Generating thematic choices for multilingual text generation

Abstract
The present paper addresses two important issues
for generating pragmatically adequate administrative forms in the context of GIST, a multilingual
text generation system for the automatic production of administrative forms in English, German
and Italian.. First, it investigates the influence of
socio-cultural factors (such as the social distance
and the social role) on the mood structures preferred by each language when requesting action from
users. Given a specific mood choice, the influence
of pragmatic constraints (such as weight, identifiability, and topicality) is also explored to account
for the different thematic realizations found in the
corpus. Second, it addresses issues of computational specification for a multilingual generation architecture by using mechanisms exploited in a
number of diverse multilingual generation projects.
1 Introduction
The GIST project addresses the development of a
multilingual generation system for the automatic
production of administrative forms in three different languages: English, German and Italian. In
this context, it is part of our task to provide a
contrastive study and computational specification
of the preferred thematic realization options which
each language selects to express the information
drawn from a common application domain dealing
mainly with pension and family benefits, unemployment and disability allowances.
Careful analysis of the forms has shown that thematic realizations occur in different mood contexts which each language prefers to express the
same speech function. For example, when requesting action from users, English forms prefer to
thematize the Predicator element (unmarked theme in imperative clauses) while Italian and German prefer to thematize participants functioning
as Affected or Agent in the transitivity structure of
the clause. Example 1, taken from a bilingual
Italian/English form, illustrates this contrast (Themes are underlined):
Example 1:
Italian: Il presente formulario, debitamente compilato, deve essere presentato o inviato al più
presto possibile alla Sede provinciale dell’ Instituto Nazionale della Previdenza Sociale (I.N.P.S)
competente per territorio.
English: Send or take the completed form to your
I.N.P.S local office as soon as possible.
These unmarked realizations, very frequent in our
multilingual corpus, are a clear indication that
thematic realization options can be partially controlled by the preferred mood selections which
each language chooses to express the same speech
function. For example, in English, unmarked theme realizations will be Subject in the declarative
clauses, wh-element in wh-interrogative clauses,
and the Predicator in imperative clauses. These
realizations are language-dependent, i.e., what is
unmarked theme in English in a given mood context is not necessarily the same in Italian and
German. However, as will be shown below, thematic realization is not fully determined by mood,
but is the result of the interplay of different factors
which impose a specific thematic structure on the
message.
Being aware of these facts, we have opted for an
analysis methodology which will allow us to contrast the language-dependent thematic realizations
occurring in specific mood contexts under language-specific interpersonal selections. We have concentrated on the thematic realizations which each
language prefers when expressing requests, a
speech act which predominates in administrative
forms.
The paper is organized as follows. First, we describe the multilingual corpus used for the study
and explain the methodology used for its analysis.
Second, we suggest how specific socio-cultural
selections (such as social distance and social role
options) preselect the mood structures where thematic realizations occur. Third, we present the
language-specific thematic realizations found in
requests, and present possible motivating factors
which contribute to control them. These are then
captured as inquiries which control the different
options in each of the proposed language-specific
thematic networks.
183
2 Corpus Analysis
The GIST corpus of English forms contains approximately 25,000 words and comprises eight forms,
all of them issued by the Department of Social
Security (DSS). The Italian/German corpus contains 36,248 words and comprises 12 bilingual
(Italian/German) and 2 bilingual (Italian/English)
forms. The documents are produced in Italian by
INPS (Instituto Nazionale Previdenza Sociale)
and adapted to include a German translation at
PAB (Provincia Autonoma di Bolzano), the local
government of the province of Bolzano.
The analysis involved addressing the following
issues:
1. Determining the interpersonal context in
which thematic realization variants occur.
This involved identifying the social distance
and the social role options which characterize
each culture.
2. Determining the preferred mood structures
which each language uses to realize requests;
dividing them into direct and indirect requests
depending on the selected mood structure, and
establishing the probability of occurrence of
one type or another in each language under
study.
3. Determining the thematic realization forms
selected by each language in the mood structures identified in 2.
4. Determining the pragmatic factors which constrain the selection of a given thematic realization occurring in a specific mood context.
As explained above, this paper concentrates on a
specific class of the basic speech function of demanding goods-and-services, namely, on requests, and the preferred thematic realization options which each language chooses in
administrative forms. However, the proposed methodology has also being employed for the comparison of preferred thematic realizations of other
speech functions which frequently appear in administrative forms.
3 Interpersonal options in administrative
forms
Determining the interpersonal context in which
thematic realizations occur involved identifying
the social distance and the social role options
which characterize each culture in the administrative setting where the forms are produced. The
English forms have been written according to the
DDU (Document Design Unit) guidelines and are
the result of extensive user testing to achieve a
warm and non-intimidating tone. These forms
tend to minimize the social distance and to level
out the social roles between the administration and
the users, and this is achieved through a number
of devices, among them by the direct expression
of requests by means of the imperative mood. In
contrast, both the Italian/German forms, produced
in a different socio-cultural setting, where the
social distance is maximal and the social roles are
hierarchic, prefer to use indirect strategies for the
realization of requests, thus opting for the declarative mood. Figure 1 illustrates these languagespecific preferences
Figure 1: Language-specific preferences in the realization of requests in the GIST corpus
While the English forms prefer the imperative
mood in 88.2 % of the cases and the declarative
only in 11.8%, the Italian/German forms opt for
the declarative in 89.3% and 92.3% of the cases,
respectively, selecting the imperative in only
10.7% and 7.7% of the cases. The difference in
these proportionalities is statistically significant
(p<0.001).
4 From mood structures to thematic
realizations
As we explained above, thematic realizations always occur in specific mood contexts, which can
be direct or indirect realizations of a given speech
function. Table 1 below illustrates the range of
thematic realizations selected by each language
and their frequency distribution when requesting
action from users.
184
As explained above, English prefers the direct
expression of requests by means of the imperative
mood (88.2%). This fact is reflected in the preferred thematic realization options selected: typically, English administrative forms select the interpersonal modal “please” followed by the
Predicator element in an imperative clause type.
This option occurs in 209 clauses from a total of
486, thus constituting the 43 % of all direct thematic realizations in the English corpus. Indirect
realizations are much more unfrequent (11.8%)
than direct ones, with a predominance of personal
forms; among them, the second person pronoun
“you” functioning as Subject in declarative clauses is the most widely used when referring to the
form-filler, and the first person plural pronoun
“we” when referring to the source of the form.
By contrast, German and Italian prefer indirect
strategies for the expression of requests due to
culture-specific requirements where the social distance and the social role relationships between
the administration and the users is maximal and
hierarchic. This fact is reflected in their preferred
thematic realizations. Leaving aside the circumstantial elements, the conditional clauses, and the
realizations by means of an infinitive – with a high
number of occurrences in both corpora for reasons
which we will explain below, both languages tend
to thematize the Affected (Patient) element as
Subject in independent declarative clauses, thus
creating a passive construction. It is well-known
that passive constructions contribute to create a
distant, impersonal impression on the reader by
not encoding agents as subjects, thus obscuring
this ‘natural’ expected mapping. The influence of
these interpersonal factors is even more apparent
in those (highly frequent) cases where the thematized Affected element is a nominalization, where
the action is hidden out as a noun and there is no
mention of the participant involved. Also, when
the Subject is the Agent, all thematic realizations
are expressed by Nominal Groups, rather than by
personal pronouns referring to the addressee
which contributes to a an impersonal and distant
tone.
The high percentage of infinitives both in Italian
(some of them preceded by “da”) and their translation into German (most of them preceded by the
Direct Object) is, again, an indication of the tendency to use impersonal non-finite forms at the
Table 1: Distribution of thematic realizations of direct and indirect requests in the three languages
Italian German English
Indirect Direct Indirect Direct Indirect Direct
S / Agent 19 Predicator 6 S / Agent 20 Predicator 6 S / You 28 Predicator 124
S/ Aff. 51 Cond.Cl. 10 S/ Aff 52 Circumst. 10 S / We 5 Please+P 209
Od/Aff 3 Circumst. 12 Od/Aff 2 Cond. Cl 3 S / NG 6 Do not+P 19
Circumst. 65 Od/Aff 3 Circumst. 53 Inter. T 1 S / They 1 Cond. Cl 89
Finite 24 Temp. Cl 2 S / Es 14 Od/ Aff. 1 Temp. Cl 16
S/ Carrier 3 Cond. Cl 29 Cond.Cl 18 Locat Cl 2
Temp.Cl 5 Temp. Cl 3 Temp. Cl 5 Final Cl 2
Cond.Cl 32 Finite 3 Conc. Cl 1 Conc. Cl 1
Final Cl 1 Textual 7 Circumst. 24
Textual T 5 NF 3
Neg+ Fin 1 Neg.+NF 1
Si + Fin 1 Neg +Aj. 1
Infinitive 48 Neg+ Infi 1
Da+ Infi 16 Od + Infi. 41
Neg+Infi. 3 Infinitive 13
TOTAL 278 TOTAL 33 TOTAL 243 TOTAL 20 TOTAL 65 TOTAL 486
% 89.3 % 10.7 % 92.3 % 7.7 % 11.8 % 88.2
S= Subject; Od= Direct Object; P= Predicator; Aff.= Affected; Circumst.=Circumstance; Temp. Cl= Temp.
Clause; Cond.Cl= Conditional Clause; Locat.Cl= Locative Clause; Final Cl= Final Clause; Conc. Cl=
Concessive Clause; Textual T= Textual Theme; Inter. T= Interpersonal Theme; Fin.= Finite; NF=
NonFinite; Infi= Infinitive.
185
beginning of sections where the writer expects an
immediate action from the reader and does not
need to provide him with details about the filling
of the form.
Other significant realizations captured on table 1
can be explained by the influence of varied factors,
among which we have isolated two: weight and
topicality. These two factors combined explain the
high percentages of conditional clauses and circumstantial elements in thematic position.
Topicality is the tendency to present in first initial
position those elements which are ‘topical’ in the
sense of shared by the speaker and the hearer in
the discourse context. Careful analysis of the initial conditionals in the three corpora has shown
that, in general, they function as a premise, as
material that the addresse is expected to take as
given (see Lavid & Taboada, 1994). The high
percentage of circumstantial elements, most of
which express condition realized by prepositional
groups in Italian and in German, is due to the same
factor. Most of these circumstantial elements pick
up information which has been mentioned before
in the discourse, thus preserving topic continuity
and contributing to smooth information flow.
Weight is the principle according to which long,
“heavy” constituents tend to come late in the clause (Leech, 1983: 65). In the case of Italian, this
results in the thematization of the Finite element
of the Verbal Group, while the long Subject is
delayed to the end of the clause. This factor accounts for 24 occurrences in the Italian forms
(11.3 %). The German translations use the rather
unfrequent construction introduced by the particle
“Es” to imitate the Italian structure in 7.4 % per
cent of the cases.
5 Representation
As illustrated in the previous sections, the combination of pragmatic factors such as weight and
topicality, on the one hand, and the social distance
and social role selections preferred by each culture, on the other, control language-dependent thematic realizations in our multilingual corpus.
Using the chooser-and-inquiry framework developed by Mann and others (Mann, 1983) and explained elsewhere (Matthiessen & Bateman,
1991), we propose to represent these factors as
inquiries (semantic choices) associated with features in each of the language-specific thematic
networks1
which we present in Figure 2, next
page.
The English theme markedness options are the
ones used in NIGEL, the systemic-functional
grammar component implemented in the PENMAN System (Matthiessen, 1993). The German
options are an adaptation of the account presented
by Steiner and Ramm (forthcoming). The Italian
options are the result of our own ongoing account
for this language. Interestingly, as the network
indicates, marked thematic options in English are
no longer dependent on mood but on transitivity.
In Italian and in German, marked thematic options
are not restricted to a limited number of transitivity
roles which are fronted in the structure of the
clause, but include a wide variety of clause ordering combinations which mark its whole structure
as marked. As an example of how choices could
be controlled by a specification of different factors
from the communicative situation, we propose the
following inquiries which control Italian unmarked thematic choices in the declarative mood context :
Choice: UNMARKED as either SUBJECT or
FINITE or CIRCUMSTANCE
Inquiries (-id): Is the Subject identifiable, and is there
more than one participant in the process?
Then: conflate THEME with Subject
Inquiry (-id): Is the Subject unidentifiable, or too long, or
is there only one participant in the process?
Then: conflate THEME with Finite
Inquiry (-id) Is the current topic a time or space
description of the current situation?
conflate THEME with CIRCUMSTANCE
Inquiry (-id) Is the current topic a participant recoverable
from the co(n)textual situation?
conflate THEME with CLITIC element
We hope this to be sufficient as an illustration of
how different contextual factors represented as
inquiries can constrain thematic choices in the
markedness options of each of the language-specific thematic possibilities. While these factors are
common to the three languages, their interaction
with thematic realizations is language-specific.
5 Summary and Conclusion
This paper has employed an intensive corpus analysis to identify the different thematic realizations
which three different languages (German, Italian,
English) select when requesting action from users
in administrative forms. The analysis has shown
that thematic realizations are partially determined
by the preferred mood structures which each
language selects on the basis of culturally-conditioned interpersonal factors such as the social distance and the social role between the interactants.
For the purposes of multilingual text generation in
the application context of the GIST system, we
have proposed to represent the language-specific
mood structures which realize direct and indirect
requests as options within a system with percentages annotating the preferred selections. These
186
mood strutures constitute the context where thematic realizations can be further specified by representing motivating factors from the communicative context as inquiries which control choices for
each of the three language-specific thematic networks.
This paper, therefore, has presented a useful methodology for the generation of thematic choices
in requests which can be fruitfully applied to other
speech acts, such as questions and statements,
which are also common in administrative forms.
In fact, the current text structurer of the GIST
system is using the proposed specification for the
multilingual generation of mood structures and
thematic choices in these three types of speech
acts, thus validating the applicability of this method in a practical computational scenario.
Figure 2: Theme markedness in English, German and Italian.
187
Note
1. The use of the system network and its translation into (typed) feature structures and of the
inquiry interface is well understood in computational systemic-functional accounts, being a
common representational means in PENMAN-style generators used in a number of
diverse multilingual generation projects. These include: DRAFTER (British EPSRC Project J19221) (Paris et al., 1995), KOMET
(GMD/IPSI) (Bateman and Teich, forthcoming), TECHDOC (Rösner and Stede,
1991), DANDELION (EP6665), and GIST
(European LRE Project 062-009) (Not and
Stock, 1994).
References
Halliday, M.A.K. (1976). System and Function in
Language. Oxford University Press, London.
edited by G.R. Kress.
Halliday, M.A.K. (1985). An Introduction to
Functional Grammar. Edward Arnold, London.
Lavid, J. and Taboada, M.T. (1994) Specification
of Evaluation Criteria for Text Quality. GIST
Internal Report INT-12. Universidad Complutense de Madrid, September, 1994.
Leech, G.N. (1983) Principles of Pragmatics.
Longman, London.
Matthiessen, C. M.I.M. (1992) Lexicogrammatical cartography: English systems. Technical
report, University of Sydney, Linguistics Department, 1992. Ongoing expanding draft.
Matthiessen, C. M.I.M. and Bateman, J. A. (1991).
Text generation and systemic-functional linguistics: experiences from English and Japanese. Frances Pinter Publishers and St.
Martin’s Press, London and New York.
Mann, W. C. (1983). An overview of the the Penman text generation system. Technical Report
ISI/RR-83-114, USC/Information Sciences
Institute, Marina del Rey, CA, 1983.
Steiner, E. and Ramm, W. (forthcoming). On Theme as a grammatical notion for German.
Functions of Language.