PyDelphin at the Command Line

While PyDelphin primarily exists as a library to be used by other software, for convenience it also provides a top-level script for some common actions. There are two ways to call this script, depending on how you installed PyDelphin:

  • delphin.sh - if you’ve downloaded the PyDelphin source files, this script is available in the top directory of the code
  • delphin - if you’ve installed PyDelphin via pip, it is available system-wide as delphin (without the .sh)

For this tutorial, it is assumed you’ve installed PyDelphin via pip and thus use the second form.

The delphin command has several subcommands, described below.

See also

The delphin.commands module provides a Python API for these commands.

convert

The convert subcommand enables conversion of various *MRS representations. You provide –from and –to arguments to select the representations (the default for both is simplemrs). Here is an example of converting SimpleMRS to JSON-serialized DMRS:

$ echo '[ "It rains." TOP: h0 RELS: < [ _rain_v_1<3:8> LBL: h1 ARG0: e2 ] > HCONS: < h0 qeq h1 > ]' \
  | delphin convert --to dmrs-json
[{"surface": "It rains.", "links": [{"to": 10000, "rargname": null, "from": 0, "post": "H"}], "nodes": [{"sortinfo": {"cvarsort": "e"}, "lnk": {"to": 8, "from": 3}, "nodeid": 10000, "predicate": "_rain_v_1"}]}]

As the default for –from and –to is simplemrs, it can be used to easily “pretty-print” an MRS (if you execute this in a terminal, you’ll notice syntax highlighting as well):

$ echo '[ "It rains." TOP: h0 RELS: < [ _rain_v_1<3:8> LBL: h1 ARG0: e2 ] > HCONS: < h0 qeq h1 > ]' \
  | delphin convert --pretty-print
[ "It rains."
  TOP: h0
  RELS: < [ _rain_v_1<3:8> LBL: h1 ARG0: e2 ] >
  HCONS: < h0 qeq h1 > ]

Some formats are export-only, such as dmrs-tikz:

$ echo '[ "It rains." TOP: h0 RELS: < [ _rain_v_1<3:8> LBL: h1 ARG0: e2 ] > HCONS: < h0 qeq h1 > ]' \
  | delphin convert --to dmrs-tikz
\documentclass{standalone}

\usepackage{tikz-dependency}
\usepackage{relsize}
[...]

See delphin convert –help for more information.

select

The select subcommand selects data from an [incr tsdb()] profile using TSQL queries. For example, if you want to get the i-id and i-input fields from a profile, do this:

$ delphin select 'i-id i-input from item' ~/grammars/jacy/tsdb/gold/mrs/
11@雨 が 降っ た ．
21@太郎 が 吠え た ．
[..]

In many cases, the from clause of the query is not necessary, and the appropriate tables will be selected automatically. Fields from multiple tables can be used and the tables containing them will be automatically joined:

$ delphin select 'i-id mrs' ~/grammars/jacy/tsdb/gold/mrs/
11@[ LTOP: h1 INDEX: e2 ... ]
[..]

The results can be filtered by providing where clauses:

$ delphin select 'i-id i-input where i-input ~ "雨"' ~/grammars/jacy/tsdb/gold/mrs/
11@雨 が 降っ た ．
71@太郎 が タバコ を 次郎 に 雨 が 降る と 賭け た ．
81@太郎 が 雨 が 降っ た こと を 知っ て い た ．

See delphin select –help for more information.

mkprof

Rather than selecting data to send to stdout, you can also output a new [incr tsdb()] profile with the mkprof subcommand. If a profile is given via the –source option, the relations file of the source profile is used by default, and you may use a --where option to use TSQL conditions to filter the data used in creating the new profile. Otherwise, the –relations option is required, and the input may be a file of sentences via the –input option, or a stream of sentences via stdin. Sentences via file or stdin can be prefixed with an asterisk, in which case they are considered ungrammatical (i-wf is set to 0). Here is an example:

$ echo -e "A dog barks.\n*Dog barks a." \
  | delphin mkprof \
      --relations ~/logon/lingo/lkb/src/tsdb/skeletons/english/Relations \
      newprof
9746   bytes  relations
67     bytes  item

Using --where, sub-profiles can be created, which may be useful for testing different parameters. For example, to create a sub-profile with only items of less than 10 words, do this:

$ delphin mkprof --where 'i-length < 10' \
                 --source ~/grammars/jacy/tsdb/gold/mrs/ \
                 mrs-short
9067   bytes  relations
12515  bytes  item

See delphin mkprof –help for more information.

process

PyDelphin can use ACE to process [incr tsdb()] testsuites. As with the art utility, the workflow is to first create an empty testsuite (see mkprof above), then to process that testsuite in place.

$ delphin mkprof -s erg/tsdb/gold/mrs/ mrs-parsed
 9746  bytes  relations
 10810 bytes  item
 [...]
$ delphin process -g erg-1214-x86-64-0-9.27.dat mrs-parsed
NOTE: parsed 107 / 107 sentences, avg 3253k, time 2.50870s

The default task is parsing, but transfer and generation are also possible. For these, it is suggested to create a separate output testsuite for the results, as otherwise it would overwrite the results table. Generation is activated with the -e option, and the -s option selects the source profile.

$ delphin mkprof -s erg/tsdb/gold/mrs/ mrs-generated
 9746  bytes  relations
 10810 bytes  item
 [...]
$ delphin process -g erg-1214-x86-64-0-9.27.dat -e -s mrs-parsed mrs-generated
NOTE: 77 passive, 132 active edges in final generation chart; built 77 passives total. [1 results]
NOTE: 59 passive, 139 active edges in final generation chart; built 59 passives total. [1 results]
[...]
NOTE: generated 440 / 445 sentences, avg 4880k, time 17.23859s
NOTE: transfer did 212661 successful unifies and 244409 failed ones

See delphin process –help for more information.

See also

The art utility and [incr tsdb()] are other testsuite processors with different kinds of functionality.

compare

The compare subcommand is a lightweight way to compare bags of MRSs, e.g., to detect changes in a profile run with different versions of the grammar.

$ delphin compare ~/grammars/jacy/tsdb/current/mrs/ \
                  ~/grammars/jacy/tsdb/gold/mrs/
11  <1,0,1>
21  <1,0,1>
31  <3,0,1>
[..]

See delphin compare –help for more information.

See also

The gTest application is a more fully-featured profile comparer, as is [incr tsdb()] itself.

repp

A regular expression preprocessor (REPP) can be used to tokenize input strings.

$ delphin repp -c erg/pet/repp.set --format triple <<< "Abrams didn't chase Browne."
(0, 6, Abrams)
(7, 10, did)
(10, 13, n’t)
(14, 19, chase)
(20, 26, Browne)
(26, 27, .)

PyDelphin is not as fast as the C++ implementation, but its tracing functionality can be useful for debugging.

$ delphin repp -c erg/pet/repp.set --trace <<< "Abrams didn't chase Browne."
Applied:!^(.+)$                \1
   In:Abrams didn't chase Browne.
  Out: Abrams didn't chase Browne.
Applied:!'            ’
   In: Abrams didn't chase Browne.
  Out: Abrams didn’t chase Browne.
Applied:Internal group #1
   In: Abrams didn't chase Browne.
  Out: Abrams didn’t chase Browne.
Applied:Internal group #1
   In: Abrams didn't chase Browne.
  Out: Abrams didn’t chase Browne.
Applied:Module quotes
   In: Abrams didn't chase Browne.
  Out: Abrams didn’t chase Browne.
Applied:!^(.+)$                \1
   In: Abrams didn’t chase Browne.
  Out:  Abrams didn’t chase Browne.
Applied:!  +
   In:  Abrams didn’t chase Browne.
  Out: Abrams didn’t chase Browne.
Applied:!([^ ])(\.) ([])}”"’'… ]*)$           \1 \2 \3
   In: Abrams didn’t chase Browne.
  Out: Abrams didn’t chase Browne .
Applied:Internal group #1
   In: Abrams didn’t chase Browne.
  Out: Abrams didn’t chase Browne .
Applied:Internal group #1
   In: Abrams didn’t chase Browne.
  Out: Abrams didn’t chase Browne .
Applied:!([^ ])([nN])[’']([tT])               \1 \2’\3
   In: Abrams didn’t chase Browne .
  Out: Abrams did n’t chase Browne .
Applied:Module tokenizer
   In:Abrams didn't chase Browne.
  Out: Abrams did n’t chase Browne .
Done: Abrams did n’t chase Browne .

See delphin repp –help for more information.

See also