Uncovering Distributional Differences between Synonyms and Antonyms in a Word Space Model

International Joint Conference on Natural Language Processing, pages 489–497, Nagoya, Japan, 14-18 October 2013. Uncovering Distributional Differences between Synonyms and Antonyms in a Word Space Model Silke Scheible, Sabine Schulte im Walde and Sylvia Springorum Institut f¨ur Maschinelle Sprachverarbeitung Universit¨at Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {scheible,schulte,riestesa}@ims.uni-stuttgart.de Abstract For many NLP applications such as In- formation Extraction and Sentiment De- tection, it is of vital importance to distin- guish between synonyms and antonyms. While the general assumption is that dis- tributional models are not suitable for this task, we demonstrate that using suitable features, differences in the contexts of syn- onymous and antonymous German adjec- tive pairs can be identified with a simple word space model. Experimenting with two context settings (a simple window- based model and a ‘co-disambiguation model’ to approximate adjective sense disambiguation), our best model signif- icantly outperforms the 50% baseline and achieves 70.6% accuracy in a syn- onym/antonym classification task. 1 Introduction One notorious problem of distributional similarity models is that they tend to not only retrieve words that are strongly alike to each other (such as syn- onyms), but also words that differ in their meaning (i.e. antonyms). It has often been argued that this behaviour is due to the distributional similarity of synonyms and antonyms: despite conveying dif- ferent meanings, antonyms also seem to occur in very similar contexts (Mohammad et al., 2013). In many applications, such as information re- trieval and machine translation, the presence of antonyms can be devastating (Lin et al., 2003). While a number of approaches have addressed the issue of synonym and antonym distinction from a computational point of view, they are usually lim- ited in some way, for example by requiring the antonymous words to co-occur in certain patterns (Lin et al., 2003; Turney, 2008), or by relying on external resources such as thesauri (Mohammad et al., 2013; Yih et al., 2012). Probably due to this strong similarity of their contexts, there have been no successful attempts so far to distinguish the two relations via a standard distributional model such as the word space model (Sahlgren, 2006). Prominent work in psycholinguistics, however, has shown that humans are able to distinguish the contexts of antonymous words, and that these are by no means interchangeable (Charles and Miller, 1991). The goal of our research is to show that using suitable features these differences can be identified via a simple word space model, rely- ing on contextual clues that govern the ability to distinguish the relations in context. For this pur- pose, we present a word space model that ex- ploits window-based features for synonymous and antonymous German adjective pairs. Next to in- vestigating the contributions of the various parts- of-speech with regard to the word space model, we experiment with two context settings: one that takes into account all contexts in which the mem- bers of the word pairs occur, and one where we approximate context disambiguation by applying ‘co-disambiguation’: establishing the set of nouns that are modified by both …

Content truncated to 500 words