Modeling Personal Biases in Language Use by Inducing Personalized Word Embeddings

Proceedings of NAACL-HLT 2019, pages 2102–2108 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2102 Modeling Personal Biases in Language Use by Inducing Personalized Word Embeddings Daisuke Oba1, Naoki Yoshinaga2, Shoetsu Sato1, Satoshi Akasaki1, Masashi Toyoda2 1The University of Tokyo, Japan 2Institute of Industrial Science, the University of Tokyo, Japan 1{oba, shoetsu, akasaki}@tkl.iis.u-tokyo.ac.jp 2{ynaga, toyoda}@iis.u-tokyo.ac.jp Abstract There exist biases in individual’s language use; the same word (e.g., cool) is used for ex- pressing different meanings (e.g., temperature range) or different words (e.g., cloudy, hazy) are used for describing the same meaning. In this study, we propose a method of modeling such personal biases in word meanings (here- after, semantic variations) with personalized word embeddings obtained by solving a task on subjective text while regarding words used by different individuals as different words. To prevent personalized word embeddings from being contaminated by other irrelevant biases, we solve a task of identifying a review-target (objective output) from a given review. To sta- bilize the training of this extreme multi-class classification, we perform a multi-task learn- ing with metadata identification. Experimental results with reviews retrieved from RateBeer confirmed that the obtained personalized word embeddings improved the accuracy of senti- ment analysis as well as the target task. Anal- ysis of the obtained personalized word embed- dings revealed trends in semantic variations re- lated to frequent and adjective words. 1 Introduction When we verbalize what we have sensed, there exist inevitable personal biases in word meanings (hereafter, (personal) semantic variations). For example, when we say “this pizza is greasy,” how greasy can vary widely among individuals. When we see the same beer, we may use different words (e.g., red, amber) to refer its color. The semantic variations will thereby cause problems not only in communicating with each other, but also in build- ing natural language processing (NLP) systems. Several studies have attempted to personalize models to improve the performance on NLP tasks such as sentiment analysis (Gao et al., 2013) and dialogue systems (Li et al., 2016; Zhang et al., 2018). All of these studies, however, tried to esti- mate subjective output from subjective input (e.g., estimating sentiment scores given by reviewers). These personalized models are thereby affected by not only semantic variations in subjective in- put but also annotation bias (deviation of outputs given by the annotators) and selection bias (devi- ation of outputs caused by the deviation of input) (§ 2). This makes it difficult to understand the pure impact of the personal semantic variations. In this study, aiming at understanding seman- tic variations and their impact on NLP tasks, we propose a method for modeling personal seman- tic variations with personalized word embeddings obtained through the review-target identification task. This task estimates the review-target (objec- tive output) from a given review (subjective input) (§ 3), and is free from annotation bias since output labels are given a priori. Also, selection bias can be suppressed by using a dataset in which the same reviewer evaluates the same …

Content truncated to 500 words