Skip to content

ErgReleases

StephanOepen edited this page Sep 13, 2016 · 29 revisions

Background

This page (a work in progress, like so many others on this wiki) collects some practical and historic information around official snapshots of the ERG, e.g. officially released versions of the grammar.

(Pre-)Release MRS Quality Control

Once new treebanked and thinned profiles are available, run a set of automated wellformedness tests on the MRSs, for example:

  $LOGONROOT/redwoods --terg --default \
    --filter syntax,lnk,fragmentation erg/trunk/mrs/12-02-06/pet.1

Optimize the Set of Quick-Check Paths

 cheap -limit=100000 -repp -cm -packing \
   -qc-unif=0 -compute-qc-unif=pet/qc.tdl english.grm < lkb/checkpaths.items 

Re-Generate the Core SEM-I

The bulk of the semantic interface (SEM-I) is auto-generated from the lexicon (recorded as ‘core.smi’):

  (in-package :mt)
  (setf semi
    (construct-semi       
     :ids t :rules t :descendp t :embedp t
     :semi (read-semi
            "~/src/logon/lingo/erg/etc/erg.smi"
            :includep nil :finalizep nil :recordp nil)
     :patches "~/src/logon/lingo/erg/etc/patches.lisp"
     :finalizep t))
  (print-semi
   semi :stream "~/src/logon/lingo/erg/etc/hierarchy.smi"
   :format :hierarchy)
  (print-semi
   semi :stream "~/src/logon/lingo/erg/etc/abstract.smi"
   :format :compact :filter "^[^_]")
  (print-semi
   semi :stream "~/src/logon/lingo/erg/etc/surface.smi"
   :format :compact :filter "^_")

The master file ‘erg.smi’ is manually maintained and includes the auto-generated entries; the separate ‘patches.lisp’ will hopefully become unnecessary after the 1214 release.

Validate the Predicates Used in Transfer Rules

Transfer rules operate on MRSs, hence should be compatible with the specifications of the Semantic Interface (SEM-I):

(loop
      for mtrs in *transfer-rule-sets*
      do
        (format t "~%~a:~%" (mtrs-id mtrs))
        (loop
            for mtr in (mtrs-mtrs mtrs)
            for eps
            = (append
               (test-semi-compliance (mtr-filter mtr))
               (test-semi-compliance (mtr-context mtr))
               (test-semi-compliance (mtr-input mtr))
               (test-semi-compliance (mtr-output mtr)))
            for preds
            = (loop
                  for ep in eps
                  for pred = (mrs:rel-pred ep)
                  when (and pred
                            (or (stringp pred) (symbolp pred))
                            (not (equal pred "never_unify")))
                  collect pred)
            when preds
            do
              (format t "  ~a: ~{~s~^ ~}~%" (mtr-id mtr) preds)))

Validate and Update the (DT) Head Table

To identify new rules (or ones deleted from the grammar), the following will compare the head table on file (by default ‘etc/rules.hds’) to the grammar inventory of constructions:

  (in-package :tsdb)
  (read-heads "~/src/logon/lingo/terg/etc/rules.hds" :test (list #'lkb::dt-test))

Validate the Mapping from Generic to Native Lexical Entries

The definitions of lexical types for generic lexical entries should be annotated with ‘equivalences’ in terms of native lexical entries and (optionally) lexical rules, e.g.

  ;; <type val="v_np*_pr-3s-unk_le">
  ...
  ;; <native> v_3s-fin_olr v_np*_le
  ;; </type>

These pseudo-XML comments (as originally introduced for the lexical type database) are interpreted at grammar load time and can be serialized into summary form (say for comparison to lexical entries defined in ‘gle.tdl’) as follows:

  (loop
      for context in *tdl-all-contexts*
      for type = (rest (assoc :type context))
      for native = (rest (assoc :native context))
      when (and type native)
      collect (cons type native) into result
      finally
        (loop
            for (type . native) in (sort result #'string< :key #'first)
            do (format t "~a: ~a~%" type native)))

In addition to the above, we should also confirm that there are ‘native’ correspondences declared for all types instantiated by generic lexical entries (with possible exceptions, e.g. the generic type for ‘year of century’ named entities, for which there is no native equivalent):

  (with-open-file (stream "~/src/logon/lingo/terg/gle.tdl")
    (loop
        for line = (read-line stream nil nil)
        for i from 1
        while line do
          (multiple-value-bind (start end starts ends)
              (ppcre:scan ".+:=[ \\t]*([^ \\t]+)" line)
            (when (and starts ends)
              (let* ((type (subseq line (aref starts 0) (aref ends 0)))
                     (match
                      (find type *tdl-all-contexts*
                       :key #'(lambda (foo) (rest (assoc :type foo)))
                       :test #'string=))
                     (native (rest (assoc :native match))))
                (when (or (null native) (string= native ""))
                  (format
                   t "missing native declaration for `~(~a~)' (line ~d)~%"
                   type i)))))))

Re-Generate CTYPE Hierarchy and Abstractions

For the various distinct configurations, maintaining the inventory of construction types requires some manual resoluton; to generate a point of reference (for the current grammar, e.g. including the speech extensions):

  (generate-ctypes :file "/tmp/ctype.tdl")

Generate Maximum Entropy (and Optionally PCFG) Models

Parse ranking models included with the grammar are trained on the standard training plus development splits, for example Sections 01 to 12 for WeScience. By default, the Maximum Entropy training scripts (re-)generate a fresh feature cache, hence the following two jobs must not run in parallel

  sbatch ${LOGONROOT}/uio/titan/redwoods \
    --redwoods --run train.wescience.lisp

  sbatch ${LOGONROOT}/uio/titan/redwoods \
    --redwoods --run train.redwoods.lisp

Unlike in the 1111 release, no PCFG model (for chart pruning in PET) is included with the 1212 release and onwards.

Update Summary Statistics of Redwoods Treebanks

Since its October 2010 release, the ERG includes a spreadsheet that summarizes key statistics of the gold-standard profiles that comprise the Redwoods Treebank. The raw data for addition to the file ‘etc/redwoods.xls’ can be generated automagically:

  (in-package :tsdb)
  (loop
      with *phenomena* = nil
      with *statistics-aggregate-dimension* = :phenomena
      with *statistics-all-rejections-p* = t
      with *tsdb-home* = (logon-directory "lingo/terg/tsdb/gold" :string)
      initially (purge-profile-cache :all)
      for db in (find-tsdb-directories)
      for name = (get-field :database db)
      do (analyze-trees name :append "/tmp/redwoods.csv" :format :csv))

Populate the Linguistic Type Database

Instructions on invoking the relevant script to populate the type database can be found at LkbLtdb.

Create tar(1) Ball for Download

In addition to being able to directly obtain tagged releases from the DELPH-IN SVN repository, there is a collection of compressed archives, one for each release, for direct download available at http://www.delph-in.net/ftp/erg/ (hosted on the main DELPH-IN server in Oslo, though one could easily re-direct this to the LinGO server). A soft link current.tgz in the download directory is what selects which version will be accessed by the Download button on the main ERG page.