Text Structuring

The second step of RST Analysis involves structuring the text. The text structuring interface of the RST-Tool allows the user to connect the segments into a rhetorical structure tree, as shown in figure 3. For presentation, the RST-Tool follows the graphical style presented in Mann & Thompson (1987).

  
Figure 3: Text Structuring Tool

Initially, all segments are unconnected, ordered at the top of the window. The user can then drag the mouse from one segment (the nucleus) to another (the satellite). Upon releasing the mouse button, the system offers a menu of relations to choose from (the user can use the relation-sets provided with the system, or provide their own).

The user can switch freely between text segmentation and text structuring mode -- to edit text, or to change segment boundaries. The system keeps track of the structure assigned so far. If the user, in editing the text, deletes a segment, the system forgets structuring information concerning that segment.

Types of Structure

Apart from simple nuclear-satellite relations, the tool allows a number of other options:

• Multi-nuclear Relations: such as JOINT and SEQUENCE (see figure 4).

• Text Scoping: sometimes we might want to treat a nucleus-satellite complex as itself the nucleus (or satellite) of a relation. The RST-Tool allows you to insert a span node over a nucleus and its sub-structure. This span-node can then fill either nucleus or satellite roles. For instance, the spans labelled ``1-3'' and ``2-3'' in figure 1 were created in this way.

• Schemas: Pure RST would seem to analyse a text into a single top-nucleus, with all the rest of the text being a satellite of that unit, or a satellite of a satellite, etc. Multi-nuclear relations alleviate this weak premise somewhat, but are not the whole solution. Being unhappy with this situation, we added the capability to use McKeon-style schemas in the RST-Tool, what are sometimes called ``story grammars''. These allow a sequence of named elements of structure, e.g., INTRODUCTION, BODY, CONCLUSIONS, BIBLIOGRAPHY, etc.(see figure 5). We find that the macro-structure of a text is best modelled with schemas, while the lower levels are captured through RST relations.

• Clause-Internal Structure: in my work on rst-based text summarisation, I have found that better quality summarisation can be achieved if we push RST-analysis into the clause, and allow the clausal adjuncts to be treated as satellites of the nuclear clause itself. For instance, (N: Edward surrendered,)(S: in 1245). Some of these adjuncts can be connected to the clause with standard RST relations, but many can not. A set of new relations, borrowed from the Systemic labelling of adjuncts (cf. Halliday 1985), has been added for this reason.

  
Figure 4: Scoping and Multi-nuclear Relations

  
Figure 5: Constituent Structure

Defining Relations

When forming a relation between two text-nodes, a menu pops up prompting the user to specify the relation-type. These relations are not built into the system, but can be edited by the user. The user modifies a text file to add, delete or modify the relation set. A sample set derived from Mann & Thompson (1987) is included with the distribution.

Other Points

Because rst-structures can become very elaborate, the RST Tool allows the user to collapse sub-trees -- hiding the substructure under a node, This makes it easier, for instance, to connect two nodes which normally would not appear on the same page of the editor.

The user can save the present state of the screen as postscript, for inclusion in Latex documents. Alternatively, a snapshot utility can be used to save selected parts of the structure in other formats. The structured text can be saved to a file, for later re-editing, or for use in variable-length document presentation.