A Treatise on the Nature of TSDB and ancillary bits and pieces

What follows are the notes of a recent Expedition into the depth of TSDB, in the months of August and September MMXIV. These notes may be incomplete, misleading, or flat-out wrong, but are provided here in the hope that they may be of some use to future travellers.

The Lay of the Land or, the database schema

TSDB is, when you get down to it, just a relational database. But instead of using a standard SQL database, the data are stored in flat text files on disk. A TSDB profile (that is, a database) is stored in a folder. The file relations describes the database schema of this profile; its syntax is described in the TSNLP User Manual (Volume 2) (p. 27). Individual relations (or tables, if you prefer) are stored in separate files with the same name as the relation. The SQL-like query language TSQL, used to query profiles, is also documented here (pp. 26-27).

A practical upshot of this idiosyncracy is that if your needs are modest, the database can be queried with standard Unix text processing tools such as cut and grep, AWK, and/or Perl. The relation files contain a single row per line, with fields separated by the '@' character.

A RedWoods-style profile contains the following relations: analysis, daughter, decision, edge, fold, item, item-phenomenon, item-set, output, parameter, parse, phenomenon, preference, result, rule, run, score, set, tree, update. Not all of these are understood by yr hmbl crrspndnt, but I will describe what I know.

An item is a sentence. This relation contains various things, including the input sentence itself. Shared information for a parse run of the sentence is stored in the parse relation, with individual parser outputs stored in the result table.

Annotations of sentences are linked to rows in the parse table (since annotation is the selection of a gold parse from a parse forest), through the tree relation. However, since the annotator may choose to only partially disambiguate a parse forest, tree only contains shared information (including a version field, so that old annotations don't have to be overwritten on subsequent annotation passes). Selected trees (in the result table) are linked to in preference (which also links back to tree, of course).

Into the Wild or, the Lisp API

If your needs are more specialised, your most likely port of call is the Lisp API. A Lisp session can be started with the M-x logon command (cf. LogonInstallation), but this runs a precompiled binary which may or may not have been updated recently. If your site has a licence for the Allegro Common Lisp compiler (cf. LogonExtras, your local LOGON bodhisattva), the system can also be started as follows:

  1. In emacs, run M-x lisp. This will start the Lisp REPL, but not LKB or itsdb .

  2. Execute (lmt) (short for Load Machine Translation system I believe)

  3. Read in a grammar with rsa(). For English (rsa :erg) or (rsa :terg) (loading either the most recent release of the ERG or the trunk version) is most likely what you want, or (rsa :deepbank) if working with DeepBank.

If you need to do slightly different things (or are merely curious), lmt() and rsa() are defined in $LOGONROOT/dot.clinit.cl.

If you want to start the system without the GUI (if you are, say, interacting with the system on a big-iron machine, across a temperamental intercontinental internet link, avoiding X forwarding and running emacs -nw in a screen [or tmux] session is likely a good idea) execute (setf (system:getenv "DISPLAY") nil) in between steps 1 and 2.

Next, you will probably want to iterate over the items in some profile. This is most easily accomplished by leveraging the GUI code (but without running the actual GUI stuff [this code clearly predates the spread of the gospel of MVC beyond the borders of the Smalltalk nation]). Observe:

(loop
    with frame = (browse-trees profile :runp nil)
    for id in (lkb:compare-frame-ids frame)
    for nil = (browse-tree profile id frame :runp nil)
    do (format t "~A~%" id))

The GUI code is really quite chatty. It can be silenced by adding the declaration with excl:*initial-terminal-io* = nil before the invocation of browse-trees().

The compare-frame class contains a number of data members that are useful. One of these we've already seen, compare-frame-ids which is a list containing the ids of the items being browsed over. Also of interest are compare-frame-in and compare-frame-out which contain parses selected and not selected, respectively, for the current item (mnemonic: the parses in the set of selected parses and those outside the set. For a full list, see the class declaration in $LOGONROOT/lingo/lkb/src/ACL_specific/compare.lsp (lines 500-621 at the time of writing. Grep for define-application-frame).

Your specific use-case may not be directly supported by compare-frame (or it may be, you just can't find it; more on this in a moment), and some of the information you require may be not be in its data members. However you will likely find code similar enough (or even identical!) to your needs in the body of browse-trees() or, more likely, browse-tree(). However both of these functions are very tightly coupled code, with data model manipulations, application control and GUI logic intermixed. In particular browse-tree() is perhaps a tall order to comprehend in its full 850-line, non-Euclidean glory, and a more expedient approach may be to pick useful bits of code from the bowels of these functions and extract them to the comfort of your own code.

A final useful trick is liberal use of the ELI mode's go to definition functionality (bound by default to C-c .), which takes you to the source location where any function, class or variable is defined.

TsdbTop (last edited 2014-09-04 10:37:33 by ArneSkjærholt)

(The DELPH-IN infrastructure is hosted at the University of Oslo)