CHAPTER 15

Conceptual Dependencies and Scripts

15.1 The Need for Scripts

The success of the conceptual dependencies theory when applied to the processing of individual sentences, and its failure to deal adequately with more extended narrative, drew attention to the need for scripts. The use of scripts provides a mechanism for representing the implicit or underlying world knowledge of a person, which does not always find expression in spoken or written narrative. Consider the small fragment of narrative below:

'John decided to take Mary to see afilm. They drove to the centre of town in his car. He found that he hadforgotten his wallet. They went for a walk in the park.'

To an NL-system which has not been provided with the appropriate world knowledge, these four sentences would appear to be about completely dissociated actions. The knowledge essential to an understanding of this narrative, which is not given explicitly, is:

That films are shown in cinemas.
That cinemas are often located near the centre of town.
That money must be paid to gain entry to a cinema.
That a wallet is a container for money.
That parks can usually be entered without payment.


When we tell a story we can assume that our human listeners have all this information before we start. The story therefore consists simply of markers and signposts which enables the listener to select from memory the appropriate sets of knowledge, and guide them through the structure of that knowledge as through a labyrinth of alternative passageways. For an NL-system, however, this knowledge must be provided explicitly, and scripts provide us with the means to do this.

15.2 Implementing Scripts with Conceptual Dependency Structures

In section 14.5 we described the algorithm for translating the input text into a conceptual dependency structure. The input to that process is provided in the form of a list INLIST, and the output is produced in the form of a list (CST ACK) which contains a number of 'CD-FORMS' or conceptual dependency structures.

The INLIST is processed word by word in left to right order. Each word is looked up in the 'lexicon', and the corresponding 'packet' is extracted. The packet contains a number of 'requests' (which are in fact demons or functions) which, when activated, construct new CD-FORMs, or modify existing CD­FORMs, or insert the completed forms into the CST ACK. Let us call this process the 'TX-CD process' (text to conceptual dependency process).

The next part of the overall process is to assign unique identifiers to the objects or 'picture producers' (PPs) represented in the CD-FORMs. World knowledge about PPs is provided by a structure called the PP-memory. This is a lookup table which for each PP provides information about its properties: (class of object, physical object, human, ...) It also describes it in functional terms (what it is used for). This represents its usual role in a script. For example, a 'chair' is an object used by people for sitting on. This information is held in the form of CD-FORM structures with formal parameters where these cannot be identified. The problem is therefore to take the PPs identified by the TX-CD process and assign to each either a unique new identifier (a 'token') or the identifier of an existing PP, which has already been analysed, or of one which is a special 'world knowledge' PP (that is, a PP corresponding to some known unique entity such as 'Mount Everest' or 'Henry VIII'). These will be held in the PP-memory.

Assigning tokens to PPs involves the resolution of references. This in turn requires the matching of CD-FORMs in order to unify the representations of PPs. In some cases resolution cannot be completed until the next stage of the overall process. For the most part, however, this part of the process takes as input the CST ACK and outputs the CST ACK with a token appended to each PP. Let us call this part of the process the 'PP- ID process' (PP identifier process).

The last part of the overall process introduces scripts. It has a dictionary of scripts, and its first task is to select the appropriate script, or set of scripts, for the narrative being analysed, based on the part already analysed. The second task is to predict what. further scripts are likely to be required. The third task is to instantiate the PPs identified in the narrative by the PP-ID process with the roles in the scripts. Once again we have the insertion of entities into case-slots, but now the case­slots are part of a rather larger scenario and the construction of these scenarios is rather more flexible and data driven. Let us call this part of the overall process the 'SC-APP process' (script application process).

The complete process therefore consists of the function calls:

TX-CD(INLIST) => CSTACK
PP-ID(CLIST) => CSTACK + tokens
SC-APP(CLIST+TOKENS) => script + instantiations