Task 1.c - Verb Categorization

The goal of the sub-task is to group verbs into semantic categories.

The data set consists of 45 verbs, belonging to 9 semantic classes. The classification scheme is inspired to the one described in P. Vinson & G. Vigliocco (2007), “Semantic Feature Production Norms for a Large Set of Objects and Events”, Behavior Research Methods, which in turn closely follows the classification proposed in Levin (1993).

We operationalize verb categorization as a clustering task. Since the data set is organized hierarchically, we will run two clustering experiments, varying the number of classes and consequently their level of generality:

• 9-way clustering - models will be tested on their ability to categorize the verbs into the most fine-grained classes of the dataset: communication ("talk"), mentalState ("know"), motionManner ("run"), motionDirection ("arrive"), changeLocation ("carry"), bodySense ("smell"), bodyAction ("eat"), exchange ("buy"), changeState ("destroy");
• 5-way clustering - models will be tested on their ability to categorize the verbs into 5 classes: cognition (superordinate of communication and mentalState), motion (superordinate of motionManner, motionDirection, changeLocation), body (superordinate of bodySense and bodyAction), exchange, and changeState;

To abstract away from differences stemming from any specific clustering method, you are asked to run your experiments with the k-means algorithm available in CLUTO. In case you can not run CLUTO on your system, the workshop organizers will carry out the clustering for you. In this case, data should be prepared according to a format that will be specified later on. Participants are also invited to experiment with other clustering methods and to compare the results with those obtained withCLUTO.

Evaluation will be carried in two stages:

1. quantitative evaluation - results will be evaluated with respect to the two measures for cluster quality available in CLUTO: purity and entropy (cf. Zhao, Y. and G. Karypis (2002), "Evaluation of Hierarchical Clustering Algorithms for Document Datasets", in CIKM 2002).

2. qualitative evaluation - verb semantic classification is notoriously hard, and any apriori classification runs the risk of being sunsatisfactorily. This is essenailly due to the high degree of polysmet tyoical of verbs as well as to theit intrinisci multimdnsonal caharatcers. Thsi second stage of of ervaluatiow ill therefore focus on specifc verbs selected as "hard cases". Participants will be asked to perform a fine-grained error analysis onnthioer systen behavor with thes eselected verbs. Details about this type of evaluation will be provided later on.

Back to Start