MetaASSIST Robust Dialogue State Tracking with Meta Learning Fanghua YeyXi WangyJie HuangzShenghui LixSamuel SternEmine Yilmazy yUniversity College London UK

2025-05-02 0 0 1.64MB 13 页 10玖币

侵权投诉

MetaASSIST: Robust Dialogue State Tracking with Meta Learning

Fanghua Ye†Xi Wang†Jie Huang‡Shenghui Li§Samuel Stern¶Emine Yilmaz†

†University College London, UK

‡University of Illinois at Urbana-Champaign, USA

§Uppsala University, Sweden

¶Afﬁniti AI, London, UK

{fanghua.ye.19, xi-wang, emine.yilmaz}@ucl.ac.uk

jeffhj@illinois.edu, shenghui.li@it.uu.se, samuel.stern@affiniti.ai

Abstract

Existing dialogue datasets contain lots of noise

in their state annotations. Such noise can hurt

model training and ultimately lead to poor gen-

eralization performance. A general framework

named ASSIST has recently been proposed to

train robust dialogue state tracking (DST) mod-

els. It introduces an auxiliary model to generate

pseudo labels for the noisy training set. These

pseudo labels are combined with vanilla labels

by a common ﬁxed weighting parameter to train

the primary DST model. Notwithstanding the

improvements of ASSIST on DST, tuning the

weighting parameter is challenging. Moreover,

a single parameter shared by all slots and all in-

stances may be suboptimal. To overcome these

limitations, we propose a meta learning-based

framework MetaASSIST to adaptively learn the

weighting parameter. Speciﬁcally, we propose

three schemes with varying degrees of ﬂexibil-

ity, ranging from slot-wise to both slot-wise and

instance-wise, to convert the weighting param-

eter into learnable functions. These functions

are trained in a meta-learning manner by taking

the validation set as meta data. Experimental

results demonstrate that all three schemes can

achieve competitive performance. Most impres-

sively, we achieve a state-of-the-art joint goal

accuracy of 80.10% on MultiWOZ 2.4.

1 Introduction

Task-oriented dialogue systems have recently be-

come a hot research topic. They act as digital per-

sonal assistants, helping users with various tasks

such as hotel bookings, restaurant reservations, and

weather checks. Dialogue state tracking (DST) is

recognized as a core task of the dialogue manager.

Its goal is to keep track of users’ intentions at each

turn of the dialogue (Mrkši´

c et al.,2017;Rastogi

et al.,2020). Tracking the dialogue state accurately

is of signiﬁcant importance, as the state informa-

tion will be fed into the dialogue policy learning

module to determine the next system action to per-

form (Manotumruksa et al.,2021). In general, the

Auxiliary DST Model Primary DST Model

User: “Hello, could you please find a hotel in the downtown for me?”

Pseudo Label

hotel

area: center

Vanilla L abel

hotel-area: east

Combined Label

hotel-area

standard augmented

EWSNCEWSNCEWSNC

#1 − #

trained on a small clean dataset

# = 0.5

Figure 1: The structure of ASSIST and MetaASSIST.

Both frameworks utilize soft labels obtained by linearly

combining pseudo labels (one-hot) and vanilla labels

(one-hot) using a weighting parameter αto enhance the

training process compared to standard training that only

relies on vanilla noisy labels. ASSIST adopts a single

αshared by all slots and all training samples, while

MetaASSIST uses slot-wise (and instance-wise) αs.

dialogue state is represented as a set of (slot,value)

pairs (Henderson et al.,2014;Budzianowski et al.,

2018). The slots for a particular task or domain are

predeﬁned (e.g., “hotel-name”). Their values are

extracted from the dialogue context.

So far, a great variety of DST models have been

proposed (Wu et al.,2019;Campagna et al.,2020;

Balaraman et al.,2021;Lee et al.,2021;Guo et al.,

2022;Shin et al.,2022;Wang et al.,2022). These

models assume that all state labels provided in the

dataset are correct, without considering the effect

of label noise. However, dialogue state annotations

are error-prone, especially considering that most

dialogue datasets (e.g., MultiWOZ Budzianowski

et al.,2018) are collected through crowdsourcing.

The presence of label noise may impair model train-

ing and lead to poor generalization performance of

the trained model, as deep neural models can easily

overﬁt noisy training data (Zhang et al.,2021).

In order to robustly train DST models from noisy

labels, Ye et al. (2022) proposed a general frame-

work dubbed ASSIST, which augments the stan-

dard model training procedure with a small clean

dataset. As shown in Figure 1, ASSIST ﬁrst trains

arXiv:2210.12397v1 [cs.CL] 22 Oct 2022

an auxiliary model on the small clean dataset and

applies this model to generate pseudo labels for

each sample in the noisy training set. Then, it lin-

early combines the pseudo labels and vanilla labels

to train the primary model. Both theoretically and

empirically, ASSIST has been shown to be effec-

tive in reducing the impact of label noise.

However, ASSIST adopts a common weighting

parameter to combine the pseudo labels and vanilla

labels for all slots and all training samples, which

is suboptimal. In reality, different slots tend to have

different noise rates (Eric et al.,2020), indicating

that the weighting parameter should be slot-wise.

On the other hand, different training samples may

also require different weighting parameters, since

whether pseudo labels or vanilla labels should be

preferred is highly dependent on speciﬁc training

instances. Furthermore, the weighting parameter is

considered a hyperparameter and thus needs to be

carefully tuned on each dataset.

To address the aforementioned limitations of AS-

SIST, we propose MetaASSIST, a meta learning-

based general framework that supports automati-

cally learning slot-wise (and instance-wise) weight-

ing parameters. Speciﬁcally, our contributions are:

•

We propose three different schemes for trans-

forming the weighting parameters into learn-

able functions. These schemes have varying

degrees of ﬂexibility, ranging from slot-wise

to both slot-wise and instance-wise.

•

We propose to train these learnable functions

through a meta-learning paradigm that takes

the validation set as meta data and adaptively

adjusts the parameters of each learnable func-

tion (as a result, the weighting parameters) by

reducing the validation loss.

•

We conduct extensive experiments to test the

effectiveness of the proposed three schemes.

All of them achieve superior performance. For

the ﬁrst time, we achieve over

80%

joint goal

accuracy on MultiWOZ 2.4 (Ye et al.,2021a).

2 Preliminaries

In task-oriented dialogue systems, the DST module

transforms users’ goals or intentions expressed in

unstructured natural languages into structured state

representations (e.g., a series of slot-value pairs).

The state representations are continually updated

in each round of the user-system interactions.

2.1 Problem Statement

More formally, we symbolize a dialogue of

turns

X={(R1, U1),...,(RT, UT)}

, where

and

denote the system response and user utterance

at turn

(

1≤t≤T

), respectively. We adopt

represent the dialogue context from the ﬁrst turn to

the

-th turn, i.e.,

Xt={(R1, U1),...,(Rt, Ut)}

Further, let

denote the set of all the predeﬁned

slots and

Bt={(s, vt)|s∈ S}

the dialogue state

at turn

. Here,

is the corresponding value of slot

at turn

. Then, the DST problem is deﬁned as

learning a dialogue state tracker F:Xt→ Bt.

As discussed earlier, annotating dialogue states

via crowdsourcing is prone to incorrect and incon-

sistent labels. These noisy annotations are likely

to adversely affect model training. We denote the

noisy state annotations as

Bt={(s, ˜vt)|s∈ S}

where

˜vt

is the noisy label of slot

at turn

. In this

work,

refers to the labels provided in the dataset

and

refers to the unknown true state annotations.

As pointed out by Ye et al. (2022), existing DST

approaches are only able to learn a suboptimal di-

alogue state tracker

F:Xt→˜

rather than the

optimal dialogue state tracker

F:Xt→ Bt

. Aim-

ing at learning a strong dialogue state tracker

F∗

better approximate

,Ye et al. (2022) proposed a

general framework ASSIST that supports training

DST models robustly from noisy labels.

2.2 Overview of ASSIST

ASSIST assumes that a small clean dataset is avail-

able. Based on this assumption, it ﬁrstly trains an

auxiliary model on the clean dataset. Then, it lever-

ages the trained model to generate pseudo labels

for each sample in the large noisy training set. The

generated pseudo labels are expected to be a good

complement to the vanilla noisy labels. Therefore,

combining the two types of labels has the poten-

tial to reduce the inﬂuence of noisy labels when

training the primary model.

Denote the generated pseudo state annotations

Bt={(s, ˘vt)|s∈ S}

, where

˘vt

represents the

pseudo label of slot

at turn

. Within the frame-

work of ASSIST, the primary model is required

to predict

and

concurrently during the train-

ing process. In other words, the target of model

training turns into learning a dialogue state tracker

F∗:Xt→C(˘

Bt,˜

Bt)

, where

C(˘

Bt,˜

Bt)

denotes a

combination of

and

. There can be different

methods to combine the generated pseudo labels

and vanilla noisy labels. The most straightforward

way is to combine them linearly, which is also the

strategy adopted in ASSIST. The linearly combined

label of slot sat turn tis formulated as:

t=α˘

vt+ (1 −α)˜

vt,(1)

where

and

are the one-hot vector representa-

tion of the pseudo label

˘vt

and vanilla noisy label

˜vt

, respectively. The parameter

α(0 ≤α≤1)

employed to control the weights of ˘

vtand ˜

vt.

Let

p(˘vt|Xt, s)

denote the likelihood of

˘vt

and

p(˜vt|Xt, s)

the likelihood of

˜vt

. Then, the likeli-

hood of the combined label vc

tis calculated as:

p(vc

t|Xt, s) = p(˘vt|Xt, s)αp(˜vt|Xt, s)(1−α).(2)

Based on this formula, the training objective of the

primary model can be derived as follows:

L=1

|Dn||S| X

Xt∈Dn

s∈S

−log p(vc

t|Xt, s)

=α

|Dn||S| X

Xt∈Dn

s∈S

−log p(˘vt|Xt, s)+

(1 −α)

|Dn||S| X

Xt∈Dn

s∈S

−log p(˜vt|Xt, s),

(3)

where Dnrepresents the noisy training set.

3 MetaASSIST: A Meta Learning-Based

Version of ASSIST

Equations

(1)

and

(3)

show that a single

is shared

by all slots when combining the pseudo labels and

vanilla labels. This is suboptimal, as the ratio of the

noise rate of pseudo labels to that of vanilla labels

tends to be different for different slots. When the

vanilla labels have higher quality than the gener-

ated pseudo labels,

should be set to a small value;

otherwise, a large

should be used. This implies

that setting

to different values for different slots

can help train the primary model more robustly. In

the following, we ﬁrst theoretically show that the

combined labels obtained via slot-wise weighting

parameters instead of a common one can better ap-

proximate the unknown true labels. Then, we elab-

orate on the proposed framework MetaASSIST.

3.1 Theoretical Justiﬁcation

Following (Ye et al.,2022), we employ the mean

squared loss to deﬁne the mean approximation error

of any corrupted labels

to their corresponding

unknown true labels vt, as formularized below:

Y¨

v=1

|Dn||S| X

Xt∈Dn

s∈S

EDc[k¨

vt−vtk2

2].(4)

Here,

refers to the small clean dataset. Both

and vtare the vector representations of labels.

Let

αs

be the slot-wise weighting parameter for

slot

. We utilize

to denote the combined label

obtained by replacing αwith αsin Eq. (1). Thus,

t=αs˘

vt+ (1 −αs)˜

vt.(5)

Same as α,αsis also bounded between 0 and 1.

Substituting the corrupted labels

in Eq.

(4)

with vs

tand vc

t, we have the following theorem:

Theorem 1.

The optimal mean approximation er-

ror with respect to the combined labels

derived

from slot-wise weighting parameters

αs

is smaller

than or equal to that of the combined labels

derived from a shared weighting parameter

, i.e.,

min

αs

Yvs≤min

αYvc.

Proof.

The conclusion is obvious as we can replace

αswith αif Yvc< Yvs, but not vice versa.

3.2 Slot-Wise Weighting Parameters as Meta

Learnable Functions

In the framework of ASSIST,

is treated as a hy-

perparameter. It needs to be meticulously tuned in

the training phase so as to help the primary model

achieve the best performance. Although it is feasi-

ble to tune a single parameter

, it would become

extremely painful to tune all the slot-wise parame-

ters. This is because multi-domain dialogues can

have dozens of or even hundreds of slots (e.g., there

are 37 slots in the MultiWOZ dataset Eric et al.,

2020). To circumvent the troublesome step of tun-

ing each slot-wise parameter

αs

of slot

, we pro-

pose to learn all these parameters automatically via

meta learning (Hospedales et al.,2021).

Speciﬁcally, we propose three different schemes

to cast the slot-wise weighting parameters as learn-

able functions, which are described in detail below:

Scheme One (S1):

The ﬁrst scheme assumes that

the parameter

αs

is fully independent of the dia-

logue context

. As a consequence of this assump-

tion, all the training samples will share the same

αs

for slot

. Given that the parameter

αs

is restricted

to fall in the range of 0 to 1, it is tricky to learn it by

gradient-based optimizers. In our implementation,

we introduce an unconstrained learnable parameter

wsand regard αsas a Sigmoid function of ws:

αs=f1(ws) = Sigmoid(ws).(6)

As thus, the parameter

rather than

αs

will be

directly optimized during the training process.

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：

MetaASSIST:RobustDialogueStateTrackingwithMetaLearningFanghuaYeyXiWangyJieHuangzShenghuiLixSamuelStern{EmineYilmazyyUniversityCollegeLondon,UKzUniversityofIllinoisatUrbana-Champaign,USAxUppsalaUniversity,Sweden{AfnitiAI,London,UK{fanghua.ye.19,xi-wang,emine.yilmaz}@ucl.ac.ukjeffhj@illinois.edu,shen...

展开>> 收起<<

MetaASSIST Robust Dialogue State Tracking with Meta Learning Fanghua YeyXi WangyJie HuangzShenghui LixSamuel SternEmine Yilmazy yUniversity College London UK.pdf

共13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！

MetaASSIST Robust Dialogue State Tracking with Meta Learning Fanghua YeyXi WangyJie HuangzShenghui LixSamuel SternEmine Yilmazy yUniversity College London UK

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: