This page in German/Diese Seite auf Deutsch

The Sentence Fairy - Die Satzfee

A main prerequisite of our system is that the children should not type but only use the mouse, as their writing skills vary too much (in a corpus studies with 1000 little stories, we found out-of-vocabulary rates of 30% due to spelling problems; moreover, keyboard typing skills shouldn't be an issue at all).

The The Sentence Fairy system consists of two main parts: (1) Teacher mode to encode new stories and (2) Pupil mode to run essay-writing exercises.

The goal of the teacher mode is to specify exercises for a new story with minimal effort. The exercises for the pupils are constructed fully automatically from a syntactic/semantic encoding of the story. As teachers cannot be expected to encode the syntactic/semantic representation directly, in our system, those representations are extracted by parsing plain text and asking for some elaboration and clarification in a dialogue mode. Thus, basically, an ordered series of pictures and the corresponding text in terms of simple main clauses have to be entered by the teacher. Moreover, (s)he answers the questions asked by the system and, finally, (s)he determines a limited set of rhetorical structure relations (RST) between pairs of sentences. Currently, we only use the temporal and causal relations. The given simple main clauses are parsed rudimentarily syntactically and semantically by a Performance Grammar parser (the PG structures are expanded by semantic features in the spirit of Minimal Recursion Semantics (MRS). Missing vocabulary is added in a dialogue to fill morpho-syntactic and semantic features. The resulting syntactic/semantic structures guide another dialogue to verify, expand or revise the constructed representation of all the sentences to be used in the Sentence Fairy system.

From this representation of a story, three different exercises for the pupil mode can be generated fully automatically in the current prototype: (1) A comprehension test for the story is performed by asking to associate sentences with pictures. (2) The second exercise is a sentence combining task. From the representation, pairs of sentences with an RST relation are selected and all syntactic realizations for all RST relations that are encoded in the Sentence Fairy's internal knowledge base make up the list of selectable items (e.g., the relation "cause" can be realized by denn, da, weil (for, because). All conjunctions are associated with their syntactic realizations in terms of main and subordinate clause realization (coordinations and relative clause constructions can also be constructed fully automatically by our system). These syntactic representations are sent to our natural language generation component, which can produce all paraphrases licenced by German word order rules. The Sentence Fairy yields all word orders for all possible constructions the pupils could enter. Based on this list, the system generates its binary feedback. Moreover, in exercise (2), anaphors have to be selected by the pupil in order to avoid repetitions. (3) Word order variation is exercised here for main clauses or combined clauses. For instance, a relative clause has to be produced by the pupils. Here again our natural language generator allows to verifiy or falsify a pupil's choice automatically. Accordingly, an appropriate feedback is selected from a fixed list of error messages or compliments.

Currently we are developing new exercises for teaching pupils (in)direct speech, because we identified this as a serious problem when we analysed corpus material of transliterated essays that had been handwritten by pupils. Moreover, we are going to integrate grammar teaching exercises with the sentence combining exercises. Our system is being evaluated with 4th grade pupils.

Literature (soon two longer articles will be published here)

• Toward accurate syntactic feedback in a writing lab for Germanspeaking elementary schoolers: A generation-based approach
  Karin Harbusch, Gergana Itsova, Ulrich Koch, and Christine Kühner
  in the Proceedings of the CALICO workshop on "Automatic Analysis of Learner Language", San Francisco, CA, USA, 2008.
• "The Sentence Fairy": NLP techniques in support of essay writing by German-speaking elementary schoolers
  Karin Harbusch, Ulrich Koch, Christine Kühner, Gergana Itsova, and Camiel van Breugel
  in the Proceedings of the iicall2 2007: Interfaces of Intelligent Computer-Assisted Language Learning, Waterloo, Canada, 2007.

For questions and comments contact:

Phone: +49 261 287 2600 or 2604
Fax:      +49 261 287 2601 or +49 261 287 100 2604

harbusch@informatik.uni-koblenz.de