Some Issues in Generation From a Semantic Representation

1. Choice of a term to express a concept

To generate a RESULT, one might start by qene­ rating independently the new state and its cause, and then connecting them by an explicit causal for­ mulation, such as a clause introduced by bec~e. With this method, a representation where the RESULT is DEAD and the ACT is an EXPLOSION would be gene­ rated as

The ICM PJt~ident clied y~:teJtday in Pa!!M bec~e anMchi.6:t6 exploded a bomb in h..L-6 hoU6e.

However, this text is far Less satisfactory than one using the verb :to rA11 and dropping becaU6e The ICM PJt~iden:t WM killed y~:teJtday in PCUW., by anMcW:t.6 who exploded a bomb -i_n hi.6 how., e.

* A word written in capital Letters is a mnemonic name for a semantic or pragmatic notion.

P verb such as :to kill is particularly appro­ priate for expressing the result DEAD because it means roughly CAUSE TO DIE · Many verbs or verbal groups in French, as in English, have causal seman­ tics similar to that of :to lzill in the sense that their subject causes a new state to occur fer their object. This is the case for ·various verbs such as :to peJt:t.UJtb, :to di.6coUJtage, :to cut, :to cook, :to pack and compounds such as :to give a hangoveJt, :to put into velt.6e, etc. Adopting the principle that it is better to express oneself as compactly as possible, the first attempt to generate a RESULT rnu~t be tc Look for a verbal expression with the appropriate causal semantics. Let us examine which factors must be taken into account to choose such an expression.

The context in which the RESULT occurs may be significant. As an illustration, consider the case of the result WOUNDED. This result can be expressed for example by :to wound or by :to injUJte. The choice between these two verbs depends upon the context, as shown in the following pair of sentences

John .6ho:t and (wounded + *injUJted)MMy A lightning boU (*Wounded + injUJted) MMy

The former one involving a crime, the Latter one an accident.

The specific action which causes the RESULT may also be relevant in determining which term best expresses the RESULT. For example, in order to cho­ ose verbs such as :to .6:t.Jtangle

A young woman WM .6:t.Jtangl!.e.d :th-i_,6 moJtning by a mentally-ill pelt.6on.

the generation process must be aware of the specific ACT which caused the death of the young woman.

Furthermore, the generation process must consi­ der whether the ACTOR is specified or not. As an illustration, suppose that, when the ACTOR is spe­ cified, the generator produces :

AnMchi.6:t.6 lzilled :the ICM PJt~iden:t

for RESULT = DEAD on a TARGET which is not famous

AnMchi.6:t6 a.6.6M.6ina:ted Pope Max r for RESULT= DEAD on a TARGET which is famous

AnMchi.6:t.6 blew up a police .6:tation ... V~ anMcw:t~ ant 6ai:t explMeJt un corm0.,.6Mia:t de police ·.. for RESULT = DEMOLISHED and ACT = EXPLOSION

*However, this analysis remains controversial in the current Literature.

Now, consider the same cases but for stories in which the ACTOR is not specified. One solution may be to keep the same verbal expression as when the ACTOR is specified, but to use it in another construction, such as the passive ccnstruction wi­ thout an agent. This is possible with t o ~f, t o cLMa.l.lJ.>.tnate. and to blow up but not with 6a.AAe o.xpf0.6eJt which cannot be passivised :

The ICH Ptr.e.-6-tde.Y'..t Wa.l.l ~e.d ··. Pope Max 7 Wa.l.l aMMJ.> .tnate.d ··· A pouc. e. J.>ta;t_.<_on wM blown up ..· *Un c.orrmiMM.tat de_ po.f_.<_c.~- a ete ~ad e.xploMA · · ·

Another possibility is to construct the verbal expression in the active form but ~here the subject is concerned with the cause of the ACT, such as a bomb. Thi s is possible with to ~ and t o blow up but not with t o a.l.lf.>a.l.l-6-tnat e. or 6a.AAe e. xploJ.>etr. which can be constructed only with a human subject :

A bomb ~e.d the ICM Ptr.e.-6-tde.nt

*A bomb a.l.lf.>a.l.l -6-tnate.d Pope Max 7 ···

A bomb blew up a pof.tc.e. J.>tat.ton ···

*Une. bomb e. a 6a.tv~xploJ.>eA un c.orrmU. J.>alt.ta-t de pouc.e. · ··

The for~ation of a passive form depends first upon the form of the verbal group involved and its syntactic construction. For example, in French, the pass ive requires, roughly, a simple verb constructed wi t h a direct object*. The form and the construction of a verbal group cannot, of course, be predicted from a language-free representation. Furthermore, even if a verbal expression fits the syntactic con­ ditions required for the passive, the formation of the passive depends not only upon the specific ver­ bal group involved but alsc upor. its complements. The following pairs, taken from (Gross 1979), il­ lustrate this point :

Max .tnhabit-6 Manhat-tan *Manhat-tan .t-6 .tnhabde.d by Max

R.tc.h polit.tc.iaY!-6 .tnhabd Manhat-tan Manhat-tan .t-6 .tnhabde.d . by IL.tc.h polit.tuaY!-6

The work done at the LADL by Gross and his col­ leagues** has shown that there is no syntactic "transformation" that can be applied blindly, i.e., without looking at which particular verb and comple­ ments are used. As a consequence, the generation process cannot employ a mechanical procedure such as : when the ACTOR is not specified, choose the same verbal expression as if the ACTOR were speci­ fied, and construct it in a passive form without an agent. Nor is there any way to predict whether a verb can take only a human subject. Therefore, the generat~on process must take into account whether the ACTOR is specified or not in choosing a term to express a RESULT.

The person or object involved in the RESULT may also determine the choice of a term to express the '< F~''LT. Pecall that the result DEAD in a terro­ ri s t crime context i s usuall y expre ssed by to aJ.>J.>a­ .6.6-tnate. if it concerns a f amous target. Furthermore, the term for expressing a RESULT may al so depend on the nu~ber of target s involved. As an illustra­ tion, consider the case of a terrori s t crime invol ­ ving two famous targets. If the results of the ACT for the two targets are identi cal, one may genera­ te the result for each target as if it were alone, and then coordinate the phrases obtained :

*As a consequence, a compound such as 6a.ttte e.xploJ.>etr. cannot be passivised. When the ACTOR is r.ot speci­ fied, the verb e.xploJ.>etr. can be used : Un c.orrmU.­ J.>atr..tat de pou c.e. a e.xploJ.>e .

**See, among others, (Gross 1975), (Boons et al. 1976), (Giry-Schneider 1978), (Negroni-Peyre 1978) (Danlos 1981), (Meunier 1981), (Salkoff 1983).

Pope Max 7 and Id.t Am.tn WeJLe. (MJ.>a.l.l-6-tnate.d + wounded ) yeJ.>teJLday .tn Patr..t-6 .

But, for the case of one target DEAD and the other one WOUNDED, this method would yield

Pope Max 7 Waf.> MJ.>M-6-<-nate.d yuteJrday -<-n Patr..t-6 and Id.t Am-i_n ~ wounded.

which i s only appropriate for a s tory involving two different assassination attempts, for example, in two different places in Paris. Therefore, another formul at i on must be used, such as :

Pope Max 7 and Id.t Am.tn WeJLe t he tatr.ge.t-6 o6 an MJ.>MJ.>.tnCLU.on at-tempt yuteJLday .tn Patr..t-6 : Pope Max 7 Wa.l.l ~e.d and I d.t Am.tn Wa.l.l wounded .

In conclusion, the factors re levant for choo­ sing a term to express a RESULT, i.e., a STATE which concerns an OBJECT because of an ACT done by an ACTOR, are the particular STAl E, OBJECT, ACT and ACTOR involved, in addition to the context in which this RES'ILT occurs. This situation is not peculiar to the choice of a term to express a RESULT. It seems likely that, for each kind of concept, all the semantic roles concerned with that concept may be factors in choosing the best term to express it.

2. Segmentation into sentences

We have seen that the choice of a term to express a RESULT, and more generall y any concept, requires us to look at numerous aspects of the se­ mantic representation. This is also true for the problem of segmentation into sentences.

One factor affecting the generation of a PER­ SON-TERRORISM story is the feature arbitrary/non­ arbitrary target. A non-arbitrary target is a person specifically aimed at, while an arbitrary one has not been individuall y chosen but represents a type of TARGET (e.g. British soldiers, policemen or right wing activists). When the TARGET is not arbitrary, it seems appropriate to produce a text such as :

Pope Max 7 ~ a.l.lf.>a.l.l-6-tnate.d yeJ.>tetr.day .tn Patr..t-6 . A bomb exploded in h.t-6 houJ.>e. .

where a sentence expressing the RESULT is simply followed by a sentence expressing the ACT. However, in cases where the victims are arbitrary, it seems more appropriate to use a form such as :

A bomb exploded i n a pol.tc.e. J.>t at.ton i n Patr..t-6 ye.-6 - t etr.day ~ng 4 pouc.eme.n .

in which the sentence expressing the RESULT occurs in the gerund form after the ACT sentence. But this kind of segmentation for the arbitrary target case does not always produce satisfactory texts. For example, if the ACT takes place in an ambush, the sentences produced by this method seem clumsy :

A bomb e_xn.forie_d ttndc.tt the_ c.M o ~ 4 pohc.e_me_n hil­ Urt(j the_m.

The following kind of text reads better ·

4 pouc.ene_n l"(!}te_ bifle_d whe_n a bomb e_xp.fode_d undeA the_-Uc c.att .

Furthermore, special procedures must he usee when there are very few details about the ACT. For example, if the only information ~iven about the ACT is the fact that it took place in an ambush or that it was a shooting, the generation must then be some­ thing Li ke :

• 4 pouc.e_me_n weAe_ lzille_d -tn an an'bw.,h.
• 4 pouc.e_me_n we.tte_ .6hot a11d !Uile_d .

where the ACT is combined with the RE~ULT in sui ta­ ble formulations.

3. Ou r generation process

Even the few examples mentioned he re show that both phrase selection and sentence segmentation are affected by many semant ic factor s present in the representation. In generating causa l sequences , we claim that the generation process cannot he a pro­ cedure made up of the three fo llowino steps :

1. Generate the ACT
2. Generate the RESULT
3. Combine the phrases obtained.

Such a procedure would suppose that the genera­ ti on process can be based on Loca l deci sions , an assumption suppo rted by (McDonald 19~2 ) amono others. But we have just shown t hat each of these operations requires many different tes t s about information s cat ­ t e red throughout the representation. Thus the gene­ ration process cannot be based on a s imple combina­ ti on of purely Local decisions about how to generat e sub-part s of a semantic rep resentation. It requires that the semantic representation be considered as a whole. Therefore our generator is based on associa­ tions between the semantic rep resentations for spe ­ cific classes of s ituations and templates of senten­ ces expressing what occur s in those si tuations . For example, the renresentati on of a s tory in which the ­ re is a non-arbitrary and non- famous TARGFT ki lled in an explosion, i s associated with the template :

TARGET (pass ive ( lz,{_f_£)) TIME GEO by ACTORWho e_xp.fo­ de_ EXPLOSIVE LOCATION.

While st0 ri es for which there i s no information about the ACT except that it is a SHOOTING are associated with templates such as :

TARGET (passive (.6 hoot and M.f.f )) TIME GEn.

These templates indicate the entire segmentation in­ to sentences of large representation structures with the terms needed to express the main concepts inclu­ ded. They reflect a di s course structure appropriate for the domain of terrori st crimes as reported in newspapers. Thi s accords with the assumption, suppor­ ted by (McKeown 1982) for example, that people have preconceived idea s about the means with which a par­ ticular story can be expressed as well as about the ways in wh ich these means can be integrated to form a (written) text. From these templates, the genera­ tor produces English text in a fairly straightfor­ ward fashion (Danlos 1983).

Our generation process does not operate by use of generation rules of a general nature. Of cour se , our claim is not that there exist no general prin­ cipl es of generation. For example, it is certain that one must not say something which has already been said or which is easily inferable from the pre­ vious phrases. eut, first, the application of such abstract principles varies from one case to anot he r under conditions which elude all formalization in the present state of knowledge. Secondly, the gene­ ration of connected text must rel y upon discour se structures. Such structures are domain and ta sk de­ pendPnt. For example, there is no common point bet ­ ween the discourse s tructures used by McKeown 1982 and ours. Let us add that our discourse structures would not be satisfactory, if they were to be used in a journa l spec ialized in bomb explos ions. Final ­ ly, a generation system must take into account the cons traints presented by the words and their uses. These constraints cannot be predicted from a Lan­ guage-free representation. As an illus tration, com­ pare

1. John a.6.6a.6.6-tnate.d Matty and .6panlze.d Pe.te_tt
2. John o.l.lMM-tnate.d Matty and wounde_d Pe.teA

The clul!'siness of (1) comes from a pragmatic fact, namely the difference of importance attached to an assassination and a spanking. It is possible to pro­ pose a general principle which forbids the coordi ­ nation of concept s which differ too much in impor­ tance. Such a principle would enable us to avoid generating awkward sentences such as (1). But the clumsiness of (2) - being interpreted as a singl e crime with two targets - comes from the uses of the words to a.6.6a.6.6-trto~e. and to wound. Thus here the awkwardness cannot be predicted from a Language-· free representation. The generator must be given the patterns of sentences where these verbs occur, and the situations that correspond to the patterns .

Any generation from a conceptual representation must rely on associations between representations and "templates" of sentences. Such templates <wh ate­ ver the form with which t ~ey arc presented) are clearly necessary, if only to indicate the construc­ tion of a verb, i.e., to relate the semantic roles (ACTOR, TARGET, etc.) to the complements of the verbs (subject, direct object, indirect object with its preposition). However, our use of templates gi­ ves far more information than those which are used only for the purpose of indicating the relation between the semantic roles and the syntactic comple­ ments (cf.(Goldman 1975) for example). This is jus­ tified by the fact that the generation process of a connected text cannot be reduced to a simple combi­ nation of phrases expressing sub-parts of the representation.

ACKNOWLEDGMENTS

I would like to thank Lawrence Birnbaum, Mark Burstein and Roderick ~c�uire for many valuable discussions and suggestions on this paper.