Pseudo-OOD training for robust language models Dhanasekar Sundararaman1 Nikhil Mehta1 Lawrence Carin1 1Duke University

2025-05-02 0 0 482.33KB 8 页 10玖币

侵权投诉

Pseudo-OOD training for robust language models

Dhanasekar Sundararaman1∗, Nikhil Mehta1∗, Lawrence Carin1

1Duke University

{ds448,nm208}@duke.edu

Abstract

While pre-trained large-scale deep models

have garnered attention as an important topic

for many downstream natural language pro-

cessing (NLP) tasks, such models often make

unreliable predictions on out-of-distribution

(OOD) inputs. As such, OOD detection is

a key component of a reliable machine learn-

ing model for any industry-scale application.

Common approaches often assume access to

additional OOD samples during the training

stage, however, outlier distribution is often un-

known in advance. Instead, we propose a

post hoc framework called POORE - POsthoc

pseudo Ood REgularization, that generates

pseudo-OOD samples using in-distribution

(IND) data. The model is ﬁne-tuned by in-

troducing a new regularization loss that sepa-

rates the embeddings of IND and OOD data,

which leads to signiﬁcant gains on the OOD

prediction task during testing. We extensively

evaluate our framework on three real-world di-

alogue systems, achieving new state-of-the-art

in OOD detection.

1 Introduction

Detecting Out-of-Distribution (

OOD

) (Goodfel-

low et al.,2014;Hendrycks and Gimpel,2016;

Yang et al.,2021) samples is vital for develop-

ing reliable machine learning systems for various

industry-scale applications of natural language un-

derstanding (NLP) (Shen et al.,2019;Sundarara-

man et al.,2020) including intent understanding

in conversational dialogues (Zheng et al.,2020;

Li et al.,2017), language translation (Denkowski

and Lavie,2011;Sundararaman et al.,2019), and

text classiﬁcation (Aggarwal and Zhai,2012;Sun-

dararaman et al.,2022). For instance, a language

understanding model deployed to support a chat

system for medical inquiries should reliably de-

tect if the symptoms reported in a conversation

∗The authors contributed equally to this work

constitute an

OOD

query so that the model may

abstain from making incorrect diagnosis (Sied-

likowski et al.,2021).

Although

OOD

detection has attracted a great

deal of interest from the research commu-

nity (Goodfellow et al.,2014;Hendrycks and Gim-

pel,2017;Lee et al.,2018), these approaches are

not speciﬁcally designed to leverage the structure

of textual inputs. Consequently, commonly used

OOD

approaches often have limited success in real-

world NLP applications. Most prior

OOD

methods

for NLP systems (Larson et al.,2019;Chen and Yu,

2021;Kamath et al.,2020) typically assume addi-

tional

OOD

data for outlier exposure (Hendrycks

et al.,2018). However, such methods risk over-

ﬁtting to the chosen OOD set, while making the

assumption that a relevant OOD set is available dur-

ing the training stage. Other methods (Gangal et al.,

2020;Li et al.,2021;Kamath et al.,2020) assume

training a calibration model, in addition to the clas-

siﬁer, for detecting

OOD

inputs. These methods

are computationally expensive as they often require

re-training the model on the downstream task.

Motivated by the above limitations, we propose

a framework called POsthoc pseudo Ood REgular-

ization (

POORE

) that generates pseudo-OOD data

using the trained classiﬁer and the In-Distribution

(

IND

) samples. As opposed to methods that use

outlier exposure, our framework doesn’t rely on any

external OOD set. Moreover,

POORE

can be eas-

ily applied to already deployed large-scale models

trained on a classiﬁcation task, without requiring

to re-train the classiﬁer from scratch. In summary,

we make the following contributions:

We propose a Mahalanobis-based context

masking scheme for generating pseudo-OOD

samples that can be used during the ﬁne-

tuning.

We introduce a new Pseudo Ood Regular-

ization (

POR

) loss that maximizes the dis-

arXiv:2210.09132v1 [cs.CL] 17 Oct 2022

tance between IND and generated pseudo-

OOD samples to improve the OOD detection.

Though extensive experiments on the three

benchmarks, we show that our approach per-

forms signiﬁcantly better than existing base-

lines.

2 Related Works

OOD Detection

. It is a binary classiﬁcation

problem that seeks to identify unfamiliar inputs

during inference from in-distribution (

IND

)

data observed during training. Standard

OOD

methods can be divided into two categories. The

ﬁrst category (Lee et al.,2018;Podolskiy et al.,

2021;Nalisnick et al.,2019;Ren et al.,2019)

corresponds to approximating a density

pI N D (x)

where density is used as a conﬁdence estimate

for binary classiﬁcation. The second category of

approaches (Hendrycks and Gimpel,2016,2017;

Li et al.,2017;Gal and Ghahramani,2016) use the

predictive probability to estimate the conﬁdence

scores. In our experiments, we compare against

approaches from both the categories.

OOD Detection in NLP.

There have been several

methods developed for

OOD

detection in NLP. Li

et al. (2021) proposed using

sub models, where

each model is trained with different masked inputs.

Kamath et al. (2020) uses an external

OOD

set

to train an additional calibration model for

OOD

detection. Most related to our proposed frame-

work is MASKER (Moon et al.,2021) that lever-

ages

IND

data to generate pseudo-

OOD

samples,

and uses self-supervision loss inspired from Devlin

et al. (2018) and predictive entropy regularization

for pseudo-

OOD

inputs. We also use BERT self-

supervision inspired keyword masking, however,

we propose a novel keyword selection criterion.

Moreover, we also introduce a novel model regu-

larization loss that directly increases the distance

of IND and pseudo-OOD samples.

3 Preliminaries and Notations

We consider a deep learning model

g◦f(x)

composed of an encoder

f:X → F

and a

classiﬁer

that maps

f(x)

to the output space,

where

x∈ X

corresponds to natural sentences

composed of a sequence of tokens

vi∈ V

i.e.

x= [v1, . . . , vT]

is the length of the sequence,

and

is the token vocabulary. For a downstream

classiﬁcation task, the class prediction is deﬁned

as p(y|x) = softmax (g(f((x))).

Architecture.

In this work, we construct

using the bi-directional Transformer architecture

(Vaswani et al.,2017). Speciﬁcally, we use the en-

coder architecture proposed in Devlin et al. (2018)

such that

f(x)

is the ﬁnal hidden representation

of the CLS token. We use a two-layer multi-layer

perceptron (MLP) as the classiﬁer g.

Mahalanobis OOD Scoring. OOD

methods typ-

ically learn a conﬁdence estimator that outputs

a score

s(x)∈R

such that

s(xind)> s(xood)

where

xind

and

xood

are sampled from

IND

dis-

tribution (

DI N D

) and

OOD

distribution (

DOOD

)

respectively. Lee et al. (2018) proposed using Ma-

halanobis distance estimator for OOD detection

that uses pre-trained features of the softmax neural

classiﬁer. Namely, given feature of a test sample

φ(x)

, the mahalanobis score

sM(x)

is computed

as follows

d(x, c) = (φ(x)−ˆµc)Tˆ

Σ−1(φ(x)−ˆµc)(1)

sM(x) = −min

cd(x, c)(2)

where

is an intermediate layer of the neural clas-

siﬁer and

denotes the class. The parameters of

the estimator {ˆµc,ˆ

Σ}denote the class-conditional

mean and the tied covariance of the IND features.

4 Post hoc Pseudo-OOD Regularization

In this section, we describe our framework called

POsthoc pseudo Ood REgularization (

POORE

which uses pseudo-OOD samples for ﬁne-tuning

a pre-trained classiﬁer. We ﬁrst describe our

masking-based approach to generate pseudo-OOD

samples from the

IND

samples available during

training. These generated pseudo-OOD samples

are used to regularize the encoder during post-hoc

training of a pre-trained classiﬁer, which leads to

improved robustness of the model towards

OOD

samples.

4.1 Masking for Pseudo-OOD Generation

We perform context masking of

IND

samples for

generating pseudo OOD samples. To generate

context-masked pseudo

OOD

samples, we ﬁrst

identify a set of tokens

v∈ K ⊂ V

that have

high attention scores and consequently, a higher

inﬂuence in model predictions. Given the set

of keywords, we perform random masking of

non-keywords in a given IND sample

to generate

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：

Pseudo-OODtrainingforrobustlanguagemodelsDhanasekarSundararaman1,NikhilMehta1,LawrenceCarin11DukeUniversity{ds448,nm208}@duke.eduAbstractWhilepre-trainedlarge-scaledeepmodelshavegarneredattentionasanimportanttopicformanydownstreamnaturallanguagepro-cessing(NLP)tasks,suchmodelsoftenmakeunreliablepr...

展开>> 收起<<

Pseudo-OOD training for robust language models Dhanasekar Sundararaman1 Nikhil Mehta1 Lawrence Carin1 1Duke University.pdf

共8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！

Pseudo-OOD training for robust language models Dhanasekar Sundararaman1 Nikhil Mehta1 Lawrence Carin1 1Duke University

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: