One of the main innovations for the new ENQUIRE interface was incremental query expansion. In previous versions of Okapi, term extraction from relevant documents took place only when query expansion was explicitly requested, and the top 20 terms were then used in a search or shown to the user. By contrast the method adopted for ENQUIRE involves dynamic term extraction.
Whenever a document is judged relevant the textual parts of it --- descriptors, title, abstract, but not date of publication or class-code --- are sent to the BSS as query strings to be ``parsed'' into their constituent terms. The query layer maintains a list of extracted terms, with a running total of the number of relevant documents in which they occur. After every relevance judgement the terms currently making up the query are re-weighted and re-ordered if appropriate, and the extracted terms are examined to see whether they have gained enough weight to be included automatically in the query. If so, they are added to the list for the interface script to display.
As with other thresholds, the decision about what test to apply here was based upon experiment. The main objective was to produce acceptable behaviour at the interface, introducing one or two new terms after each relevance judgement in a way that would be easily controllable by the user. For the current experiments, the first criterion is that the term must have occurred in at least 2 relevant documents. The next is that its weight must be the average weight of the existing query terms.
These criteria will undoubtedly be fine-tuned following more intensive data-collection and analysis. There are some obvious flaws at present, for instance if the initial query contains only one term it is virtually impossible for any others to achieve a high enough weight to be added to it. Likewise the list tends to become resistant to latecomers once half-a-dozen relevance judgements have been made. At the same time, extraction from the INSPEC abstracts seems sometimes to bring up very frequent words which are too general in meaning to be useful, suggesting that this expansion process might benefit from a very comprehensive stop-word list.