ReSel N-ary Relation Extraction from Scientiﬁc Text and Tables by Learning to Retrieve and Select Yuchen Zhuang1 Yinghao Li1 Jerry Junyang Cheung1 Yue Yu1

2025-04-29 0 0 1.4MB 15 页 10玖币

侵权投诉

ReSel: N-ary Relation Extraction from Scientiﬁc Text and Tables by

Learning to Retrieve and Select

Yuchen Zhuang1, Yinghao Li1, Jerry Junyang Cheung1, Yue Yu1,

Yingjun Mou1, Xiang Chen2, Le Song3,4, Chao Zhang1

1Georgia Institute of Technology, Atlanta, USA

2Adobe Research, San Jose, USA 3MBZUAI, Abu Dhabi, UAE 4BioMap, Beijing, China

1{yczhuang,yinghaoli,jzhang3027,yueyu,ymou32,

chaozhang}@gatech.edu

2xiangche@adobe.com 3le.song@mbzuai.ac.ae

Abstract

We study the problem of extracting N-ary rela-

tion tuples from scientiﬁc articles. This task is

challenging because the target knowledge tu-

ples can reside in multiple parts and modali-

ties of the document. Our proposed method

RESEL decomposes this task into a two-stage

procedure that ﬁrst retrieves the most relevant

paragraph/table and then selects the target en-

tity from the retrieved component. For the

high-level retrieval stage, RESEL designs a

simple and effective feature set, which cap-

tures multi-level lexical and semantic similar-

ities between the query and components. For

the low-level selection stage, RESEL designs

a cross-modal entity correlation graph along

with a multi-view architecture, which mod-

els both semantic and document-structural re-

lations between entities. Our experiments on

three scientiﬁc information extraction datasets

show that RESEL outperforms state-of-the-art

baselines signiﬁcantly. 1

1 Introduction

Scientiﬁc information extraction (SciIE) (Augen-

stein et al.,2017;Luan et al.,2018;Jiang et al.,

2019), the task of extracting scientiﬁc concepts

along with their relations from scientiﬁc literature

corpora, is important for researchers to keep abreast

of latest scientiﬁc advances. A key subtask of SciIE

is the

-ary relation extraction problem (Jia et al.,

2019;Jain et al.,2020), which aims to extract the

relations of different entities as

-ary knowledge

tuples. This problem is challenging because the

entities of the knowledge tuples often reside in

multiple sections (e.g., abstracts, experiments) and

modalities (e.g., paragraphs, tables, ﬁgures) of the

document. Effective scientiﬁc

-ary relation ex-

traction requires not only understanding the seman-

tics of different modalities, but also performing

Our code is available on

https://github.com/

night-chen/ReSel.

Query 1: <Dependency Parsing, English PTB, Arc-

hybrid (dyn, α=0.75), LAS>

Answer 1: 91.42

Support 1:

Query 2: <Dependency Parsing, English PTB, Arc-

hybrid (dyn, α=0.75), UAS>

Answer 2: 93.56

Support 2:

Scientific Literature Query Examples

Figure 1: Illustration of the multi-modal scientiﬁc N-

ary relation extraction problem on the SciREX dataset.

document-level inference based on interleaving sig-

nals such as co-occurrences, co-references, and

structural relations, as shown in Figure 1.

Document-level

-ary relation extraction has

been studied in literature (Jia et al.,2019;Jain et al.,

2020;Viswanathan et al.,2021;Liu et al.,2021).

Some works (Zeng et al.,2020;Tu et al.,2019) use

graph-based approaches to model long-distance

relations in the document with the focus on text

only. However, for scientiﬁc articles, an equally if

not more important data structure is the table, as

scientiﬁc results are often reported in tables and

then referred to and discussed in text. There are

also works that pre-train large-scale transformer

models on massive table and text pairs (Yin et al.,

2020;Herzig et al.,2020). These methods are de-

signed for question answering, which are strong

at retrieving answers that semantically match the

query but fall short in inferring ﬁne-grained entity-

level N-ary relations. Besides, to perform well on

SciIE, they usually require large task-speciﬁc data

to ﬁne-tune the pre-trained model, especially for

long documents that contain many candidates. But

in practice, such large-scale annotation data can be

expensive and labor-intensive to curate. Therefore,

extracting

-ary relations jointly from scientiﬁc

text and tables still remains an important but chal-

arXiv:2210.14427v1 [cs.CL] 26 Oct 2022

lenging problem.

We propose RESEL, a hierarchical

trieve-and-

Sel

ection model for multi-modal and document-

level SciIE. In RESEL, we pose the

-ary relation

extraction problem as a question answering task

over text and tables (Figure 1). RESEL then decom-

poses the challenging task into two simpler sub-

tasks: (1) high-level component retrieval, which

aims to locate the target paragraph/table where the

ﬁnal target entity resides, and (2) low-level entity

extraction, which aims to select the target entity

from the chosen component.

For high-level component (i.e., paragraph or ta-

ble) retrieval, we design a feature set that com-

bines the strengths of two classes of retrieval

methods: (1) sparse retrieval (Aizawa,2003;

Robertson and Zaragoza,2009) that represents the

query-candidate pairs as high-dimensional sparse

vectors to encode lexical features; (2) dense re-

trieval (Karpukhin et al.,2020) that leverages la-

tent semantic embeddings to represent query and

candidates. We design sparse and dense retrieval

features for query-component pairs by augmenting

BERT (Devlin et al.,2019)-based semantic similar-

ities with entity-level semantic and lexical similar-

ities, allowing for training an accurate high-level

retriever using only a small amount of labeled data.

The low-level entity extraction stage aims to

infer

-ary entity relations from complex and

noisy signals across paragraphs and tables. In

this stage, we ﬁrst build a cross-modal entity-

correlation graph, which encodes different entity-

entity relations such as co-occurrence, co-reference,

and table structural relations. While most of the

existing methods (Zheng et al.,2020;Zeng et al.,

2020) use BERT embeddings as node representa-

tions, we ﬁnd BERT embeddings limited in dis-

tinguishing adjacent table cells or similar entities.

This issue is even more severe when the BERT em-

beddings are propagated on the graph. To address

this, we design a new bag-of-neighbors (BON) rep-

resentation. It computes the lexical and semantic

similarities between each candidate entity and its 1-

hop neighbors. We then feed the BON features into

a graph attention network (GAT) to capture both

neighboring semantics and structural correlations.

Such GAT-learned features and BERT-based em-

beddings are treated as two complementary views,

which are co-trained with a consistency loss.

We summarize our key contributions as follows:

(1) We propose a hierarchical retrieve-and-select

learning method that decomposes

-ary scientiﬁc

relation extraction into two simpler subtasks; (2)

For high-level component retrieval, we propose a

simple but effective feature-based model that com-

bines multi-level semantic and lexical features be-

tween queries and components; (3) For low-level

entity extraction, we propose a multi-view architec-

ture, which fuses graph-based structural relations

with BERT-based semantic information for extrac-

tion; (4) Extensive experiments on three datasets

show the superiority of both the high-level and low-

level modules in RESEL.

2 Related Work

Component Retrieval

For component retrieval,

traditional sparse retrieval methods such as TF-

IDF (Aizawa,2003) and BM25 (Robertson and

Zaragoza,2009) focus on keyword-level match-

ing but ignore entity semantics. Recently, pre-

trained language models have also been used to

represent queries and documents in a learned

space (Karpukhin et al.,2020) and have been ex-

tended to handle tabular context (Herzig et al.,

2021;Ma et al.,2022). However, these methods

mainly focus on passage-level retrieval, and can-

not well capture ﬁne-grained entity-level seman-

tics (Zhang et al.,2020;Su et al.,2021). Such

an issue makes them suboptimal for encoding nu-

anced terms and descriptions in scientiﬁc articles.

In contrast, RESEL leverages both component- and

entity-level semantic and lexical features that help

the model better understand the correlations be-

tween components and queries.

N-ary Relation Extraction

Many existing

methods (Jia et al.,2019;Jain et al.,2020;

Viswanathan et al.,2021) treat

-ary relation ex-

traction as a binary classiﬁcation problem and pre-

dict whether the composition of

entities in the

document are valid or not. However, the candidate

space grows exponentially with N, and the perfor-

mance of the binary classiﬁers can be largely inﬂu-

enced by the number and quality of negative tuples.

Some other methods (Du et al.,2021;Huang et al.,

2021) formulate the problem as role-ﬁller entity ex-

traction and propose BERT-based generative mod-

els to extract the correct entities for each element of

the

-ary relation. None of these methods consider

-ary relation across modalities. Lockard et al.

(2020) leverages the layout information for extract-

ing relations from web pages. However, the layout

information in science articles are less prominent

and harder to be utilized.

3 Problem Formulation

In SciIE, we aim to extract information from a cor-

pus of

scientiﬁc articles. Each article, denoted

, is a sequence of

|D|

components, where each

component

Ci∈ D

i∈N[0,|D|)

can be either a

paragraph or a table. A table is ﬂattened and con-

catenated with its caption as a sequence of words.

Given a document

, we have a set of queries

with number

|Q|

. Each query contains

N−1

elements

Qj= [ej,1,· · · , ej,N−1], j ∈N[0,|Q|)

and the task is to extract the correct

-th ele-

ment from document

to form a valid

-ary rela-

tion. We assume a dataset

{xk, yk}M

k=1

that can be

used to learn such a

-ary relation extractor, each

sample includes a document and a set of queries

xk= (Dk,Qk)

, and each ground-truth label

indicates the target entity in the document Dk.

4 The RESEL Method

4.1 Component Retriever

In Stage I, we design a high-level model to retrieve

the most relevant paragraphs or tables that contain

the ﬁnal answer. We ﬁrst use BERT to embed the

paragraphs/tables into sequences of vectors (de-

tails in Appendix A.1). We encode the

-th query

Qj= [ej,1,· · · , ej,N−1]

, into query embedding

h(Qj)

and get the corresponding element embed-

dings of the query

h(ej,a), a ∈N[0,N )

. Similar to

the query encoder, we encode the

-th component,

, as component embedding

h(Ci)

, and the aver-

aged entity embeddings

h(mi,b)

, where

mi,b ∈Ci

indicates the

-th entity extracted from

. With

the encoded sequences of vectors, we compute the

different views of features for the component-query

pair

(Ci, Qj)

as follows to take advantage of both

the entity-level matching signals and component-

level semantic signals, which are complementary:

Component-Level Semantic Features (CS).

The ﬁrst view extracts the semantic features for

component-query pairs from two different an-

gles: (1) Embedding-Based Similarity: the cosine

similarities

fcs-1(Ci, Qj)

between component and

query embeddings. (2) Entailment-Based Score:

the classiﬁcation score

fcs-2(Ci, Qj)

between

and

calculated by feeding them both into a

BERT binary sequence classiﬁer as a concatenated

sequence (Nogueira and Cho,2019;Nie et al.,

2019). We concat these two scalar features as the

ﬁrst view fcs(Ci, Qj).

Entity-Level Semantic Features (ES).

The sec-

ond view computes entity-level cosine similarities

fes(mi,b, ej,a)

between the component entity em-

beddings

h(mi,b)

and the query elements embed-

dings

h(ej,a)

. With all these similarity scores,

we apply a max-pooling operation over all com-

ponent entities

mi,b

, and use the obtained max-

imum

fes(Ci, ej,a) = maxmi,b∈Cifes(mi,b, ej,a)

to represent the relation between the component

and one query element

ej,a

. Then, we gather

the relation scores

fes(Ci, ej,a)

as the ﬁnal entity-

level semantic feature vector:

fes(Ci, Qj) =

[fes(Ci, ej,1),· · · , fes(Ci, ej,N-1)]T.

Entity-Level Lexical Features (EL).

Our third

view extracts lexical features between component

entities and the query elements. We compute three

text similarities (Appendix A.2): (1) Levenshtein

Distance (Levenshtein et al.,1966); (2) the length

of Longest Common Substring; (3) the length of

Longest Common Subsequence. As the metrics vary

in scale according to the length of the strings, we

use the normalized metrics

fel(mi,b, ej,a)∈[0,1]3

via dividing by involved string lengths. Simi-

lar to ES features, we perform max-pooling to

obtain the relation scores between the compo-

nent and a single query element,

fel(Ci, ej,a) =

maxmi,b∈Cifel(mi,b, ej,a)

and concatenate the re-

sults as entity-level lexical features:

fel(Ci, Qj) =

[fel(Ci, ej,1)T⊕ · · · ⊕ fel(Ci, ej,N −1)T]T.

We aggregate the features to predict which com-

ponent has the highest probability to contain the

ﬁnal answer. As the features in the three views

share the same scale range and similar dimension-

ality, we just concatenate these features together

fh= [fT

cs ⊕fT

es ⊕fT

el ]T

, and train one uniﬁed

classiﬁer over fhfor component retrieval.

4.2 Entity Extractor

In Stage II, we use the predictions from Stage I

to restrict the searching space for low-level entity

extraction.

4.2.1 Multi-Modal Entity-Level Graph

To model document-level entity correlations, we

construct a multi-modal entity correlation graph

G= (V,E)

, where

V={v1, v2,· · · , v|V|}

de-

notes the entity nodes, and

E={E(vi,vj)|vi, vj∈

V;i, j ∈N[1,|V|]}

denotes the edges between them.

Each node

vi∈ V

represents a paragraph en-

tity or a table cell. We construct different edge

types to model the intra- and inter-modality rela-

tions to encode the entity correlation across modal-

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：

ReSel:N-aryRelationExtractionfromScienticTextandTablesbyLearningtoRetrieveandSelectYuchenZhuang1,YinghaoLi1,JerryJunyangCheung1,YueYu1,YingjunMou1,XiangChen2,LeSong3;4,ChaoZhang11GeorgiaInstituteofTechnology,Atlanta,USA2AdobeResearch,SanJose,USA3MBZUAI,AbuDhabi,UAE4BioMap,Beijing,China1fyczhuang,yi...

展开>> 收起<<

ReSel N-ary Relation Extraction from Scientiﬁc Text and Tables by Learning to Retrieve and Select Yuchen Zhuang1 Yinghao Li1 Jerry Junyang Cheung1 Yue Yu1.pdf

共15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！

ReSel N-ary Relation Extraction from Scientiﬁc Text and Tables by Learning to Retrieve and Select Yuchen Zhuang1 Yinghao Li1 Jerry Junyang Cheung1 Yue Yu1

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: