Covid vaccine is against Covid but Oxford vaccine is made at Oxford Semantic Interpretation of Proper Noun Compounds Keshav KolluruyGabriel StanovskyzMausamy

2025-05-06 0 0 549.75KB 14 页 10玖币

侵权投诉

“Covid vaccine is against Covid but Oxford vaccine is made at Oxford!”

Semantic Interpretation of Proper Noun Compounds

Keshav Kolluru†Gabriel Stanovsky‡Mausam†

†Indian Institute of Technology Delhi

keshav.kolluru@gmail.com, mausam@cse.iitd.ac.in

‡The Hebrew University of Jerusalem

gabriel.stanovsky@mail.huji.ac.il

Abstract

Proper noun compounds, e.g., “Covid vac-

cine”, convey information in a succinct man-

ner (a “Covid vaccine” is a “vaccine that immu-

nizes against the Covid disease”). These are

commonly used in short-form domains, such

as news headlines, but are largely ignored in

information-seeking applications. To address

this limitation, we release a new manually an-

notated dataset, PRONCI, consisting of 22.5K

proper noun compounds along with their free-

form semantic interpretations. PRONCI is

60 times larger than prior noun compound

datasets and also includes non-compositional

examples, which have not been previously ex-

plored. We experiment with various neural

models for automatically generating the se-

mantic interpretations from proper noun com-

pounds, ranging from few-shot prompting to

supervised learning, with varying degrees of

knowledge about the constituent nouns. We

ﬁnd that adding targeted knowledge, particu-

larly about the common noun, results in per-

formance gains of upto 2.8%. Finally, we

integrate our model generated interpretations

with an existing Open IE system and observe

an 7.5% increase in yield at a precision of

85%. The dataset and code are available at

https://github.com/dair-iitd/pronci.

1 Introduction

Proper noun compounds (PNCs) (Breban et al.,

2019)

are grammatical constructions where a

proper noun is followed by a common noun, for ex-

ample: Covid vaccines or Buddhist monks. These

often serve as a compact way to convey information

about an already known entity, omitting predicates

that are interpreted by the reader using surround-

ing context, common sense, and world knowledge.

For example, a reader is likely to interpret that

“Buddhist monks” are “religious people who are

buddhists”. In other cases, PNCs are used to iden-

tify speciﬁc entities, and do not provide additional

1also referred to as proper noun modiﬁed compounds.

information. For example, Watergate scandal and

Kawasaki disease do not have any implicit relation

between the proper and common noun as they refer

to a speciﬁc instance of a scandal and a disease.

Table 1provides additional examples.

Thanks to their brevity, PNCs are commonly

used to shorten descriptions in space-constrained

domains, such as news articles headlines (Breban

et al.,2019). However, we ﬁnd that prior work

on compound noun interpretations only considered

cases where the constituents are common nouns

(e.g. baby oil), thus missing all of the information

conveyed in proper noun compounds (Shwartz and

Waterson,2018;Hendrickx et al.,2013).

To address this limitation in current systems, we

begin by deﬁning the task of PNC interpretation

as two subsequent stages (Section 3). The ﬁrst

stage requires identifying whether a given PNC is

compositional or not, while the second stage is the

generation of an interpretation, where applicable.

In Section 4, we present PRONCI, a crowd-

sourced dataset over Wikipedia containing 22.5K

proper noun compounds and their annotated se-

mantic interpretations. Candidates PNCs are found

using syntactic parsing, and are then presented to

crowdworkers who are asked to interpret them. Our

annotation interface marks whether workers needed

to read the full sentence, thus identifying PNCs

whose interpretation relies on context. We will

make the PRONCI dataset publicly available to

spur future research into PNCs.

In Section 5, we develop two approaches for

PNC interpretation: (1) a multi-task neural model

that performs classiﬁcation and sequence gener-

ation in two distinct stages and (2) a text-to-text

approach, using a sequence-to-sequence model for

both classiﬁcation and generation. In addition, we

experiment with different methods for injecting

various sources of world knowledge, which seems

crucial for the task, using external resources like

Wikipedia and WordNet (Fellbaum,2010), that

arXiv:2210.13039v1 [cs.CL] 24 Oct 2022

Type Example Semantic Interpretations

Proper NC Shakespeare biography is a biography about Shakespeare

(Proper-Common)London theatre is a theatre in London ;is a theatre located in London

Concorde airplane [NON-CMP] (Non-Compositional)

Notre-Dame cathedral [NON-CMP] (Non-Compositional)

Common NC nursing job is a job in nursing ﬁeld ;is a job involving nursing

(Common-Common)oil price is price paid for the oil

Table 1: Examples of common and proper noun compounds along with their semantic interpretations (“;” separates

multiple interpretations). [NON-CMP] indicates the absence of implicit relation between the constituent nouns.

give relevant information or deﬁnitions about the

PNCs, that help in improving performance.

For evaluating the generated interpretations, we

propose a combination of classiﬁcation-based met-

ric and generation metrics to properly handle both

the interpretable and non-interpretable cases, re-

spectively (Section 6). Since multiple correct inter-

pretations are possible for a PNC, we use learned

metrics such as BLEURT (Sellam et al.,2020), that

is ﬁnetuned on human-annotated preferences.

Finally, we show that training on PRONCI yields

models that can readily beneﬁt extrinsic down-

stream application in the task of Open Information

Extraction (Banko et al.,2007), thus widely extend-

ing their coverage (Section 8). Our approach ﬁrst

automatically extracts PNC interpretations using

our models, then introduces it explicitly back into

an Open IE extraction using a sequence to sequence

model, thus giving an interpretation-integrated ex-

traction. We then apply a high precision rule to

generate new relations which leads to a 7.5% in-

crease in yield at an estimated precision of 85% on

the added extractions, when compared to extrac-

tions generated from the original sentences them-

selves. A major advantage of this approach is that

it is agnostic to the Open IE system being used. To

conclude, our main contributions are:

We introduce the PRONCI dataset, contain-

ing interpretation for 22.5K proper noun com-

pounds and their semantic interpretations.

We develop multi-task and generation based

neural baselines that can leverage external

knowledge for achieving higher performance.

We design metrics for evaluating the quality

of generated semantic interpretations.

We demonstrate the usefulness of the gener-

ated interpretations in a downstream applica-

tion by using them to augment the expressivity

of Open IE systems.

2 Related Work

Noun compounds are commonly used in English

language, constituting 3.9% of the tokens in the

Reuters corpus (Baldwin and Tanaka,2004). They

can be arbitrary length phrases, such as split air

conditioner, but most prior work on interpreting

noun compounds has primarily looked at two word

noun compounds of the type noun-noun, where

both are common nouns. To the best of our knowl-

edge challenges in interpretation where the ﬁrst

word is a proper noun (i.e., proper noun com-

pounds) have not been addressed, although their

functional analysis and prevalence in certain do-

mains have been studied in linguistics (Rosen-

bach,2007;Alexiadou,2019;Breban et al.,2019).

We brieﬂy summarise the various types of noun-

compound interpretations in literature and discuss

their uses in applications.

Types of interpretation:

Various types of in-

terpretations for noun compounds have been ex-

plored, covering classiﬁcation, ranking and gen-

eration. Prior literature has frequently posed the

interpretation as a

classiﬁcation

task, where the

classes can belong to abstract labels (Fares,2016),

semantic frame elements (Ponkiya et al.,2018)

or prepositions (Lauer,1995) However, none of

these schemes can cover all range of possible noun

compounds, thus limiting their expressivity and

coverage. SemEval 2010 Task 9 (Butnariu et al.,

2009) annotates human preferences for a set of

25-30 templatized paraphrases for each of the 250

training and 300 testing noun compounds. The task

is framed as producing an accurate score for each

paraphrase that

ranks

them in the correct order.

SemEval 2013 Task 4 (Hendrickx et al.,2013) re-

leased a dataset of noun compounds and annotated

free paraphrases for each compound. Participating

models were evaluated by matching and scoring

the generated predictions with the gold set.

Ponkiya et al. (2020) is the current state of art

which poses the problem as generation of masked

tokens using a pretrained T5 model (Raffel et al.,

2020) to get free paraphrase interpretations in a

completely unsupervised manner. This leads to bet-

ter performance than techniques that use the avail-

able training data. However, with the PRONCI

dataset, we do ﬁnd that supervised models do out-

perform zero-shot models due to the scale.

Applications:

Noun compound interpretations

have been helpful in translation of noun compounds

by either using a one-to-one mapping of interpreted

prepositions (Paul et al.,2010) or using recursive

translation patterns (Balyan and Chatterjee,2015).

In Question Answering systems, they have been

used for disambiguating different types of noun-

noun compounds in passage analysis (Ahn et al.,

2005). They have also been useful for normalizing

text that can help textual entailment (Nakov,2013)

and as auxiliary semantic annotation modules to

improve parsing (Tratz,2011). In this work, we

show their use in the task of Open IE.

Open Information Extraction

(Open IE)

(Banko et al.,2007;Mausam,2016;Kolluru et al.,

2020b) involves extracting a set of tuples from the

sentence where each ﬁeld of the tuple contains

phrases from the sentence itself. This makes it

ontology-agnostic and allows it to be used for

creation of domain agnostic Open Knowledge

Bases (Broscheit et al.,2020;Vashishth et al.,

2018;Gupta et al.,2019). The relations are often

verb-based (Fader et al.,2011) but can also be

noun-mediated (Pal and Mausam,2016) or involve

implicit information (Soderland et al.,2015).

Fader et al. (2011) relied on high precision rules

to extract a wide variety of verb-mediated relations.

Soderland et al. (2015) uses dependency paths

for generating high precision extractions based on

three implicit relations, has job title,has city and

has nationality.Pal and Mausam (2016) consid-

ers noun mediated relations that can be extracted

from compound noun phrases while dealing with

challenges involved with denonyms and compound

relational nouns. However, none of them consider

implicit relations present in noun compounds.

Moreover, recent state of art Open IE systems

like OpenIE6 (Kolluru et al.,2020a) and Gen2OIE

(Kolluru et al.,2022) rely on bootstrapped exam-

ples (generated using OpenIE4 (Pal and Mausam,

2016;Christensen et al.,2011)) for training. There-

Task Instructions

1. Your goal is to describe the relation between the two

words by ﬁlling in the blanks.

2. You can write up to ﬁve words (or less!) 3. The

resulting relation should form a valid English sentence

(see below for an example).

4. You can consult an example sentence as additional

context, but the relation you write should be inferred only

from the two words, and not by additional information.

5. If it is a name, entity, location or if you can’t describe

the relation between the words, please leave the relation

blank.

Examples

1. Coke Spokesman is a worker of Coke.

2. Leake government is located in Leake.

3. Capitol Hill

Pitfalls

1. Coke Spokesman employment Coke.

The relation should form a valid sentence.

2. Leake government has a failed goverment.

The relation should be inferred by the words themselves

and not by additional context.

Table 2: Instructions for the task along with examples

and common pitfalls that are provided to the human

workers from AMT for constructing PRONCI dataset.

fore they only generate extractions that contain

phrases from the text and miss the cases where the

content words are implicit. OpenIE6 (Kolluru et al.,

2020a) adopts a pipeline approach to integrate con-

junction splitting into Open IE outputs, where co-

ordination analysis and sentence splitting is per-

formed as a preprocessing step, and the Open IE

extractions are generated from the split sentences

which are then merged.

3 Problem Deﬁnition

Interpretations of noun compounds are meant to

expose the expressed implicit relation. Free-form

paraphrases as interpretations provide ﬂexibility for

expressing relations implied in noun compounds,

overcoming the limitations associated with choos-

ing from a ﬁxed set of classes or templates at the

cost of a possibly non-consolidated representation,

i.e., where similar-meaning noun compounds are

represented differently. Hence, we deﬁne semantic

interpretation of a PNC as a free-form paraphrase

that exposes the implicit relation between the con-

stituent nouns, if any relation exists, else identify it

as non-compositional ([NON-CMP]).

SemInt(pnc) = (Paraphrase,if reln. exists

[NON-CMP],if reln. absent

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：

CovidvaccineisagainstCovidbutOxfordvaccineismadeatOxford!SemanticInterpretationofProperNounCompoundsKeshavKolluruyGabrielStanovskyzMausamyyIndianInstituteofTechnologyDelhikeshav.kolluru@gmail.com,mausam@cse.iitd.ac.inzTheHebrewUniversityofJerusalemgabriel.stanovsky@mail.huji.ac.ilAbstractPropernou...

展开>> 收起<<

Covid vaccine is against Covid but Oxford vaccine is made at Oxford Semantic Interpretation of Proper Noun Compounds Keshav KolluruyGabriel StanovskyzMausamy.pdf

共14页,预览3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！

Covid vaccine is against Covid but Oxford vaccine is made at Oxford Semantic Interpretation of Proper Noun Compounds Keshav KolluruyGabriel StanovskyzMausamy

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: