Definite Description Lexical Choice: taking Speaker’s Personality into account Alex Gwo Jen Lan, Ivandr´e Paraboni University of S˜ao Paulo, School of Arts, Sciences and Humanities S˜ao Paulo, Brazil alex.lan@usp.br, ivandre@usp.br Abstract In Natural Language Generation (NLG), Referring Expression Generation (REG) lexical choice is the subtask that provides words to express a given input meaning representation. Since lexical choices made in real language use tend to vary greatly across speakers, computational models of lexicalisation have long addressed the issue of human variation in the REG field as well. However, studies of this kind will often rely on large collections of pre-recorded linguistic examples produced by every single speaker of interest, and on every domain under consideration, to obtain meaning-to-text mappings from which the lexicalisation model is built. As a result, speaker-dependent lexicalisation may be impractical when suitable annotated corpora are not available. As an alternative to corpus-based approaches of this kind, this paper argues that differences across human speakers may be at least partially influenced by personality, and presents a personality-dependent lexical choice model for REG that is, to the best of our knowledge, the first of its kind. Preliminary results show that our personality-dependent approach outperforms a standard lexicalisation model (i.e., based on meaning-to-text mappings alone), and that the use of personality information may be a viable alternative to strategies that rely on corpus knowledge. Keywords: Referring Expressions, Lexical choice, Personality, Big Five 1. Introduction In Natural Language Generation (NLG), lexical choice is understood as the task of selecting words to express an input meaning representation. This paper focuses on the particular subtask of definite descriptions lexical choice, that is, the generation step that follows Referring Expres- sion Generation (REG) content selection (Krahmer and van Deemter, 2012) in a traditional NLG architecture (Reiter and Dale, 2000). The input to the lexicalisation task is a set of meanings (or properties) represented as attribute-value pairs to be ex- pressed in surface form, and the output is a word string. For simplicity, in what follows we shall focus on the choice of words that realise input properties, and we will leave aside issues of linearisation, agreement and others. Let us consider the goal of producing a lexicalisation for a possible description of the person illustrated in Figure 1, taken from Face Place1 images (Righi et al., 2012). Figure 1: An example image from Face Place. In a context of this kind, an underlying REG algorithm may produce a description as in {<gender-male>,<hair.style-curly>}. 1Stimulus images courtesy of Michael J. Tarr, Center for the Neural Basis of Cognition and Department of Psychology, Carnegie Mellon Univ. Funding provided by NSF award 0339122. The task of the lexicalisation model in this case is to as- sign words to these properties, which in the present exam- ple may result in a word string, as in ‘the man with curly hair’. Existing approaches to definite description lexicalisation will often generate a single, fixed surface realisation from the given input. Human descriptions, on the other hand, show much greater linguistic …

Content truncated to 500 words