Focusing on Context is NICE Improving Overshadowed Entity Disambiguation Vera Provatorova1 Simone Tedeschi23 Svitlana Vakulenko4

2025-04-27 0 0 510.79KB 10 页 10玖币

侵权投诉

Focusing on Context is NICE:

Improving Overshadowed Entity Disambiguation

Vera Provatorova1, Simone Tedeschi 2,3, Svitlana Vakulenko4

Roberto Navigli2and Evangelos Kanoulas1

1University of Amsterdam, 2Sapienza University of Rome

3Babelscape, Italy, 4Amazon Alexa AI

{v.provatorova, e.kanoulas}@uva.nl

{tedeschi, navigli}@diag.uniroma1.it

svitlana.vakulenko@gmail.com

Abstract

Entity disambiguation (ED) is the task of map-

ping an ambiguous entity mention to the corre-

sponding entry in a structured knowledge base.

Previous research showed that entity overshad-

owing is a signiﬁcant challenge for existing

ED models: when presented with an ambigu-

ous entity mention, the models are much more

likely to rank a more frequent yet less contextu-

ally relevant entity at the top. Here, we present

NICE, an iterative approach that uses entity

type information to leverage context and avoid

over-relying on the frequency-based prior. Our

experiments show that NICE achieves the best

performance results on the overshadowed enti-

ties while still performing competitively on the

frequent entities.

1 Introduction

Entity disambiguation (ED) is the task of mapping

an ambiguous entity mention to the corresponding

entry in a structured knowledge base. Despite ED

being a well-known task, recent work has shown

that the existing methods are still far from achiev-

ing human-level performance: in particular, the

case of entity overshadowing remains a big chal-

lenge. An entity

overshadows

if the two enti-

ties share the same surface form

, and

is more

common than

, i.e., has a higher prior probability

to be linked to

(Provatorova et al.,2021). For ex-

ample, when given the sentence “

Michael Jordan

published a paper on machine learning” and the

task of linking Michael Jordan either to the bas-

ketball player (a frequent entity) or to the scientist

(an overshadowed entity), a human will correctly

choose the latter, while a typical model is likely

to ignore the context and give the wrong yet more

popular answer due to over-relying on prior prob-

ability. Figure 1shows another example of entity

overshadowing: the entity Rome (TV series) is

overshadowed by Rome (city).

According to previous research, current ED sys-

tems are prone to over-relying on prior probability

Rome_(city) output (REL, WAT)

Rome_(TV_show) correct answer

Hinds played Caesar in Rome .

Figure 1: An example of entity overshadowing: two

popular ED systems (REL and WAT) predict the most

frequent entity (Rome the city) instead of the correct

answer in this context (Rome the TV series).

instead of focusing on context information, which

causes them to underperform on overshadowed en-

tities. In benchmarking experiments performed

by Provatorova et al. (2021), all ED systems under

evaluation appeared to have a large performance

gap between Top and Shadow subsets of the Shad-

owLink dataset, where Top contains most frequent

entities and Shadow contains their overshadowed

counterparts. The results of a human evaluation

experiment in the same study indicate that the chal-

lenge of entity overshadowing is unique to auto-

mated ED methods: human participants achieved

equally good results at disambiguating entities sam-

pled from Top and Shadow. These ﬁndings call for

further research in the ﬁeld of ED, with the goal

of building a method that outperforms existing sys-

tems on overshadowed entities while still achieving

competitive results on standard datasets.

Interestingly, the best results on Shadow in

the benchmarking experiments were achieved by

AIDA (Hoffart et al.,2011), an unsupervised col-

lective entity disambiguation method: while still

affected by overshadowing, this method appeared

to be the best at capturing the context information

in comparison with modern neural appproaches.

Speciﬁcally, AIDA relies on two main sources of

context information: semantic similarity between

an entity and its context and graph-based related-

ness between the candidate entities of different

arXiv:2210.06164v1 [cs.CL] 12 Oct 2022

mentions. Our study continues this line of work,

incorporating modern neural methods to measure

semantic similarity and adding novel heuristics to

improve candidate ﬁltering and collective disam-

biguation.

We introduce NICE (NER

-enhanced Iterative

Combination of Entities), a combined entity dis-

ambiguation algorithm designed to tackle the chal-

lenge of entity overshadowing by focusing on three

aspects of context-based information: entity types,

entity-context similarity and entity coherence. The

pipeline of NICE includes a NER-enhanced candi-

date ﬁltering module designed to improve robust-

ness on overshadowed entities (Section 2.1), a pre-

scoring module that calculates semantic similarity

between a candidate entity and a mention in con-

text, and an unsupervised iterative disambiguation

algorithm that maximises entity coherence (Section

2.3), combining the relatedness scores between can-

didate entities with the scores of the semantic simi-

larity module (Sections 2.3-2.4). To the best of our

knowledge, our study is the ﬁrst attempt to build an

entity disambiguation method designed speciﬁcally

to tackle the problem of entity overshadowing.

We perform a systematic evaluation of the NICE

method, and use our experimental results to answer

the following research questions:

RQ1:

Does focusing on context information im-

prove ED performance on overshadowed entities?

RQ2:

Does focusing on context information in-

stead of relying on mention-entity priors in ED al-

low to maintain competitive performance on more

frequent entities?

RQ3:

In what ways do the different aspects of

context information contribute to ED performance

on overshadowed entities?

We hope that our work will encourage further

studies concerning overshadowed entities. The

source code of the NICE method is provided as

supplementary material and will be released pub-

licly upon acceptance.

2 The NICE method

Our method is based on the assumption that the

main challenge in disambiguating overshadowed

entities stems from over-relying on entity common-

ness, and therefore switching the focus to the con-

text (entity relatedness) can improve the perfor-

mance. We consider three main ways of extracting

1Named Entity Recognition (Yadav and Bethard,2018)

information from the context: (1) using mention-

entity similarity to predict entity types and improve

candidate ﬁltering, (2) using word embeddings en-

hanced with entity types to measure semantic sim-

ilarity between an entity and its context, and (3)

using entity-entity similarity to make sure that the

entity disambiguating decisions within one docu-

ment are coherent (collective disambiguation).

2.1 Candidate ﬁltering

Adding the step of ﬁltering candidate entities

before disambiguation brings the beneﬁts of re-

duced inference time and potential improvements

in accuracy. To perform this step in the NICE

method, we follow the work of Tedeschi et al.

(2021) by using entity type information. Given

an entity mention

surrounded by textual context

(contlef t, contright)

and a list of candidate entities

cands ={e1, . . . ., en}

, we use a NER classiﬁer to

predict the top-k possible entity types of

. Then,

we discard all candidate entities that have an entity

type not matching any of these kclasses:

candsfiltered ={ei:type(ei)∈ˆ

T|ei∈cands},

where

is the set of top-k predicted entity types. If

the conﬁdence score of the NER classiﬁer is above

a threshold value

, only one class is used instead

. In the current setup of the NICE method, the

number of top predicted classes is

k= 3

and the

conﬁdence threshold value is

t= 1

, which means

that the classiﬁer always outputs the top-3 entity

classes. Figure 2shows an example of NER-based

candidate ﬁltering.

To obtain the entity types for the candidates, we

use the Wiki2NER dictionary provided by Tedeschi

et al. (2021)

. Then, instead of using the NER clas-

siﬁer as provided by Tedeschi et al. (2021), which

has been trained only on the AIDA training set and

therefore may be biased towards frequent entities

as well, we introduce a reﬁned version of it, which

is more robust to overshadowing. Speciﬁcally, we

ﬁlter the training set of BLINK (Wu et al.,2020)

by discarding the entries where the ground truth

answer has the highest popularity score among all

candidate entities. Then, we use the 2M remaining

data entries to ﬁne-tune the classiﬁer. The moti-

vation behind ﬁne-tuning the classiﬁer rather than

training it from scratch is to achieve an improve-

ment in recognising overshadowed entities without

2https://github.com/Babelscape/ner4el

BLINK is a dataset for ED consisting of 9M entries ex-

tracted from Wikipedia.

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：

FocusingonContextisNICE:ImprovingOvershadowedEntityDisambiguationVeraProvatorova1,SimoneTedeschi2;3,SvitlanaVakulenko4RobertoNavigli2andEvangelosKanoulas11UniversityofAmsterdam,2SapienzaUniversityofRome3Babelscape,Italy,4AmazonAlexaAI{v.provatorova,e.kanoulas}@uva.nl{tedeschi,navigli}@diag.uniroma1....

展开>> 收起<<

Focusing on Context is NICE Improving Overshadowed Entity Disambiguation Vera Provatorova1 Simone Tedeschi23 Svitlana Vakulenko4.pdf

共10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！

Focusing on Context is NICE Improving Overshadowed Entity Disambiguation Vera Provatorova1 Simone Tedeschi23 Svitlana Vakulenko4

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: