Towards Pragmatic Production Strategies for Natural Language Generation Tasks Mario Giulianelli

2025-05-06 0 0 208.07KB 7 页 10玖币

侵权投诉

Towards Pragmatic Production Strategies

for Natural Language Generation Tasks

Mario Giulianelli

Institute for Logic, Language and Computation

University of Amsterdam

m.giulianelli@uva.nl

Abstract

This position paper proposes a conceptual

framework for the design of Natural Lan-

guage Generation (NLG) systems that follow

efﬁcient and effective production strategies

in order to achieve complex communicative

goals. In this general framework, efﬁciency is

characterised as the parsimonious regulation

of production and comprehension costs while

effectiveness is measured with respect to

task-oriented and contextually grounded

communicative goals. We provide concrete

suggestions for the estimation of goals, costs,

and utility via modern statistical methods,

demonstrating applications of our framework

to the classic pragmatic task of visually

grounded referential games and to abstractive

text summarisation, two popular generation

tasks with real-world applications. In sum, we

advocate for the development of NLG systems

that learn to make pragmatic production

decisions from experience, by reasoning about

goals, costs, and utility in a human-like way.

1 Introduction

Novelists choose the right words to keep readers

engaged and enthused, good journalists can con-

vey facts clearly and convincingly, while poets

may want to surprise the reader. Teachers adapt

their explanations to the level of their students,

and the language of parents changes with the pro-

ﬁciency of their children, with the same objects

described ﬁrst using simpliﬁed funny expressions

(‘moo moo’) and then more informative and dis-

criminative names (‘cow’,‘calf’). Using language

to communicate successfully requires effort. On

the side of the language producer, it is ﬁrst of all ef-

fortful to come up with words that truthfully corre-

spond to one’s communicative intent. Then, words

must be actually produced, e.g. said out loud or

typed on a keyboard. At the same time, the pro-

ducer has to take into consideration whether the

comprehender—for whom, too, linguistic commu-

nication is costly—will be able to infer the original

intent. Comprehenders make efforts to pay atten-

tion to the utterance they are being addressed with,

to interpret it, and to infer their interlocutor’s com-

municative intent. Fortunately, these efforts are

often not in vain. They allow people to exchange

knowledge, ideas, plans, and to achieve goals.

This paper introduces a conceptual framework

for Natural Language Generation (NLG) in vari-

ably complex communicative scenarios, which re-

lies on three main notions:

communicative goals

production and comprehension

costs

, and

utility

We deﬁne these notions formally and then, in two

case studies, we provide suggestions for their op-

erationalisation in classic NLG tasks. In sum, we

model humans as decision makers striving for efﬁ-

cient and effective communication, and argue that

human-like linguistic behaviour emerges as a result

of reasoning about goals, costs, and utility. Learn-

ing to navigate the complex decision space deﬁned

by these notions is still an open problem: we dis-

cuss possible promising directions.

2 Doing Things with Words

Communication always comes with a goal: speak-

ers use words to change the state of the world. In

this section, we give a characterisation of commu-

nicative goals, discuss the types of effort (or costs)

necessary to achieve goals, and describe the re-

wards associated with successful communication.

2.1 Communicative Goal

What do speakers do with words? The communica-

tive goal (or communicative intent) of a speaker

can be formulated as a function of the current state

of the world w∈W:

Gs:W→W, w 7→ w∗(1)

where

w∗

is the intended future state of the world.

Speaker

and audience

are included in

they can be both conceptualised, and there is ev-

idence that they are processed (Brown-Schmidt

arXiv:2210.12828v1 [cs.CL] 23 Oct 2022

et al.,2015), as parts of the state of the world. For

communication to be successful, the audience must

be able to reconstruct the original communicative

goal: their decoded transformation of the world,

Da:W→W

, must be such that

Da(w)≈Gs(w)

Communicative goals shape and constrain a

speaker’s production choices: different utterance

types typically correspond to different goals. The

communicative goal of a referring utterance (“The

black and white cat”), for example, is a state of

the world where the audience is able to identify an

entity in context. The transformation

required

to achieve

w∗

is a change of attention by the

audience. Statements (“The Sun is a star”) are

typically used when the purpose of an interaction is

pure information transmission—e.g., when giving

a scientiﬁc talk. In this case, the communicative

goal is a state of the world in which the audience

holds new beliefs, the ones intended by the speaker.

is a transformation of the belief state of the

audience, and the communicative goal is achieved

when

Da(w)≈Gs(w)= w∗

. All utterance types—

e.g., questions, directives, and performatives—can

be seen as strategies to achieve communicative

goals. The same utterance type, and even the same

utterance, can fulﬁl different goals: a blatantly

false statement (“It never rains in Amsterdam”)

can be used for comedic effect rather than for

conveying facts. For simplicity, in the rest of this

paper, we describe utterances as having a single

communicative goal. Often, however, different

goals are associated with the same utterance at

the same time: a teacher can use a question (“Are

you sure this is the right answer?”) to inform

their student that their answer is incorrect, while

showing a positive attitude towards them—thereby

striving for both epistemic and social utility. Our

framework naturally generalises over such cases;

when multiple communicative goals are involved,

states of the world can be designed accordingly.2

2.2 Production Costs

Given the current and the intended future state

of the world,

and

w∗=Gs(w)

, a speaker en-

codes the communicative goal

Gs(w)

into a mental

representation of the intended state of the world:

We sometimes refer to

Da(w)

and

Gs(w)

and

as in our formulation these functions are always applied to the

current state of the world w.

To account for epistemic and social utility, for example,

states of the world can be deﬁned to include the audience

belief state as well as their emotional state.

Es(Gs(w))= e

. To use a slightly different vocabu-

lary, this is the speaker’s conception of the intended

environment state. The speaker then realises

an utterance

which is presented to the audience:

Rs(e) = r

. Two types of cost are associated with

the encoding and realisation processes. Because

the encoding process is inevitably lossy—mental

representations are compressed representations of

the real state of the world—the speaker makes an

effort to reduce information loss; we refer to this

as the

encoding cost CE

. The cost associated with

executing a bit of behaviour

meant to be per-

ceived by the audience (e.g., speaking, writing, or

typing) is the

realisation cost CR

. Both costs af-

fect the decision making process of speakers. In

addition, the speaker is inﬂuenced by the expected

comprehension costs of the audience.

2.3 Comprehension Costs

The speaker’s communicative goal

Gs(w)

is not

observable by comprehenders. Given a state of the

world

and the speaker’s behaviour

, compre-

henders process

into a reconstruction of the orig-

inal mental representation,

Pa(r) = e0≈e

, from

which they decode the speaker’s communicative

goal:

Da(e0)= w0≈Gs(w)

. Two types of cost are

associated with the comprehension of an utterance.

Speaker and comprehender are different individu-

als and therefore have different ways of encoding

communicative goals into messages (Connell and

Lynott,2014). In the absence of a perfect model of

the speaker’s encoding mechanism, reconstructing

is a lossy and effortful process; we denote the

corresponding cost as

processing cost

. The

second cost results from interpreting

in context—

i.e., decoding from

the state of the world intended

by the speaker. In other words, this is the effort

required to ground the message in the environment.

We refer to it as the

decoding cost CD

. It is impor-

tant to note that although processing and decoding

costs are on the side of comprehenders, speakers

estimate them and take them into account when

making production decisions.

2.4 Utility

In what ways is the decision making process

of speakers affected by these costs? Speakers

are thought to be driven by efﬁciency concerns

(Zipf,1949;Jaeger and Tily,2011): they strive

to minimise the collaborative effort required

to achieve their communicative goals (Clark

and Wilkes-Gibbs,1986;Clark and Schaefer,

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

TowardsPragmaticProductionStrategiesforNaturalLanguageGenerationTasksMarioGiulianelliInstituteforLogic,LanguageandComputationUniversityofAmsterdamm.giulianelli@uva.nlAbstractThispositionpaperproposesaconceptualframeworkforthedesignofNaturalLan-guageGeneration(NLG)systemsthatfollowefcientandeffectiv...

展开>> 收起<<

Towards Pragmatic Production Strategies for Natural Language Generation Tasks Mario Giulianelli.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Towards Pragmatic Production Strategies for Natural Language Generation Tasks Mario Giulianelli

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: