Gender and Animacy

PNG.GENDER and PNG.ANIMACY are neceesary especially for blocking overgeneration. For example, 他-he;她-she;它-it.

Semantically Empty Items

In order to create semantically void items in generation without overgeneration, trigger.mtr should contains generator rules like the followings.

了_pfv_gr := arg0e_gtr &
[ CONTEXT.RELS <! [ ARG0.E.ASPECT perfective ] !>,
  FLAGS.TRIGGER "了_pfv" ]. 

吗_qp_gr := arg0e_gtr &
[ CONTEXT.RELS <! [ ARG0.SF ques ] !>,
  FLAGS.TRIGGER "吗_qp" ]. 

在_r_gr := arg0e_gtr &
[ CONTEXT.RELS <! [ ARG0.E.ASPECT imperfective ] !>,
  FLAGS.TRIGGER "在_r" ].

被_v_gr_1 := arg0e_gtr &
[ CONTEXT.RELS <! [ ARG1 individual & #i ] !>,
  FLAGS [ SUBSUME < #i >,
          TRIGGER "被_v" ] ].

Note that unwanted parse trees in parsing also cause problems in generation. For example, “*狗 在 在 在 … 叫” is ill-formed, and it should not be generated as well as parsed.

generation.ignore

There can be some items (including lexical entries and some lexical/phrasal rules) that should be ignored in generation process. Discarding such items in generation contributes to the speed of generation as well as the felicity of generated sentences. When LKB is used, the ignored items are specified in lkb/globals.lsp

;;; Blocking Generation (lexicons)
(setf *duplicate-lex-ids*
  '(
    了_pfv_robust 着_dur_robust 过_exp_robust
    ))

When ACE is used, the items are listed up in generation.ignore (under the root of Zhong). Note that generation.ignore should be evoked by the ACE configuration file.

generation-ignore-signs   := "generation.ignore".

STYLE feature

This flag feature exists outside of SYMSEM and plays a role to generate only proper sentences following the theorum parsing robustly, generating striclty. This means that even a little bit awkward sentence should be parsed, but the sentence should also be filtered out in generation. For example, although “张三 去 着” may not sound good to Chinese native speakers, our grammar provides a parse tree for the sentence with a flag [STYLE robust]. Nonetheless, our grammar does not generate such a sentence, and the STYLE feature keeps the sentence from being generated.

sign :+
 [ STYLE style,
   IDIOM bool ].
style := avm &
 [ WRITTEN luk ].
strict := style &
 [ WRITTEN + ].
robust := style.
spoken := robust & 
 [ WRITTEN - ].
unproductive := robust.
dialect := robust.

In order to realize parsing robustly, generating striclty, there are two types of roots: roots.tdl and roots-strict.tdl. The former, an ordinary one, works in parsing, and the latter with [STYLE strict] in generation. Thus, when we compile a grammar for parsing, we can create an ACE data file (e.g. hans.dat) with roots.tdl. On the other hand, we can use roots-strict.tdl for creating the other ACE data file for generation separately (e.g. hans-strict.dat).

LE ZHE GUO

There are three basic aspect markers in Mandarin Chinese; namely 了 (le), 着 (zhe), and 过 (guo). Not all these three items can be necessarily attached to all verbs. That is, a verb can choose these markers, and the choice seems to be registered into the lexical information of each verb. For example, “张三 去 着” does not sound perfect, because 去 does not co-occur well with 着. If we do not impose any constraint on this choice, there would be a number of overgenerated sentences. For this purpose, there is a head feature for verb: LE-ZHE-GUO.

lzg := avm. 
le := lzg.
zhe := lzg.
guo := lzg.
no-lzg := lzg.
le+zhe := le & zhe.
le+guo := le & guo.
zhe+guo := zhe & guo.
le+zhe+guo := le & zhe & guo.

verb :+ 
 [ LZG lzg ].

Each marker is divided into two subtypes; one for an ordinary attachment, and the other for a robust parsing. For instance, 着_dur_robust in the following can be attached to 去 whose LE-ZHE-GUO feature value is le+guo (i.e. 了 +, 着 -, 过 +).

着_dur := adv-dur-v-post &
  [ STEM <"着">,
    SYNSEM.LOCAL.CAT.HEAD.MOD < [ LOCAL.CAT.HEAD.LZG zhe ] > ].

着_dur_robust := adv-dur-v-post &
  [ STEM <"着">,
    STYLE robust,
    SYNSEM.LOCAL.CAT.HEAD.MOD < [ LOCAL.CAT.HEAD.LZG le+guo ] > ].

Note that 着_dur_robust has [STYLE robust]. When we parse and generate sentence in a strict way, this item will be sorted out. In addition, this item will be filtered out in generation because of the following rule in trigger.mtr.

着_dur_robust_gr := arg0e_gtr &
[ CONTEXT.RELS <! [ PRED "non_existing_rel" ] !>,
  FLAGS.TRIGGER "着_dur_robust" ]. 

Additionally, 着_dur_robust is included in the list specified in generation.ignore and globals.lsp.

Punctuations

The current grammar regards only sentences with punctuations as regular (strict) ones. When we use a strict version, all sentences show up with punctuation markers depending on the sentence force (SF).

period-marker := punctuation &
  [ STEM < "。" >,
    SYNSEM.LOCAL.CONT.HOOK.INDEX.SF prop ].

That implies punctuation markers in Chinese should not be ignored by repp.

Questions

1. Can a plain sentence be used for expressing questions? (e.g. 张三 叫 了 => 张三 叫 了 吗)

This is contrained by STYLE and punctuation markers explained before.

2. Can a sentence without 了 be generated into a sentence with 了? (e.g. 张三 叫 => 张三 叫 了) Likewise, can a sentence with 了 be generated into a sentence without 了? (e.g. 张三 叫 了 => 张三 叫)

No! A plain sentence 张三 叫 does not have to be paraphrased into a perfective one 张三 叫 了, and vice versa. In order to handle this constraint, a lexical rule was added in lrules.tdl. Now main-verb-lex and some related verbal items inherently have [MC na], and this feature makes the verb by itself cannot constitute a sentence (or clause).

non-aspect-lex := aspect-lex-rule &
 [ SYNSEM.LOCAL [ CAT.MC +,
                  CONT.HOOK.INDEX.E.ASPECT non-aspect ] ].

Last update: 2014-09-11 by SanghounSong [edit]