Ordering adverbs by their scaling effect on adjective intensity

Proceedings of Recent Advances in Natural Language Processing, pages 545–554, Hissar, Bulgaria, Sep 7–9 2015. Ordering adverbs by their scaling effect on adjective intensity Josef Ruppenhofer∗, Jasper Brandes∗, Petra Steiner∗, Michael Wiegand† ∗Hildesheim University Hildesheim, Germany {ruppenho|brandesj|steinerp}@uni-hildesheim.de †Saarland University Saarbr¨ucken, Germany michael.wiegand@lsv.uni-saarland.de Abstract In recent years, theoretical and computa- tional linguistics has paid much attention to linguistic items that form scales. In NLP, much research has focused on order- ing adjectives by intensity (tiny < small). Here, we address the task of automatically ordering English adverbs by their intensi- fying or diminishing effect on adjectives (e.g. extremely small < very small). We experiment with 4 different methods: 1) using the association strength be- tween adverbs and adjectives; 2) exploit- ing scalar patterns (such as not only X but Y); 3) using the metadata of product re- views; 4) clustering. The method that per- forms best is based on the use of metadata and ranks adverbs by their scaling factor relative to unmodified adjectives. 1 Introduction Being able to recognize the intensity associ- ated with scalar expressions is a basic capability needed for tackling any NLP task that can be re- duced to textual entailment. For instance, as il- lustrated by de Marneffe et al. (2010), when inter- preting dialogue (A: Was it good? B: It was ok / great / excellent.), a yes/no question involving a gradable predicate may require understanding the entailment relations between that predicate and another contained in the answer. Another applica- tion is within sentiment analysis, where assessing the strength of subjective expressions (e.g. good < great < excellent) is one of the central tasks besides subjectivity detection and polarity classi- fication (Rill et al., 2012b; Sheinman et al., 2013; de Melo and Bansal, 2013; Ruppenhofer et al., 2014, inter alia). It is also well known that sub- jective adjectives are frequently modified by ad- verbs that increase (very expensive) or decrease (fairly expensive) their intensity. As Benamara et al. (2007) have shown, it is useful to take such adverbial intensification into account when pre- dicting document-level sentiment scores. How- ever, Benamara et al. (2007) used human-assigned scores to model adverbs’ effect on adjectives. As far as we know, there is no well-established automatic method that can determine for degree adverbs what their effect will be on the intensity of various adjectives. In this paper, we explore sev- eral methods on English data that might be used towards that purpose, evaluating them against a new gold standard data set that we collected. All new resources that were created in the context of our investigation will be made publicly available. The remainder of this paper is structured as fol- lows. We present our data in §2. We describe the construction of our gold standard in §3 and the methods we use in §4. This is followed by the presentation of our experiments and results in §5. We discuss related work in §6 and conclude in §7. 2 Data For our experiments we use two large corpora (Ta- …

Content truncated to 500 words