WordGuess: Using Associations for Guessing, Learning and Exploring Related Words

WORDGUESS Using Associations for Guessing, Learning and Exploring Related Words Cennet O˘guz1, Andr´e Blessing2, Jonas Kuhn2, Sabine Schulte im Walde2 1German Research Center for Artificial Intelligence (DFKI), Saarland Informatics Campus 2Institute for Natural Language Processing, University of Stuttgart ceog01@dfki.de, {blessiae,jonas,schulte}@ims.uni-stuttgart.de Abstract This paper presents WORDGUESS, a game- with-a-purpose vocabulary training where –in order to guess a target word (such as snow)– the player is offered associations of that tar- get word (such as winter, white, cold). The game relies on existing association norms and co-occurrence information to establish an en- tertaining way of deepening the player’s learn- ing and understanding of vocabulary and of associative relatedness between words in the vocabulary. WORDGUESS comes with data in English and German and can be extended with data from further languages. From an application-oriented point of view the play- ers’ data enables us to induce conditions and weights for word association and to quantify contextual relationships, which is useful for many NLP purposes such as ontology induc- tion and anaphora resolution. 1 Introduction Games-with-a-purpose (GWAP) offer enjoyable entertainment to players and at the same time al- low researchers in Natural Language Processing (NLP) to collect and explore cognitive and (com- putational) linguistic facets of human-generated data. While the term GWAP has been coined by Von Ahn and Dabbish (2008), the underlying idea has been pursued across linguistic levels and across NLP purposes for much longer. To provide a few examples across research fields, the adventure and interactive fiction games by Gabsdil et al. (2002) rely on natural-language question-answering dia- logues to explore inference systems with reference resolution, syntactic ambiguities, and scripted dia- logues; Chamberlain et al. (2008) exploit collabo- rative work to identify relationships between words and phrases in web data; OntoGame (Siorpaes and Hepp, 2008) matches classes in an ontology with Wikipedia articles; Hladk´a et al. (2009) propose a gamified annotation approach for coreference reso- lution; Guillaume et al. (2016) design ZombiLingo, a game for syntactic dependency annotation. We present a GWAP-style game implementation called WORDGUESS1 where associations of a tar- get word are offered to players in order to guess the target word. For example, associations such as win- ter, white, cold provide hints to players when guess- ing the target word snow. Our game is a web-based and mobile-based application whose aim is to learn and understand word-association and word-context relationships: previous research has shown that associations and corpus co-occurrence are related (Church and Hanks, 1990; de Deyne and Storms, 2008a; Schulte im Walde et al., 2008, i.a.); we plan to explore their connections and differences in more depth. In this vein, (i) we vary associations obtained from humans, and context-based words induced from corpus co-occurrence; (ii) we pro- vide a multilingual gaming environment in order to understand the conditions across languages and relational patterns between native and second lan- guages; and (iii) we offer the players to choose between levels of difficulty (i.e., providing more or less cues). The obtained data enables us to induce conditions and weights for …

Content truncated to 500 words