On the Interpolation of Contextualized Term-based Ranking with BM25 for Query-by-Example Retrieval

2025-04-24 0 0 2.68MB 10 页 10玖币

侵权投诉

On the Interpolation of Contextualized Term-based Ranking

with BM25 for ery-by-Example Retrieval

Amin Abolghasemi

m.a.abolghasemi@liacs.leidenuniv.nl

Leiden University

Leiden, Netherlands

Arian Askari

a.askari@liacs.leidenuniv.nl

Leiden University

Leiden, Netherlands

Suzan Verberne

s.verberne@liacs.leidenuniv.nl

Leiden University

Leiden, Netherlands

ABSTRACT

Term-based ranking with pre-trained transformer-based language

models has recently gained attention as they bring the contextu-

alization power of transformer models into the highly ecient

term-based retrieval. In this work, we examine the generalizability

of two of these deep contextualized term-based models in the con-

text of query-by-example (QBE) retrieval in which a seed document

acts as the query to nd relevant documents. In this setting — where

queries are much longer than common keyword queries — BERT

inference at query time is problematic as it involves quadratic com-

plexity. We investigate TILDE and TILDEv2, both of which leverage

BERT tokenizer as their query encoder. With this approach, there

is no need for BERT inference at query time, and also the query

can be of any length. Our extensive evaluation on the four QBE

tasks of SciDocs benchmark shows that in a query-by-example

retrieval setting TILDE and TILDEv2 are still less eective than

a cross-encoder BERT ranker. However, we observe that BM25

could show a competitive ranking quality compared to TILDE and

TILDEv2 which is in contrast to the ndings about the relative

performance of these three models on retrieval for short queries

reported in prior work. This result raises the question about the

use of contextualized term-based ranking models being benecial

in QBE setting. We follow-up on our ndings by studying the score

interpolation between the relevance score from TILDE (TILDEv2)

and BM25. We conclude that these two contextualized term-based

ranking models capture dierent relevance signals than BM25 and

combining the dierent term-based rankers results in statistically

signicant improvements in QBE retrieval. Our work sheds light

on the challenges of retrieval settings dierent from the common

evaluation benchmarks. It could be of value as future work to study

other contextualized term-based ranking models in QBE settings.

CCS CONCEPTS

•Information systems →Retrieval models and ranking

;

Eval-

uation of retrieval results.

KEYWORDS

Query-by-example retrieval, term-based retrieval, Transformer mod-

els, BERT-based ranking

This work is licensed under a Creative Commons Attribution

International 4.0 License.

ICTIR ’22, July 11–12, 2022, Madrid, Spain.

ACM ISBN 978-1-4503-9412-3/22/07.

https://doi.org/10.1145/3539813.3545133

ACM Reference Format:

Amin Abolghasemi, Arian Askari, and Suzan Verberne. 2022. On the In-

terpolation of Contextualized Term-based Ranking with BM25 for Query-

by-Example Retrieval. In Proceedings of the 2022 ACM SIGIR International

Conference on the Theory of Information Retrieval (ICTIR ’22), July 11–12,

2022, Madrid, Spain. ACM, New York, NY, USA, 10 pages. https://doi.org/10.

1145/3539813.3545133

1 INTRODUCTION

Query-by-Example (QBE) retrieval is an Information Retrieval (IR)

setting in which a seed document

acts as the query to represent

the user’s information need and the retrieval engine searches over

a collection of the same type of documents [

]. This

retrieval setup is typical in professional, domain-specic tasks such

as legal case law retrieval [

], patent prior art search [

and scientic literature search [

]. While using a document

as a query could become challenging due to its length and complex

semantic structure, prior work has shown that traditional term-

based retrieval models like BM25 [

] are highly eective when

used in QBE retrieval [1, 2, 28].

Recently, deep contextualized term-based retrieval models have

gained attention as they bring the contextualization power of the

pre-trained transformer-based language models into the highly

ecient term-based retrieval. Examples of such models are Deep-

Impact [

], SPLADE [

], SPLADEv2 [

], TILDE [

], TILDEv2

[

], COIL [

], and uniCOIL [

]. Here, we specically investi-

gate TILDE, which is a term independent likelihood model, and

its follow-up TILDEv2 which is a deep contextualized lexical exact

matching model.

TILDE and TILDEv2, which are introduced as term-based re-

ranking models, follow a recent paradigm in term-based retrieval

where term importance is pre-computed with scalar term weights.

Besides, to predict the relevance score, both of these models use the

BERT tokenizer as their query encoder which means that they do

not need to perform any BERT inference at query time to encode the

query. However, leveraging tokenizer-based encoding of the query

trades o the query representation and therefore eectiveness with

higher eciency at inference time [

]. While the eectiveness of

these models is evaluated on tasks and benchmarks where we have

short queries, e.g., MSMARCO Passage Ranking [

] and the TREC

DL Track [

], in this paper, we evaluate them in the aforementioned

QBE retrieval setting where queries are much longer than common

keyword queries. In this regard, we address the following research

questions:

RQ1

How eective are TILDE and TILDEv2 in query-by-example

retrieval?

Throughout this paper, we use the term “document” to refer to a unit of retrieval [

]

arXiv:2210.05512v1 [cs.IR] 11 Oct 2022

ICTIR ’22, July 11–12, 2022, Madrid, Spain. Amin Abolghasemi, Arian Askari, and Suzan Verberne

A specic direction in answering RQ1 is to investigate the rank-

ing quality of TILDE and TILDEv2 in comparison with the eective

cross-encoder BERT ranker [

], which is described in section

2.4. We are interested in this direction for two reasons. First, the

cross-encoder BERT ranker exhibits quadratic complexity in both

space and time with respect to the input length [

] and this is ag-

gravated in QBE where we have long queries. TILDE and TILDEv2,

however, do not need any BERT inference at query time. Second,

due to the maximum input length of BERT, cross-encoder BERT

ranker, which uses the concatenation of the query and the docu-

ment, might not cover the whole query and document tokens in a

QBE setting, whereas in TILDE and TILDEv2, the query can be of

any length and documents are covered up to the maximum length

of BERT.

Additionally, since TILDEv2 pre-computes the term weights only

for those tokens existing in the documents, one risk is that it might

aggravate the vocabulary mismatch problem. A typical approach to

address this issue is to use document expansion methods. Zhuang

and Zuccon

[33]

use TILDE as their document expansion model

for TILDEv2. We adopt that approach for our task and further

investigate the impact of token-based document expansion with

TILDE on the ranking quality of TILDEv2 in a QBE retrieval setting.

Apart from comparing TILDE and TILDEv2 to the cross-encoder

BERT ranker, we also make a comparison to traditional lexical

matching models (BM25 and Probabilistic Language models), which

have been shown as strong baselines on QBE tasks in prior work [

28]:

RQ2

What is the eectiveness of traditional lexical matching mod-

els with varying tokenization strategies in comparison to

TILDE and TILDEv2?

To answer RQ2 we will investigate the eect of using the BERT

tokenizer [

] as pre-processing for traditional term-based retrieval

models. By doing so, we are aligning the index vocabulary of tradi-

tional models with that of TILDE and TILDEv2, which could make

our comparison more fair.

We will see in the Section 4 that BM25 shows a competitive

ranking quality in comparison to TILDE and TILDEv2 in our QBE

benchmark. Because of the similar quality on average, we are in-

terested to see if the relevance signals of TILDE and TILDEv2 are

dierent from that of BM25, to nd out if the methods are comple-

mentary to each other. To this aim, we will investigate the following

research question:

RQ3

To what extent do TILDE and TILDEv2 encode a dierent

relevance signal from BM25?

To address the question above, as it is described in details in

Section 3.3, we will analyze the eect of the interpolation of the

scores of TILDE and TILDEv2 with BM25.

Since TILDE and TILDEv2 are introduced as re-ranking models,

we use four dierent tasks from the SciDocs evaluation benchmark

[

] as a domain-specic QBE benchmark. This benchmark uses

scientic paper abstracts as the query and documents. The retrieval

setting in these tasks suits as a re-ranking setup because of the

number of documents to be ranked for each query. Since that we

are working in a domain-specic evaluation setting, we will also

address the following research question:

RQ4

To what extent does a highly tailored domain-specic pre-

trained BERT model aect the eectiveness of TILDE and

TILDEv2 in comparison to a BERTbase model?

In summary our main contributions in this work are three-fold:

•

We show that two recent transformer-based lexical models

(TILDE and TILDEv2) are less eective in Query-by-Example

retrieval than was expected based on results reported for ad

hoc retrieval. This indicates that QBE retrieval is structurally

dierent from other IR settings and requires special attention

for methods development;

•

We show that the relevance signals of TILDE and TILDEv2

can be complementary to that of BM25 as interpolation of the

methods leads to an improvement in ranking eectiveness;

•

We also investigate interpolations of BM25 with TILDE and

TILDEv2 in an ideal setting where the optimal interpolation

weight is known a priori, and by doing so, we show that

more stratied approaches for the interpolation could result

in higher gains from the interpolation of BM25 with TILDE

and TILDEv2.

In section 2 we describe the retrieval models used in this work.

In section 3 we provide details about our methods and experiments

and in section 4 we analyze the results and discuss the answers to

our research questions. Section 5 is dedicated to to further analysis

of the results, and nally, in Section 6 we provide the conclusion.

The code used in this paper is available at: https://github.com/

aminvenv/lexica

2 BACKGROUND: RETRIEVAL MODELS

In this section, we briey introduce the retrieval models that we

implement and evaluate in our experiments.

2.1 Traditional lexical matching models

BM25. For BM25 [

], we use the implementation by Elastic-

with the parameters

𝑘=

75, and

𝑏=

1, which was tuned

over the validation set.

Probabilistic Language Models. For language modeling (LM) based

retrieval [

], we use the built-in similarity functions of Elas-

ticsearch for the implementation of language model with Jelinek

Mercer (JM) smoothing [32].

2.2 Term Independent Likelihood Model:

TILDE

TILDE is a tokenizer-based term-based retrieval model which fol-

lows a term independence assumption and formulates the likelihood

of a query as follows:

TILDE-QL(𝑞|𝑑)=

|𝑞|



𝑖

𝑙𝑜𝑔(𝑃𝜃(𝑞𝑖|𝑑)) (1)

in which

𝑞

is the query, and

𝑑

is the document. As Figure 1 shows,

to compute the relevance score, the text of a document

𝑑

is fed

as the input for BERT and the log probability for each token is

estimated by using a language modeling head on top of the BERT

[CLS] token output. In other words, we are pre-computing the

2https://github.com/elastic/elasticsearch

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：

OntheInterpolationofContextualizedTerm-basedRankingwithBM25forQuery-by-ExampleRetrievalAminAbolghasemim.a.abolghasemi@liacs.leidenuniv.nlLeidenUniversityLeiden,NetherlandsArianAskaria.askari@liacs.leidenuniv.nlLeidenUniversityLeiden,NetherlandsSuzanVerbernes.verberne@liacs.leidenuniv.nlLeidenUnivers...

展开>> 收起<<

On the Interpolation of Contextualized Term-based Ranking with BM25 for Query-by-Example Retrieval.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！

On the Interpolation of Contextualized Term-based Ranking with BM25 for Query-by-Example Retrieval

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: