Creating a model of likely word tags