9.33.1.Phrases, Phrase Sets and Phrase Classes

An annotation phrase is a pair of a key (a string that is supposed to be a word or phrase of some natural language) and a value (of any type if its serialization is shorter than 2 kilobytes or a string shorter than 10 megabytes). The key phrase will be divided into words and normalized by language-specific functions used by free text search; after normalization it should contain from one to four words. The associated value may be very long in principle, but it will add noticeable overheads related to memory allocation and copying so it is recommended to keep it short if performance is somehow important. It is usual to make the value as short as an ID in some external "detailed" table, not only for speed but to support multiple synonyms for one thing or names of a thing in different languages.

Annotation phrases are grouped in "phrase sets" and every "phrase set" belongs to some "phrase class".

Phrase classes are enumerated in a DB.DBA.SYS_ANN_PHRASE_CLASS table.

  APC_ID integer not null primary key,
  APC_NAME varchar(255) unique,         -- unique name for use in API/UI
  APC_OWNER_UID integer,                -- references SYS_USERS (U_ID), NULL if the record writable for any reader
  APC_READER_GID integer,               -- references SYS_USERS (U_ID), NULL if the record is readable for public
  APC_CALLBACK varchar,

Phrase sets and phrase classes may contain confidential information. E.g., a phrase set may be used to check that a given outgoing document does not mention titles of confidential projects or names of persons that should act anonymously for public. So any application may specify access restrictions when a phrase class is created. The application usually sets APC_CALLBACK to name of some of its functions (but it may be any string) and APC_APP_ENV to value of any type (application may pass it to the APC_CALLBACK function as one of arguments). Phrase match processor only keeps these data in memory, unchanged, and returns to the application as a part of text processing report, so they can be used for any purpose.

Phrase class describes an access to some application-specific callback, but it does not mention any phrases at all. Individual phrases are grouped into phrase sets. Each phrase set belongs to exactly one phrase class and is restricted to one language handler. It does not necessarily mean that it should consist of phrases of one natural language, because many language handlers support mix of languages, but it may apply some restrictions.

Phrase sets are enumerated in a DB.DBA.SYS_ANN_PHRASE_SET table.

  APS_ID integer not null primary key,
  APS_NAME varchar(255) unique,         -- unique name for use in API/UI
  APS_OWNER_UID integer,                -- references SYS_USERS (U_ID), NULL if the record writable for any reader
  APS_READER_GID integer,               -- references SYS_USERS (U_ID), NULL if the record is readable for public
  APS_APC_ID integer not null,          -- references SYS_ANN_PHRASE_CLASS (APC_ID)
  APS_LANG_NAME varchar not null,       -- name of language handler that is used to split texts of phrases
  APS_APP_ENV any,
  APS_SIZE any,                         -- approximate number of phrases inserted in the set (actual or estimate for future)
  APS_LOAD_AT_BOOT integer not null     -- flags whether phrases should be loaded at boot time.

APS_APP_ENV can be of any type (application may pass it to the APC_CALLBACK function as one of arguments). Like APC_APP_ENV, phrase match processor only keeps it in memory and returns in a text processing report.

The APS_SIZE tweaks the amount of memory consumed by a phrase set when it's loaded. The average "price" of placing a phrase to a set is 2 bytes per phrase. Memory amounts are less than a kilobyte while APS_SIZE is below 256, less than 64 kilobytes while APS_SIZE grow up to 32 thousands of phrases and reaches its maximum of 64 megabytes when APS_SIZE reaches its practical limit of 32 million phrases. Only rough similarity to the actual number of inserts is needed, say, same order of magnitude. The exaggerated value of APS_SIZE will allocate a bit more memory and may slightly improve the speed. It is not practical to set APS_SIZE much smaller than it should be because it will cause frequent table lookups and disk buffers in use will overweight any memory savings in the phrase match processor. If there are numerous phrase sets that are used only occasionally, do not decrease APS_SIZE, instead set APS_LOAD_AT_BOOT to zero for them.

Note that if phrase set is edited intensively during server run then the number of inserted phrases is important, an effect of phrase removal is visible only after server restart. E.g. if a phrase set is frequently cleaned and refilled with new phrases then it is much better to delete an obsolete set and create a new one.

Individual annotation phrases are stored in a table DB.DBA.SYS_ANN_PHRASE, that should not be modified by applications directly by data manipulation statements. The content of the table is used to build special search structures in memory and SQL operations like INSERT and DELETE can not keep in-memory structures in sync with the content of the table. DBA may read the table but should not update; in case of occasional update all phrase sets that contain edited phrases may become unusable until server restart.

create table DB.DBA.SYS_ANN_PHRASE
  AP_APS_ID integer not null,           -- references SYS_ANN_PHRASE_SET (APS_ID),
  AP_CHKSUM integer,                    -- phrase check-sum
  AP_TEXT varchar,                      -- original text
  AP_LINK_DATA any,                     -- Associated data about links etc.
  AP_LINK_DATA_LONG long varchar,       -- Same as AP_LINK_DATA but for long content, one of two is always NULL
  primary key (AP_APS_ID, AP_CHKSUM, AP_TEXT)

The "almost direct" way of writing to the table is via BIF ap_add_phrases . It gets two arguments, the integer ID of annotation phrase set and a vector of descriptions of phrases that should be edited in that phrase set. Every item of vector of descriptions is in turn vector of one or two values; first value is the text of the phrase, second value is associated application specific data, the absence of second value indicates that the phrase should be removed. If same text of phrase appears in the vector of description more than once, and associated data differ then any version of data can be stored for future use; it is the roll of dice because the vector is reordered for faster processing.