A Multichannel Convolutional Neural Network For Cross-language Dialog State Tracking

0.18 MB·

by  Hongjie Shi

PublishedasaconferencepaperatIEEESLT2016
A MULTICHANNEL CONVOLUTIONALNEURAL NETWORK
FOR CROSS-LANGUAGEDIALOGSTATE TRACKING
HongjieShi,Takashi Ushio,MitsuruEndo,KatsuyoshiYamagamiandNoriakiHorii
InteractiveAIResearch Group,PanasonicCorporation,Osaka, Japan
7
1 ABSTRACT hand-craftedrule-based approaches[2, 3].This is a very un-
0
favorablesituationbecause buildinghand-annotatedtraining
2 The fifth Dialog State Tracking Challenge (DSTC5) intro-
corpus is very expensive, time-consuming and requires hu-
duces a new cross-language dialog state tracking scenario,
n man experts. Not to mentionthe collection of a new corpus
a wheretheparticipantsareaskedtobuildtheirtrackersbased
foreachlanguageotherthanEnglishifweneedtobuildtrack-
J ontheEnglishtrainingcorpus,whileevaluatingthemwiththe
ersforanewlanguage.
3 unlabeledChinesecorpus. Althoughthecomputer-generated
TheDSTC5proposedanewchallengebasedonusingthe
2 translationsforbothEnglishandChinesecorpusareprovided
rapidlyadvancingmachinetranslation(MT)technology,one
in the dataset, these translations contain errors and careless
] maybeabletoadaptthebuilttrackertoanewlanguagewith
L useofthemcaneasilyhurttheperformanceofthebuilttrack-
limitedtrainingordevelopmentcorpusinthatlanguage. We
C ers.Toaddressthisproblem,weproposeamultichannelCon-
find this idea very attractive because not only it can reduce
. volutionalNeuralNetworks(CNN)architecture,inwhichwe
s the costofnew languageadaptation,butalso it providesthe
treat English and Chinese language as different input chan-
c
possibility of building a tracker with cross-languagecorpus.
[ nelsofonesingleCNN model. IntheevaluationofDSTC5,
For example it can be very useful for developingthe tourist
wefoundthatsuchmultichannelarchitecturecaneffectively
1 informationsystemsbecauseonemayhavecorpuscollected
v improvethe robustness against translation errors. Addition-
fromdifferentlanguagespeakers(i.e. touristsfromdifferent
7 ally,ourmethodforDSTC5ispurelymachinelearningbased
countries): for each language,the amountof corpusmay be
4 and requires no prior knowledge about the target language.
2 very limited, but together it can be large enough for a good
We consider this a desirable property for building a tracker
6 training.Ontheotherhand,althoughthemachinetranslation
inthecross-languagecontext,asnoteverydeveloperwillbe
0 technologyhasachievedgreatprogressrecently,the transla-
. familiarwithbothlanguages.
1 tionqualityisstillnotsatisfactory[4]. Aconventionalmono-
0 Index Terms— Convolutional neural networks, multi- lingualtrackertrainedonthecomputer-generatedtranslations
7 channelarchitecture,dialogstatetracking,dialogsystems may lead to an imperfect model, and it can only accept the
1
translationsfromotherlanguagesasinputwhichwillalsode-
:
v gradetheperformance.
1. INTRODUCTION
i To addressthese problems,we proposea modelthatcan
X
be trained with different languages at the same time, and
r Dialog state tracking is one of key sub-tasksof dialogman-
a agement, whose goal is to transfer human utterances into a use both original utterances and their translations as input
source for the dialog state tracking. In such way, we can
slot-value representation (dialog state) that is easy for com-
avoidbuildingthetrackeronlybasedoncomputer-generated
putertoprocessandtracktheinformationthatappearedinthe
translations, and maximize the use of all possible input lan-
dialog. Toprovideacommontestbedforthistask,theseries
guages to increase the robustness to translation errors. This
of Dialog State Tracking Challenges (DSTC) was initiated
paper is organized as follows. Sect.2 briefly describes the
[1]. Thischallengehasalreadybeenheldforfourtimes,dur-
datasetandthedialogstatetrackingproblem;Sect.3presents
ingwhichitprovidedaveryvaluablesharedrecourseforthe
an overviewof our methodand explainsin detailour multi-
researchin this field and helped to improvethe state-of-the-
channel CNN model. Sect.4 presents the evaluation results
art. Sincetheforthchallenge(DSTC42015),thetargetofdi-
withanalysisanddiscussion. Sect.5concludesourworkand
alogstatetrackinghasbeenshiftedfromhuman-machinedia-
proposesfutureimprovements.
logtohuman-humandialog,whichsignificantlyincreasesthe
difficultyofthedialogstatetrackingtaskbecauseofthevari-
etyandambiguityinthehuman-humandialog.Onelessonwe 2. DATASETANDPROBLEMDESCRIPTION
learnedfromDSTC4 isthe difficultyof buildinga highper-
formancetrackerforhuman-humandialogwithverylimited The fifth Dialog State Tracking Challenge (DSTC5) uses
trainingcorpus,nomatterwhetherusingmachinelearningor thewholedataset(includingtrain/dev/testdatasets)fromthe
Copyright2016IEEE.Publishedinthe2016IEEEWorkshoponSpokenLanguageTechnology(SLT2016).
3. METHOD
Table1. Exampleofatranscriptionanddialogstatelabelfor
asub-dialogsegmentinthetopicof‘Accommodation’.
InDSTC4weproposedamethodwhichisbasedonthecon-
volutional neural networks originally proposed by Kim [5].
Transcription Dialogstatelabel
Guide:Let’strythisone,okay? Bythismethodwewereabletoachievethebestperformance
Tourist:Okay. fortrackingtheINFO slot. TheCNNmodelwe usedinthis
Guide:It’sInnCrowdBackpackersHostelinSingapore.If
methodwasmodifiedfromtheoriginbyaddingastructureof
youtakeadormbedperpersononlytwentydollars.Ifyou
takearoom,it’stwosinglebedsatfiftyninedollars INFO:Pricerange multi-topicconvolutionallayer,sothatitcanbetterhandlethe
Tourist:Um.Wow,that’sgood.
informationpresentedindifferentdialogtopics. Thismodel
Guide: Yah,thepricesarebasedonperpersonperbedor PLACE:
dorm.Butthisoneisroom.Soitshouldbefiftynineforthe InnCrowd Back- is characterizedby its high performancefor limited training
tworoom.Soyou’reactuallypayingabouttendollarsmore packersHostel data, because it can be trained across various topics. More
perpersononly.
Tourist: Ohokay. That’s-thepriceisreasonableactually. detailsaboutthismulti-topicmodelcanbefoundin[2].
It’sgood. In DSTC5 the training data is 75% more than DSTC4,
thereforethesituationoflimitedtrainingdataisimproved.In
orderto focusmoreon thenewcross-languageproblemand
DSTC4 as the training dataset. This dataset contains 35 di- keepour methodsimple, instead ofusing the more complex
alog sessions on tourist information for Singapore collected multi-topicmodelweproposedlasttime,wetrainedindivid-
ualCNNmodelforeachslot-topiccombination. Thatis,for
from English speakers. Besides the training dataset, a de-
velopment set which includes 2 dialog sessions collected examplethe‘INFO’slotinthetopicof‘FOOD’andthesame
fromChinesespeakersisprovidedfortestingandtuningthe ‘INFO’ slot in the topic of ‘SHOPPING’ will be trained by
two independentmodels. This is the major difference from
trackers’ cross-language performance before the final eval-
uation. Both the training and developmentsets are labelled thelast time, wherewe trainedonesinglemodelforall top-
ics. With this new scheme, we can set the hyperparameters
with the dialog state tags and come with 5-best hypothesis
of English or Chinese translations by a machine translation in each model for each slot/topic to be exactly the same, so
systems. In the evaluationphase ofthe challenge, a Test set thatourmethodisscalable,universallyapplicableandeasyto
tune.Fig1isasimplediagramillustratingourmethod.
including 8 unlabeled Chinese dialogs is distributed to each
participant,andall predictionresultssubmittedbyeach par-
ticipantareevaluatedbycomparingwiththetruelabels. The (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)
(cid:7)(cid:8)(cid:9)(cid:10)(cid:11)(cid:10)(cid:11)(cid:12)(cid:6)(cid:13)(cid:14)(cid:8)(cid:15)(cid:16)(cid:17)
test dataset also includes 5-best English translations which
aregeneratedbythesamemachinetranslationsystemasthe (cid:20)(cid:10)(cid:23)(cid:7)(cid:24)(cid:8)(cid:24)(cid:25)(cid:6)(cid:26)(cid:27)(cid:6)(cid:7)(cid:14)(cid:15)(cid:10)(cid:13)(cid:6)(cid:19)(cid:20)(cid:21)(cid:21)(cid:1)(cid:22)
training/developmentdataset.
(cid:3)(cid:8)(cid:9)(cid:10)(cid:11)(cid:10)(cid:11)(cid:12)(cid:6)(cid:13)(cid:14)(cid:8)(cid:15)(cid:16)(cid:17)
The dialog state in this challengeis defined by the same (cid:10)(cid:11)(cid:6)(cid:7)(cid:14)(cid:15)(cid:10)(cid:13)(cid:18)(cid:6)(cid:19)(cid:20)(cid:21)(cid:21)(cid:1)(cid:22)
ontology used in DSTC4, which contains 5 topic branches
(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)
with different slot sets (Table2). These topic-slot combina-
tions indicate the most important information mentioned in
(cid:28)(cid:16)(cid:23)(cid:7)(cid:10)(cid:13)(cid:29)(cid:9)(cid:11)(cid:11)(cid:24)(cid:23)(cid:6)
thattopic,forexamplethe‘CUISINE’slotunderthetopicof (cid:4)(cid:30)(cid:30)(cid:6)(cid:28)(cid:14)(cid:25)(cid:24)(cid:23)
(cid:31)(cid:14)(cid:8)(cid:6)(cid:17)(cid:23)(cid:14)(cid:7)(cid:18)(cid:6)(cid:19) (cid:30)(cid:20)(cid:21)(cid:22)
‘FOOD’referstothecuisinetypes,whilethe‘STATION’slot
(cid:28)(cid:16)(cid:23)(cid:7)(cid:10)(cid:13)(cid:29)(cid:9)(cid:11)(cid:11)(cid:24)(cid:23) (cid:19) (cid:30)(cid:20)(cid:21)(cid:22)(cid:6)(cid:18)(cid:6)
forthetopicof‘TRANSPORTATION’referstothetrainsta- (cid:4)(cid:30)(cid:30)(cid:6)(cid:28)(cid:14)(cid:25)(cid:24)(cid:23) %&(cid:9)(cid:23)(cid:16)(cid:24)'((cid:6)&(cid:9)(cid:23)(cid:16)(cid:24))((cid:6)*+
tions. Intotalthereare30suchtopic-slotcombinations,and !(cid:7)(cid:7)(cid:24)(cid:8)(cid:9)(cid:11)(cid:13)(cid:24)(cid:17)$(cid:10)(cid:7)(cid:29)(cid:6) (cid:31)(cid:14)(cid:8)(cid:6)(cid:17)(cid:23)(cid:14)(cid:7)(cid:18)(cid:6)(cid:19)(cid:4)! (cid:2) (cid:30)"(cid:22) (cid:19)(cid:4)! (cid:2) (cid:30)"(cid:22)(cid:6)(cid:18)(cid:6)
(cid:7)(cid:8)(cid:9)(cid:11)(cid:17)(cid:23)(cid:9)(cid:7)(cid:10)(cid:14)(cid:11)(cid:17) %&(cid:9)(cid:23)(cid:16)(cid:24),((cid:6)&(cid:9)(cid:23)(cid:16)(cid:24)-((cid:6)*+
allpossiblevaluesforeachtopic-slotaregivenasalistinthe (cid:10)(cid:11)(cid:6)(cid:7)(cid:14)(cid:15)(cid:10)(cid:13)(cid:18)(cid:6)(cid:19)(cid:20)(cid:21)(cid:21)(cid:1)(cid:22) (cid:1)
(cid:1)
(cid:19)(cid:1) (cid:2)#(cid:22)(cid:6)(cid:18)
ontology. The main task of DSTC5 is to predict the proper
%&(cid:9)(cid:23)(cid:16)(cid:24)(cid:5)((cid:6)&(cid:9)(cid:23)(cid:16)(cid:24).((cid:6)*+
value(s)foreachslotgiventhecurrentutteranceanditstopic,
(cid:28)(cid:16)(cid:23)(cid:7)(cid:10)(cid:13)(cid:29)(cid:9)(cid:11)(cid:11)(cid:24)(cid:23)
withalldialoghistorypriortothatturn(e.g.Table1). (cid:4)(cid:30)(cid:30)(cid:6)(cid:28)(cid:14)(cid:25)(cid:24)(cid:23)
(cid:31)(cid:14)(cid:8)(cid:6)(cid:17)(cid:23)(cid:14)(cid:7)(cid:18)(cid:6)(cid:19)(cid:1) (cid:2)#(cid:22)
(cid:7)(cid:8)(cid:9)(cid:10)(cid:3)(cid:11)(cid:12)(cid:5)(cid:2)(cid:13)(cid:14)(cid:6)(cid:4)(cid:13)(cid:15)(cid:4)(cid:16)(cid:17) (cid:21)(cid:22)(cid:4)(cid:3)(cid:11)(cid:23)(cid:17)(cid:4)(cid:3)(cid:14)(cid:5)(cid:12)(cid:9)(cid:4)(cid:5)(cid:6)
Table2. Listofslotsforeachtopic. (cid:18)(cid:17)(cid:4)(cid:6)(cid:17)(cid:14)(cid:11)(cid:16)(cid:19)(cid:8)(cid:17)(cid:20) (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6) (cid:18)(cid:2)(cid:8)(cid:17)(cid:19)(cid:8)(cid:17)(cid:20)
Topic SLOT
Food INFO, CUISINE, TYPEOFPLACE, DRINK, PLACE,
MEALTIME,DISH,NEIGHBOURHOOD Fig.1. Overviewofourmethod(forthe‘FOOD’topic).
Attraction INFO, TYPEOFPLACE, ACTIVITY, PLACE, TIME,
NEIGHBOURHOOD
Shopping INFO, TYPEOFPLACE, PLACE, NEIGHBOURHOOD,
TIME
Transportation INFO,FROM,TO,STATION,LINE,TYPE,TICKET 3.1. Motivation
Accommodation INFO,TYPEOFPLACE,PLACE,NEIGHBOURHOOD
ThebiggestchallengeofDSTC5isthatthetrainingandtest
corporaareoriginallycollectedindifferentlanguages. Since
bothcomputer-generatedChineseandEnglishtranslationare whereSisthesigmoidfunction;⊕istheconcatenationoper-
provided in the training and test dataset respectively, one atorandhˆchn = (hˆ1,...,ˆhm)isthepenultimatelayerofthe
straightforward approach is to train a model with English n-thchannel.
corpusand use it for the English translation in the test data.
NoticethatintheoriginalpaperofKim,amulti-channel
Alternatively,a modeltrainedon Chinese translationsin the
architecturehasalsobeenproposed.Themaindifferencebe-
training dataset can be used for Chinese utterances in the
tweenourmodelandtheirmodelisthatweusedifferentsets
test data. However, both methods will waste the originally
offiltersforeachchannel,whileintheirmodelthesamefilter
collected utterances either in the training or the test data.
setareappliedtoallchannels. Thereasonforthismodifica-
In order to fully utilize the corpus resource in both English
tion is that the word embedding for different languages can
and Chinese languages, we proposed the following multi-
channel model which can be regarded as a combination of
bothEnglishandChinesemodels.
3.2. Modelarchitecture (cid:1)(cid:2)(cid:3)(cid:4)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)
Ourmodelisinspiredbythemultichannelconvolutionalneu-
ral networks commonly used in the image processing [6]. (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:9)(cid:10)(cid:11)(cid:12)
Instead of RGB channels used for color images, we apply (cid:9)(cid:3)(cid:6)(cid:10)(cid:5)(cid:8)(cid:7)(cid:11)(cid:7)(cid:12)(cid:13)(cid:8)
(cid:9)(cid:3)(cid:6)(cid:10)(cid:5)(cid:14)(cid:7)(cid:11)(cid:7)(cid:12)(cid:13)(cid:14)
eachinputchanneltoeachdifferentlanguagesource. Inthis (cid:9)(cid:3)(cid:6)(cid:10)(cid:5)(cid:15)(cid:7)(cid:11)(cid:7)(cid:12)(cid:13)(cid:16)
(cid:1)(cid:2)(cid:3)(cid:4)(cid:4)(cid:5)(cid:6)(cid:7)(cid:14)
(cid:9)(cid:3)(cid:6)(cid:10)(cid:5)(cid:17)(cid:7)(cid:11)(cid:7)(cid:12)(cid:13)(cid:18)
model,theinputofeachchannelisatwodimensionalmatrix,
(cid:9)(cid:3)(cid:6)(cid:10)(cid:5)(cid:19)(cid:7)(cid:11)(cid:7)(cid:12)(cid:13)(cid:15)
eachrowofwhichistheembeddingvectorofthecorrespond- (cid:9)(cid:3)(cid:6)(cid:10)(cid:5)(cid:20)(cid:7)(cid:11)(cid:7)(cid:12)(cid:13)(cid:8)
(cid:9)(cid:3)(cid:6)(cid:10)(cid:5)(cid:21)(cid:7)(cid:11)(cid:7)(cid:12)(cid:13)(cid:14)
ingword:
—w —
1
 
—w —
2 (cid:1)(cid:2)(cid:3)(cid:4)(cid:4)(cid:5)(cid:6)(cid:7)(cid:15)
s= .. , (1)
 . 
 
 —wn— 
wherewi ∈ Rk istheembeddingvectorforthei-thwordin Input layer: Convolutional Max-pooling Fully-connected
embedding matrix layer layer sigmoid layer
the inputtext. This 2-dimensionalarray s is a matrix repre-
sentationoftheinputtext. Weusedthreedifferentwordem-
beddinginourmodel—twoforChineseandoneforEnglish. Fig.2. MultichannelCNNmodelarchitectureforthreeinput
Thedetailsoftheseembeddingwillbeexplainedlaterinthe channels.
Sect. 3.3. For each channel, a feature map h ∈ Rn−d+1 is
obtainedbyconvolvingatrainablefilterm ∈ Rd×k withthe (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:3)(cid:6)(cid:7)(cid:8)(cid:7)(cid:8)(cid:9)(cid:4)(cid:10)(cid:2)(cid:3)(cid:11)(cid:12)(cid:13)(cid:4)(cid:14)(cid:15)(cid:8)(cid:9)(cid:16)(cid:7)(cid:13)(cid:17)(cid:18)(cid:19)
embeddingmatrixs∈Rn×k usingthefollowingequation:
(cid:12)(cid:7)(cid:5)(cid:9)(cid:10)(cid:16)(cid:10)(cid:11)(cid:17)(cid:14)(cid:18)(cid:15)
(cid:10)(cid:19)(cid:20)(cid:10)(cid:15)(cid:15)(cid:5)(cid:9)(cid:21)(cid:11)(cid:19)(cid:14)(cid:15)(cid:10)(cid:3)
h=f(m∗s+b). (2) (cid:22)(cid:9)(cid:21)(cid:3)(cid:5)(cid:16)(cid:7)(cid:11)(cid:4)(cid:14)(cid:11)(cid:12)(cid:7)(cid:5)(cid:9)(cid:10)(cid:16)(cid:10)
(cid:1)(cid:23)(cid:11)(cid:16)(cid:24)(cid:16)(cid:4)(cid:10)(cid:19)
(cid:12)(cid:7)(cid:5)(cid:9)(cid:10)(cid:16)(cid:10)(cid:11)(cid:6)(cid:7)(cid:8)(cid:18)(cid:8)(cid:6)(cid:4)(cid:10)(cid:18) (cid:1)(cid:2)(cid:3) (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:9)(cid:9)(cid:10)(cid:3)(cid:11)
Heref isanon-linearactivationfunction1;∗istheconvolu- (cid:10)(cid:19)(cid:20)(cid:10)(cid:15)(cid:15)(cid:5)(cid:9)(cid:21)(cid:11)(cid:19)(cid:14)(cid:15)(cid:10)(cid:3) (cid:1)(cid:2)(cid:4) (cid:12)(cid:13)(cid:13)(cid:11)(cid:1)(cid:14)(cid:15)(cid:10)(cid:3)
tion operator and b = (b,...,b) ∈ Rn is a bias term. The (cid:1)(cid:2)(cid:5)
maximum value of this feature map hˆ = max{h} is then (cid:22)(cid:9)(cid:21)(cid:3)(cid:5)(cid:16)(cid:7)(cid:11)(cid:2)(cid:4)(cid:4)(cid:10)(cid:18)(cid:8)(cid:9)(cid:6)(cid:10)(cid:16) (cid:22)(cid:9)(cid:21)(cid:3)(cid:5)(cid:16)(cid:7)(cid:11)(cid:17)(cid:14)(cid:18)(cid:15)
(cid:10)(cid:19)(cid:20)(cid:10)(cid:15)(cid:15)(cid:5)(cid:9)(cid:21)(cid:11)(cid:19)(cid:14)(cid:15)(cid:10)(cid:3)
selected by the max-pooling layer. This is the process how
onefilter extractsonemostimportantfeaturefromthe input (cid:25)(cid:18)(cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:15)(cid:11)(cid:5)(cid:9)(cid:27)(cid:2)(cid:4)
matrix. Inthismodel,multiplefiltersareusedineachchan-
(cid:1)(cid:2)(cid:3)(cid:4)(cid:20)(cid:21)(cid:22)(cid:21)(cid:16)(cid:2)(cid:11)(cid:23)(cid:21)(cid:8)(cid:5)(cid:24)(cid:5)(cid:21)(cid:13)(cid:5)(cid:4)(cid:10)(cid:2)(cid:3)(cid:11)(cid:12)(cid:13)(cid:4)(cid:14)(cid:25)(cid:17)(cid:7)(cid:8)(cid:21)(cid:13)(cid:21)(cid:18)(cid:19)
neltoextractmultiplefeatures. Thesefeaturesthenformthe
poolinglayerwhicharepassedtoafullyconnectedlayerfor (cid:12)(cid:7)(cid:5)(cid:9)(cid:10)(cid:16)(cid:10)(cid:11)(cid:17)(cid:14)(cid:18)(cid:15)
(cid:10)(cid:19)(cid:20)(cid:10)(cid:15)(cid:15)(cid:5)(cid:9)(cid:21)(cid:11)(cid:19)(cid:14)(cid:15)(cid:10)(cid:3)
prediction.
(cid:12)(cid:7)(cid:5)(cid:9)(cid:10)(cid:16)(cid:10)(cid:11)(cid:2)(cid:4)(cid:4)(cid:10)(cid:18)(cid:8)(cid:9)(cid:6)(cid:10)(cid:16)
The idea of multichannel model is to connect those ex-
(cid:12)(cid:7)(cid:5)(cid:9)(cid:10)(cid:16)(cid:10)(cid:11)(cid:6)(cid:7)(cid:8)(cid:18)(cid:8)(cid:6)(cid:4)(cid:10)(cid:18) (cid:1)(cid:2)(cid:3) (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:9)(cid:9)(cid:10)(cid:3)(cid:11)
tracted featuresfrom differentchannelsbeforethe final out- (cid:10)(cid:19)(cid:20)(cid:10)(cid:15)(cid:15)(cid:5)(cid:9)(cid:21)(cid:11)(cid:19)(cid:14)(cid:15)(cid:10)(cid:3) (cid:1)(cid:2)(cid:4) (cid:12)(cid:13)(cid:13)(cid:11)(cid:1)(cid:14)(cid:15)(cid:10)(cid:3)
put, so that the model can use richer information obtained (cid:1)(cid:2)(cid:5)
from different channels. The fully connect layer in multi- (cid:12)(cid:7)(cid:5)(cid:9)(cid:10)(cid:16)(cid:10)(cid:11)(cid:4)(cid:14)(cid:11)(cid:22)(cid:9)(cid:21)(cid:3)(cid:5)(cid:16)(cid:7) (cid:22)(cid:9)(cid:21)(cid:3)(cid:5)(cid:16)(cid:7)(cid:11)(cid:17)(cid:14)(cid:18)(cid:15)
(cid:1)(cid:23)(cid:11)(cid:16)(cid:24)(cid:16)(cid:4)(cid:10)(cid:19) (cid:10)(cid:19)(cid:20)(cid:10)(cid:15)(cid:15)(cid:5)(cid:9)(cid:21)(cid:11)(cid:19)(cid:14)(cid:15)(cid:10)(cid:3)
channelmodelfollowstheequation:
(cid:25)(cid:18)(cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:15)(cid:11)(cid:5)(cid:9)(cid:27)(cid:2)(cid:4)
y=S w·(hˆch1⊕hˆch2⊕hˆch3)+b , (3)
(cid:0) (cid:1)
1Weusedrectifiedlinearunit(ReLU)forthisactivationfunction. Fig.3. Pre-processingoftheinpututterances.
varygreatly, forexamplethe same (or nearlythe same) em- (0.0956/0.0635)inAccuracyand15%(0.4519/0.3945)inF-
beddingvectorindifferentlanguagemodelsmaycorrespond measure with the sub-dialog evaluation. Our submitted five
toirrelevantwordswith verydifferentmeanings. Usingdif- entriesaretheresultsof5differenthyperparameterssettings
ferent sets of filters ensures that proper features can be ex- whicharedeterminedbyaroughgridsearch3,andthoseset-
tracted in each channel no matter how the word embedding tingsaresummarizedinTable4. Comparedtheseresultswith
variesamongdifferentlanguages. eachother,onecaneasilytellthatamongthesehyperparam-
eters the dropoutrate is a key factor. The dropoutis known
asatechniqueforreducingoverfittinginneuralnetworks[9],
3.3. Embeddingmodels
andinourcasereducingthedropoutratealwaysimprovesthe
The word2vec [7] is one of the most common methods for PrecisionwhiledegradingtheRecallscore. Oneexplanation
producing word embeddings. In DSTC5, we applied this forthisisthatanover-fittedmodelonlyoutputsthesamela-
method and trained three different models with different bels for the data which are very similar to the training data,
training corpus. The details of these models are listed as andthereforedecreasesitsgeneralizationtounseendata. On
below: the other hand, further decreasing the dropout rate does not
improve the overall performance, whose results and param-
1. English wordmodel: 200-dimensionword2vecmodel
eter settingsarealso shownin the table as‘AdditionalExpt.
trained on English Wikipedia, with all text split by
#5&6’.
space and all letters lowercased. This model contains
253854Englishwords.
2. Chinesewordmodel: 200-dimensionword2vecmodel Table 3. Evaluationresults(subdialog-level)onDSTC5test
trained on Chinese Wikipedia, with all text split by dataset.4
wordboundaryusing‘jieba’module2. Thismodelcon-
Tracker Accuracy Precision Recall F-measure
tains457806Chinesewordsand53743Englishwords
Multichannel#3 0.0956 0.5643 0.3769 0.4519
appearedintheChineseWikipedia. Multichannel#4 0.0872 0.5427 0.3842 0.4499
Multichannel#0 0.0964 0.5217 0.3849 0.4430
3. Chinese character model: 200-dimension word2vec Multichannel#1 0.0712 0.4340 0.4196 0.4267
Multichannel#2 0.0681 0.4216 0.4303 0.4259
modeltrainedonChineseWikipediawithalltextsplit
Team4-entry2 0.0635 0.3768 0.4140 0.3945
into single Chinese character. This model contains
Team1-entry4 0.0612 0.3811 0.3548 0.3675
12145 Chinese characters and 53743 English words Team6-entry1 0.0383 0.4063 0.3124 0.3532
appearedintheChineseWikipedia. Team5-entry0 0.0520 0.3637 0.3044 0.3314
baseline2 0.0222 0.1979 0.1774 0.1871
The reason why we trained two modelsforChinese lan- AdditionalExpt#5 0.0949 0.5786 0.3689 0.4505
AdditionalExpt#6 0.0888 0.5677 0.3712 0.4489
guage is because identifying word boundaries in Chinese is
not a trivial task. For Chinese, the smallest element with
meaning(word)variesfromone single Chinese characterto
severalconcatenatedChinesecharacters,andthetaskforChi-
nesewordsplittingusuallyinvolvesparsingthesentenceand
the state-of-the-artmethodstill cannotachieveperfectaccu- Table4. Mainhyperparametersettingsforeachentry.
racy. For this reason the Chinese word model may contain Entry#inTable3 #3 #4 #0 #1 #2 #5 #6
incorrectvocabulariesandisnotcapableofhandlingunseen Dropoutrate 0.4 0.5 0.6 0.8 0.8 0.2 0
Number of filters m ∈
Chinesecharactercombinations. Ontheotherhand,theChi- R1×kforeachchannel 1000 600 1000 1400 1000 1000 1000
nesecharactermodeldoesnotrelyonwordsegmentationso Number of filters m ∈
that the model is error-free, and also it can easily deal with R2×kforeachchannel 1000 600 1000 1400 1000 1000 1000
Learningrate 0.001
unseen words. However, since the Chinese character model
Weightdecaycoefficient
0.0005
ignoresthewordboundaries,theresultingembeddingvector forL2regularization
maynotbeabletoreflecttheprecisemeaningofeachword. Trainingdata Trainingcorpus+1-bestChinesetranslation
&Developmentcorpus+1-bestEnglishtranslation
Trainingepochs 100
Testinput Testcorpus+1-bestEnglishtranslation
4. RESULTS
The results of the proposed method along with the scores
of other teams are shown in the Table 3. Our multichannel
CNN model achieves the best score among all 9 teams: the
result of entry-3 outperforms the second best team by 50% 3Aguideforsettingthesehyperparameterscanbefoundin[8]
4Thesearetheevaluationresultsusing‘Schedule2’describedinthechal-
2https://github.com/fxsjy/jieba lengehandbook[1].
Table5. Exampleofpredictedlabelsbydifferentmodels.
Chinese character Multichannel model
Transcription (translation) Chinese word model English word model Model combination
model (= True label)
INFO: Exhibit
INFO: Exhibit
Guide: (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)们(cid:9)(cid:10)(cid:11)(cid:12)(cid:13)(cid:14)(cid:5)(cid:15)
PLACE: Singapore
(you can walk to the botanical garden where we INFO: Exhibit
PLACE: Singapore Botanic Gardens, PLACE: Singapore
go for a walk .)
PLACE: Singapore Botanic Gardens, Bukit TimahNature Botanic Gardens
(cid:16)(cid:17)赏(cid:18)(cid:19)(or the flowers .) PLACE: Singapore
Botanic Gardens Bukit TimahNature Reserve,
Botanic Gardens,
Reserve, Orchard Road, ACTIVITY: Walking
Tourist: (cid:20)(cid:19)(er.) National Orchid Garden
HortPark, National Orchid Garden
(cid:20)(cid:10)(cid:11)-(er, plants)
National Orchid Garden
ACTIVITY: Walking
4.1. Multichannelmodel&singlechannelmodel&model exampleisdescribedusingdifferentfeaturesetsthatprovide
combination different,complementaryinformationabouttheinstance.The
fullyconnectedlayerin the multichannelmodelfurtherpro-
Toinvestigatebyhowmuchtheproposedmultichannelarchi-
videsan optimizationto use this informationfor the predic-
tecture contributesto these results, we comparedthe perfor-
tion,andthereforetheresultingmodelcaninprinciplebetter
mancebetweenthemultichannelandordinarysinglechannel
dealwiththetranslationerrorsappearedindifferentchannels.
CNN models. For this comparison, we trained three differ-
Table5isoneoftheexamplesthatdemonstratethisidea.
ent monolingual single channel CNN models using each of
In this particular sub-dialog segment, none of the 3 single
theembeddingmodelsmentionedinSection3.3. Thesemod-
channelmodelsisabletooutputthecorrectlabels,whilethe
els used the same parameter setting as ‘multichannel#3’ in
multichannelmodelgivesthe correctprediction. As seen in
the Table 4, and were trained only on the 1-best machine
this example, the model combination behaves like a simple
translation results. Fig.4 shows the comparing results: the
voting,whichmeansitonlypicksupthelabelsthataresup-
Chinese character model achieves the best overall accuracy
portedbymajorityofthesinglechannelmodels. Themulti-
amongsinglechannelmodels,whilethemultichannelmodel
channelmodel,ontheotherhand,isabletoselectivelychoose
outperformsallthreesinglechannelmodels.
which languagesource to trust more for each particular slot
In the earlier DSTC, a simple model combination tech-
orvalue. Asaresult,thelabelof‘Walking’iscorrectlypre-
niquehasbeenusedtofurtherimprovethepredictiveperfor-
dicteddespiteitonlyappearingintheEnglishmodel’soutput,
mance, where the final outputis computedby averagingthe
whilethe‘Exhibit’labeliscorrectlyrejectedeventhoughitis
scoresoutputbydifferentmodels[10]. We also appliedthis
supportedbytwosinglechannelmodelsoutofthree.
method to combine the output from all three single channel
However the real situation is more complex. When we
models,andtheresultisalsoshownintheFig.4. Thissimple
look at the overall predictive accuracy for each slot (Fig.5),
modelcombinationmethoddoesnotperformas goodas the
we can find that the performance for each model varies on
multichannel model, but considering its simplicity, we still
slots. We consider this is due to the ambiguity caused by
consideritasagoodalternativetoimprovetheperformance.
machine translation which varies on different subjects. For
example, as a time expression in English, 96% of the word
“evening” and 43% of the word “night” are translated into
(cid:0)(cid:6)(cid:11)(cid:7)
(cid:0)(cid:6)(cid:11)
(cid:0)(cid:6)(cid:0)(cid:10)
(cid:0)(cid:6)(cid:0)(cid:9)
(cid:0)(cid:6)(cid:0)(cid:8)
(cid:0)(cid:6)(cid:0)(cid:7) !&
(cid:0) !%
(cid:12)(cid:13)(cid:14)(cid:15)(cid:16)(cid:17)(cid:16) (cid:12)(cid:13)(cid:14)(cid:15)(cid:16)(cid:17)(cid:16)(cid:27)(cid:25)(cid:20)(cid:26) (cid:28)(cid:15)(cid:22)(cid:23)(cid:14)(cid:17)(cid:13)(cid:27)(cid:25)(cid:20)(cid:26) (cid:29)(cid:25)(cid:26)(cid:16)(cid:23) (cid:29)(cid:31)(cid:23)(cid:21)(cid:14)(cid:18)(cid:13)(cid:19)(cid:15)(cid:15)(cid:16)(cid:23) !$
(cid:18)(cid:13)(cid:19)(cid:20)(cid:19)(cid:18)(cid:21)(cid:16)(cid:20)(cid:17)(cid:14)(cid:15)(cid:22)(cid:23)(cid:16)(cid:17)(cid:14)(cid:15)(cid:22)(cid:23)(cid:16)(cid:18)(cid:13)(cid:19)(cid:15)(cid:15)(cid:16)(cid:23) (cid:17)(cid:14)(cid:15)(cid:22)(cid:23)(cid:16)(cid:18)(cid:13)(cid:19)(cid:15)(cid:15)(cid:16)(cid:23) (cid:18)(cid:25)(cid:24)(cid:30)(cid:14)(cid:15)(cid:19)(cid:21)(cid:14)(cid:25)(cid:15) (cid:24)(cid:25)(cid:26)(cid:16)(cid:23) !#
(cid:18)(cid:13)(cid:19)(cid:15)(cid:15)(cid:16)(cid:23)(cid:24)(cid:25)(cid:26)(cid:16)(cid:23) (cid:24)(cid:25)(cid:26)(cid:16)(cid:23) (cid:24)(cid:25)(cid:26)(cid:16)(cid:23) !"

Fig.4. Overallpredictiveaccuracyofdifferentmodels.
4.2. Discussion ’()*+,+-(./.-0+/123+4 ’()*+,+52/3123+4
6*74),(52/3123+4 123+4’289)*.0)2*
1:40)-(.**+4123+4
Wethinktheaboveresultscanbepartiallyexplainedfromthe
pointofviewofensemblelearning.Inamultichannelmodel, Fig.5. Comparisonof accuracyforeach slotin the topicof
each channel provides a different view of the data, and an ‘Attraction’.
thesameChineseword“wanshang”. AlthoughthisChinese Koichiro Yoshino, “The Fifth Dialog State Tracking
word does have both meanings of “evening” and “night” in Challenge,” inProceedingsofthe2016IEEEWorkshop
Chinese, there are more precise Chinese terms representing onSpokenLanguageTechnology(SLT),2016.
eachword.ThisEnglishtoChinesetranslationambiguityim-
[2] HongjieShi,TakashiUshio,MitsuruEndo,Katsuyoshi
mediatelyincreasesthedifficultyofidentifyingthevaluesof
Yamagami, and Noriaki Horii, “Convolutional neural
EVENING and NIGHT in Chinese, which leads to the poor
networks for multi-topic dialog state tracking,” Pro-
performanceoftheChinesemodelintheslotofTIME.
ceedings of the 7th International Workshop on Spoken
Anotherproblemisthatthetranslationqualityoftenvaries
DialogueSystems(IWSDS),2016.
byreversingthetranslationdirection,duetothedifferencein
inflections, word order and grammars [11]. Since the train- [3] Franck Dernoncourt, Ji Young Lee, Trung H Bui, and
ingcorpusonlycontainsonetranslationdirection(Englishto Hung H Bui, “Robust dialog state tracking for large
Chinese), the multichannelmodelis by no meansoptimized ontologies,” arXivpreprintarXiv:1605.02130,2016.
forthereversetranslationdirection. Thismaycausethemul-
[4] Ilya Sutskever, Oriol Vinyals, and Quoc V Le, “Se-
tichannelmodelto have bias on certain channels, and it can
quence to sequence learning with neural networks,”
explainwhyincertainslotsthemodelcombinationthattreats
in Advances in neural information processing systems,
eachchannelequallyworksbetter.Amoresophisticatedway
2014,pp.3104–3112.
totrainourmultichannelmodelshouldbefirstlytrainingthe
modelwithonetranslationdirectionandthenfine-tuningthe [5] YoonKim,“Convolutionalneuralnetworksforsentence
modelwiththeother.UnfortunatelythisisdifficultinDSTC5 classification,” arXivpreprintarXiv:1408.5882,2014.
becausethedevelopmentdatasetthatcanbeusedforthefine
tuningistoolimited. [6] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hin-
ton, “Imagenet classification with deep convolutional
neural networks,” in Advances in neural information
5. CONCLUSION
processingsystems,2012,pp.1097–1105.
We proposed a multichannel convolutional neural network, [7] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey
inwhichwetreatmultiplelanguagesasdifferentinputchan- Dean, “Efficientestimation of word representationsin
nels. This multichannelmodel is found to be robustagainst vectorspace,” arXivpreprintarXiv:1301.3781,2013.
thetranslationerrorsandoutperformsanyofthesinglechan-
[8] Ye Zhang and Byron Wallace, “A sensitivity analy-
nelmodels. Furthermore,ourmethoddoesnotrequireprior
sis of (and practitioners’ guide to) convolutional neu-
knowledgeaboutnewlanguages,andthereforecanbeeasily
ralnetworksforsentenceclassification,” arXivpreprint
applied to available corpusresourcesof differentlanguages.
arXiv:1510.03820,2015.
This not only can reduce the cost for the adaption to a new
language,butalsooffersthepossibilitytobuildmultilingual
[9] GeoffreyEHinton,NitishSrivastava,AlexKrizhevsky,
dialogstatetrackerswithlarge-scalecross-languagecorpora.
IlyaSutskever,and Ruslan R Salakhutdinov, “Improv-
Inthisworkweappliedthreedifferentembeddingmodels,
ingneuralnetworksbypreventingco-adaptationoffea-
whilethereisonemorewedidnottry—theEnglishcharac- turedetectors,” arXivpreprintarXiv:1207.0580,2012.
termodel. Thereareseveralcharacter-awarelanguagemod-
elsproposedrecently,whicharesuperiorindealingwithsub- [10] MatthewHenderson,BlaiseThomson,andSteveYoung,
word information, rare words and misspelling [12, 13]. We “Word-based dialog state tracking with recurrent neu-
believethatintegratingthemintothemultichannelmodelisa ralnetworks,” inProceedingsofthe15thAnnualMeet-
promisingresearchdirection. ingoftheSpecialInterestGrouponDiscourseandDia-
On the other hand, since our method is purely machine logue(SIGDIAL),2014,pp.292–299.
learningbased,itcannothandleunseenlabelsinthetestdata.
[11] DavidVilar,JiaXu,LuisFernandodHaro,andHermann
Thisisaveryimportantissueespeciallyforalargeontology,
Ney, “Error analysis of statistical machine translation
becauseofthedifficultyinobtaininglargetrainingcorpusthat
output,” inProceedingsofLREC,2006,pp.697–702.
covers all concepts. To overcome this disadvantage, future
workshouldincludecombiningmachinelearningwithother [12] Yoon Kim, Yacine Jernite, David Sontag, and Alexan-
approaches,suchashand-craftrules,dataargumentationand derMRush,“Character-awareneurallanguagemodels,”
soon. arXivpreprintarXiv:1508.06615,2015.
[13] Xiang Zhang, Junbo Zhao, and Yann LeCun,
6. REFERENCES “Character-level convolutional networks for text
classification,” in Advances in Neural Information
[1] Seokhwan Kim, Luis Fernando D’Haro, Rafael E.
ProcessingSystems,2015,pp.649–657.
Banchs, Jason Williams, Matthew Henderson, and