Conditional Chow-Liu tree structures for modeling discrete-valued vector time series

Conditional Chow-Liu Tree Structures for Modeling Discrete-Valued

Vector Time Series

Sergey Kirshner,Padhraic Smyth, School of Information and Computer Science University of California,Irvine

Irvine,CA92697-3425

{skirshne,smyth}@https://www.wendangku.net/doc/8316569407.html,

Andrew W.Robertson

International Research Institute

for Climate Prediction(IRI),

The Earth Institute at Columbia University,

Palisades,NY10964

awr@https://www.wendangku.net/doc/8316569407.html,

Abstract

We consider the problem of modeling

discrete-valued vector time series data us-

ing extensions of Chow-Liu tree models to

capture both dependencies across time and

dependencies across variables.Conditional

Chow-Liu tree models are introduced,as an

extension to standard Chow-Liu trees,for

modeling conditional rather than joint den-

sities.We describe learning algorithms for

such models and show how they can be used

to learn parsimonious representations for the

output distributions in hidden Markov mod-

els.These models are applied to the impor-

tant problem of simulating and forecasting

daily precipitation occurrence for networks of

rain stations.To demonstrate the e?ective-

ness of the models,we compare their perfor-

mance versus a number of alternatives using

historical precipitation data from Southwest-

ern Australia and the Western United States.

We illustrate how the structure and param-

eters of the models can be used to provide

an improved meteorological interpretation of

such data.

1Introduction

In this paper we consider the problem of model-ing discrete-time,discrete-valued,multivariate time-series.For example,consider M time-series where each time-series can take B values.The motivating appli-cation in this paper is modeling of daily binary rainfall data(B=2)for regional spatial networks of M sta-tions(where typically M can vary from10to100). Modeling and prediction of rainfall is an important problem in the atmospheric sciences.A common appli-cation,for example,is simulating realistic daily rain-fall patterns for a90-day season,to be used as input for detailed crop-modeling simulations.A number of statistical methods have been developed for modeling daily rainfall time-series at single stations—?rst-order Markov models and various extensions(also known as “weather generators”)have proven quite e?ective for single-station rainfall modeling in many geographic re-gions.However,there has been less success in devel-oping models for multiple stations that can generate simulations with realistic spatial and temporal corre-lations in rainfall patterns(Wilks and Wilby1999). Direct modeling of the dependence of the M daily ob-servations at time t on the M observations at time t?1requires an exponential number of parameters in M.This is clearly impractical for most values of M of interest.In this paper we look at the use of hid-den Markov models(HMMs)to avoid this problem—an HMM uses a K-valued hidden?rst-order Markov chain to model time-dependence,with the M outputs at time t being conditionally independent of everything else given the current state value at time t.The hid-den state variable in an HMM serves to capture tempo-ral dependence in a low-dimensional manner,i.e.,with O(K2)parameters instead of being exponential in M. From a scienti?c viewpoint,an attractive feature of the HMM is that the hidden states can be interpreted as underlying“weather states”(Hughes et al.1999, Robertson et al.to appear).

Modeling the instantaneous multivariate dependence of the M observations on the state at time t would require B M parameters per state if the full joint dis-tribution were modeled.This in turn would defeat the purpose of using the HMM to reduce the number of parameters.Thus,approximations such as assuming conditional independence(CI)of the M observations are often used in practice(e.g.,Hughes et al.1999), requiring O(KMB)parameters.

While the HMM-CI approach is a useful starting point it su?ers from two well-known disadvantages for an application such as rainfall modeling:(1)the assumed conditional independence of the M outputs on each

other at time t can lead to inadequate characterization of the dependence between the M time-series (e.g.,unrealistic spatial rainfall patterns on a given day),(2)the assumed conditional independence of the M outputs at time t from from the M outputs at time t ?1can lead to inadequate temporal dependence in the M time-series (e.g.,unrealistic occurrences of wet days during dry spells).

In this paper we investigate Chow-Liu tree structures in the context of providing improved,yet tractable,models to address these problems in capturing out-put dependencies for HMMs.We show how Chow-Liu trees can be used to directly capture dependency among the M outputs in multivariate HMMs.We also introduce an extension called conditional Chow-Liu trees to provide a class of dependency models that are well-suited for modeling multivariate time-series data.We illustrate the application of the proposed methods to two large-scale precipitation data sets.

The paper is structured as follows.Section 2formally describes existing models and our extensions.Section 3describes how to perform inference and to learn both the structure and parameters for the models.Section 4describes an application and analyzes the performance of the models.Finally,Section 5summarizes our con-tributions and outlines possible future directions.

2Model Description

We begin this section by brie?y reviewing Chow-Liu trees for multivariate data before introducing the con-ditional Chow-Liu tree model.We then focus on vector time-series data and show how the conditional Chow-Liu tree model and hidden Markov models can be com-bined.2.1

Chow-Liu Trees

Chow and Liu (1968)proposed a method for approxi-mating the joint distribution of a set of discrete vari-ables using products of distributions involving no more than pairs of variables.If P (x )is an M -variate dis-tribution on discrete variables V = x 1,...,x M

,the Chow-Liu method constructs a distribution T (x )for which the corresponding Bayesian and Markov net-work is tree-structured.If G T =(V,E T )is the Markov network associated with T ,then

T (x )= (u,v )∈E T T (x u ,x v )

v ∈V T (x v )degree (v )

= (u,v )∈E T

T (x u ,x v )T (x v )T (x u )

v ∈V

T (x v ).(1)The Kullback-Leibler divergence KL (P,T )between

Algorithm ChowLiu (P )

Inputs:Distribution P over domain V ;procedure MWST(weights )that outputs a maximum weight spanning tree over V

https://www.wendangku.net/doc/8316569407.html,pute marginal distributions P (x u ,x v )and P (x u )?u,v ∈V

https://www.wendangku.net/doc/8316569407.html,pute mutual

information

values

I (x u ,x v

)?u,v ∈V

3.E T =MWST ({I (x u ,x v )})

4.Set T (x u ,x v )=P (x u ,x v )?(u,v )∈E T Output:T

Figure 1:Chow-Liu algorithm (very similar to Meil?a and Jordan 2000)

distributions P and T is de?ned as

KL (P,T )= x

P (x )log P (x )

T (x ).

Chow and Liu showed that in order to minimize

KL (P,T )the edges for the tree E T have to be se-lected to maximize the total mutual information of the edges (u,v )∈E T I (x u ,x v )where mutual information between variables x u and x v is de?ned as

I (x u ,x v )= x u x v P (x u ,x v

)log P (x u ,x v )P (x u )P (x v ).(2)This can be accomplished by calculating mutual infor-mation I (x u ,x v )for all possible pairs of variables in

V ,and then solving the maximum spanning tree prob-lem,with pairwise mutual information from Equation 2as edge weights (e.g.,Cormen et al.1990).Once the edges are selected,the probability distribution T on the pairs of vertices connected by edges is de?ned to be the same as P :

?(x u ,x v )∈E T

T (x u ,x v )=P (x u ,x v ),

and the resulting distribution T minimizes KL (P,T ).

Figure 1outlines the algorithm for ?nding T .If each of the variables in V takes on B values,?nding the tree T has complexity O M 2B 2 for the mutual information calculations and O M 2 for ?nding the minimum spanning tree,totaling O M 2B 2

.(Meil?a (1999)proposed a faster version of the Chow-Liu al-gorithm for sparse high-dimensional data.)In prac-tice,P is often an empirical distribution on the data,so that the calculation of pairwise counts of variables (used in calculating mutual information)has complex-ity O T M 2B 2

where T is the number of vectors in the data.

The advantages of Chow-Liu trees include(a)the ex-istence of a simple algorithm for?nding the optimal tree,(b)the parsimonious nature of the model(the number of parameters is linear in the dimensionality of the space),and(c)the tree structure T can have a simple intuitive interpretation.While there are other algorithms that retain the idea of a tree-structured dis-tribution,while allowing for more complex dependen-cies(e.g.,thin junction trees,Bach and Jordan2002), these algorithms have higher time complexity than the original Chow-Liu algorithm and do not guarantee op-timality within the model class for the structure that is learned.Thus,in the results in this paper we focus on Chow-Liu trees under the assumption that they are

a generally useful modeling technique.

2.2Conditional Chow-Liu Forests

It is common in practice(e.g.,in time-series and in re-gression modeling)that the data to be modeled can be viewed as consisting of two sets of variables,where we wish to model the conditional distribution P(x|y)of one set x on the other set y.We propose an extension of the Chow-Liu method to model such conditional distributions.As with Chow-Liu trees,we want the corresponding probability distribution to be factored into a product of distributions involving no more than two variables.Pairs of variables are represented as an edge in a corresponding graph with nodes correspond-ing to variables in V=V x∪V y.However,since all of the variables in V y are observed,we are not interested in modeling P(y),and do not wish to restrict P(y)by making independence assumptions about the variables in V y.The structure for an approximation distribution T will be constructed by adding edges such as not to introduce paths involving multiple variables from V y. Let G F=(V,E F)be a forest,a collection of disjoint trees,containing edges E x between pairs of variables in V x and edges E y connecting variables from V x and V y,E F=E x∪E y.If the probability distribution T(x|y)has G F for a Markov network,then similar to Equation1:

T(x|y)=

(u,v)∈E x

T(x u,x v)

T(x u)T(x v)

v∈V x

T(x v)

×

(u,v)∈E y

T(y u,x v)

T(y u)T(x v)

.

We will again use KL-divergence between conditional distributions P and T as an objective function:

KL(P,T)=

y P(y)

x

P(x|y)log

P(x|y)

T(x|y)

.

It can be shown that the optimal probability distribu-

Algorithm CondChowLiu(P)

Inputs:Distribution P over domain V x∪V y;pro-cedure MWST(V,weights)that outputs a maxi-mum weight spanning tree over V

1.(a)Compute marginal distributions

P(x u,x v)and P(x u)?u,v∈V x

(b)Compute marginal distributions P(y u)

and P(y u,x v)?u∈V y,v∈V x

2.(a)Compute mutual information values

I(x u,x v)?u,v∈V x

(b)Compute mutual information values

I(y u,x v)?u∈V y,v∈V x

(c)Find u(v)=arg max u∈V

y

I(y u,x v)?v∈

V x

(d)Let V =V x∪{v },and set I

x v ,x v

=

I

y u(v),x v

?v∈V x

3.(a)E T =MWST(V ,I)

(b)E x={(u,v)|u,v∈V x,(u,v)∈E T }

(c)E y={(u(v),v)|v∈V x,(v,v )∈E T }.

4.(a)Set T(x u,x v)=P(x u,x v)?(u,v)∈E x

(b)Set T(y u,x v)=P(y u,x v)?(u,v)∈E y Output:T

Figure2:Conditional Chow-Liu algorithm

tion T with corresponding Markov network G F is

?(u,v)∈E x,T(x u,x v)=P(x u,x v)

and

?(u,v)∈E y,T(y u,x v)=P(y u,x v).

As with the unconditional distribution,we wish to?nd pairs of variables to minimize

KL(P,T)=

v∈V x

H[x v]?H[x|y]

?

(u,v)∈E x

I(x u,x v)?

(u,v)∈E y

I(y u,x v)

where H[x v]denotes the entropy of P(x v),and H[x|y]denotes the conditional entropy of P(x|y). Both H[x]and H[x|y]are independent of E F,so as in the unconditional case,we need to solve a maximum spanning tree problem on the graph with nodes V y∪V x while not allowing paths between vertices in V y(alter-natively,assuming all nodes in V y are connected).

The algorithm for learning the conditional Chow-Liu (CCL)distribution is shown in Figure2.Due to the re-strictions on the edges,the CCL networks can contain

(a)

(b)

(c)

Figure 3:Conditional CL forest for a hypothetical dis-tribution with (a)1component (b)2components (c)3

components.

=

Figure 4:Graphical model for a hypothetical CCLF disconnected tree components (referred to as forests).These CCL forests can consist of 1to min {|V y |,|V x |}components.(See Figure 3for an illustration.)2.2.1

Chain CL Forests

We now return to our original goal of modeling time-dependent data.Let R t = R 1t ,...,R M

t be a multi-variate (M -variate)random vector of data with each component taking on values {0,...,B ?1}.By R 1:T we will denote observation sequences of length T .A simple model for such data can be constructed using conditional Chow-Liu forests.For this chain Chow-Liu forest model (CCLF),the data for a time point t is modeled as a conditional Chow-Liu forest given data at point t ?1(Figure 4):

P (R 1:T )=

T t =1

T (R t |R t ?1)

where T (R t =r |R t ?1=r )== (u,v )∈E V

T (R u t =r u ,R v t =r v )T (R v t =r v )T (R u t =r u ) v ∈R t T (R v

t =r v )× (u,v )∈E Ii T R v t =r v |R u t ?1

=r u v t .Note that learning the structure and parameters of CCLF requires one pass through the data to collect the counts and calculate joint probabilities of the pairs of variables,and only one run of the CondChowLiu tree algorithm.2.3

Hidden Markov Models

An alternative approach to modeling R 1:T is to use a hidden-state model to capture temporal dependence.Let S t be the hidden state for observation t ,taking on one of K values from 1to K ,where S 1:T denotes sequences of length T of hidden states.

A ?rst-order HMM makes two conditional indepen-dence assumptions.The ?rst assumption is that the hidden state process,S 1:T ,is ?rst-order Markov:

P (S t |S 1:t ?1)=P (S t |S t ?1)

(3)

and that this ?rst-order Markov process is homoge-neous in time,i.e.,the K ×K transition probability matrix for Equation 3does not change with time.The second assumption is that each vector R t at time t is independent of all other observed and unobserved states up to time t ,conditional on the hidden state S t at time t ,i.e.,

P (R t |S 1:t ,R 1:t ?1)=P (R t |S t ).

(4)

Specifying a full joint distribution P (R t |S t )would re-quire O (B M )joint probabilities per state,which is

clearly impractical even for moderate values of M .In practice,to avoid this problem,simpler models are often used,such as assuming that each vector com-ponent R j

t is conditionally independent (CI)of the other components,given the state S t ,i.e.,P (R t |S t )=

P (R 1t ,...,R M

t |S t )= M j =1P (R j t |S t ).We will use this HMM-CI as our baseline model in the experimental re-sults section later in the paper—in what follows below we explore models that can capture more dependence structure by using CL-trees.2.4

Chow-Liu Structures and HMMs

We can use HMMs with Chow-Liu trees or conditional Chow-Liu forests to model the output variable given

the hidden state.HMMs can model temporal struc-ture of the data while the Chow-Liu models can cap-ture “instantaneous”dependencies between multivari-ate outputs as well as additional dependence between vector components at consecutive observations over time that the state variable does not capture.By combining HMMs with the Chow-Liu tree model and with the conditional Chow-Liu forest model we

obtain HMM-CL and HMM-CCL models,respectively.

The set of parameters Θfor these models with K hid-den states and B -valued M -variate vector sets consists

=

T (R )T (R )S S =2S =3t

t

t

23t

t

R 1

t

R 2

t R 4t

R 3t R 1

t R 2

t R 3t

R 4t R 4t R 1

t R 2

t R 3t

S =1T (R )t

1

t

=

T (R |R )T (R |R )T (R |R )Figure 5:Graphical model interpretation of a hypothetical HMM-CL (left)and HMM-CCL (right)

of a K ×K transition matrix Γ,a K ×1vector Πof

probabilities for the ?rst hidden state in a sequence,and Chow-Liu trees or conditional forests for each hid-den state T ={T 1,...,T K }.Examples of graphical model structures for both the HMM-CL and HMM-CCL are shown in Figure 5.The likelihood of Θcan then be computed as

L (Θ)=P (R 1:T |Θ)=

S 1:T

P (S 1:T ,R 1:T |Θ)

=

S 1:T

P (S 1|Θ)

T t =2

P (S t |S t ?1,Θ)

×T t =1

P (R t |S t ,R t ?1,Θ)

=

K i 1=1

πi 1T i 1(R 1)T t =2K i t =1

γi t ?1i t T i t (R t |R t ?1)

with P (R t |S t ,R t ?1,Θ)=P (R t |S t ,Θ)

and

T i (R t |R t ?1)=T i (R t )for the HMM-CL.

For the value for the hidden state S t ?1=i ,the proba-bility distribution P (R t |Θ)is just a mixture of Chow-Liu trees (Meil?a and Jordan 2000)with mixture coef-?cients (γi 1,...,γiK )equal to the i -th row of the tran-sition matrix Γ.

As a side note,since the output part of the HMM-CCL contains dependencies on observations at the previous time-step,the model can be viewed as a form of au-toregressive HMM (Rabiner 1989).

3

Inference and Learning of HMM-based Models

In this section we discuss both (a)learning the struc-ture and the parameters of the HMM-CL and HMM-

CCL models discussed above,and (b)inferring proba-bility distributions of the hidden states for given a set

of observations and a model structure and its param-eters.We outline how these operations can be per-formed for the HMM-CL and HMM-CCL.For full de-tails,see Kirshner et al.(2004).3.1

Inference of the Hidden State Distribution

The probability of the hidden variables S 1:T given

complete observations R 1:T can be computed as

P (S 1:T |R 1:T )=P (S 1:T ,R 1:T )

S 1:T

P (S 1:T ,R 1:T ).

The likelihood (denominator)cannot be calculated directly since the sum is exponential in T .How-ever,the well-known recursive Forward-Backward pro-cedure can be used to collect the necessary informa-tion in O T K 2

M without exponential complexity (e.g.,Rabiner 1989).Since the Forward-Backward al-gorithm for HMM-CLs and HMM-CCLs is not very di?erent from standard HMMs,we will omit the de-tails.3.2

Learning

Learning in HMMs is typically performed using the Baum-Welch algorithm (Baum et al.1970),a vari-ant of the Expectation-Maximization (EM)algorithm (Dempster et al.1977).Each iteration of EM consists of two steps.First (E-step),the estimation of the pos-terior distribution of latent variables is accomplished by the Forward-Backward routine.Second (M-step),the parameters of the models are updated to maximize the expected log-likelihood of the model given the dis-tribution from the M-step.The structures of the trees

are also updated in the M-step.

The parameters Πand Γare calculated in the same manner as for regular HMMs.Updates for T 1,...,T K are computed in a manner similar to that for mixtures of trees (Meil?a and Jordan 2000).Suppose R 1:T =r 1:T .Let T i denote the Chow-Liu tree for S t =i under the updated model.It can be shown (Kirshner et al.2004)that to improve the log-likelihood one needs to maximize

K i =1

T τ=1

P (S τ=i |R 1:T =r 1:T ) T

t =1

P i (r t )log T i (r t )

where P i (r t )=P (S t =i |R 1:T =r 1:T )

T τ=1

P (S τ=i |R 1:T =r 1:T ).This can be accomplished by separately learning Chow-Liu struc-tures for the distributions P i ,the normalized posterior distributions of the hidden states calculated in the E-step.The time complexity for each iteration is then O T K 2M for the E-step and O T K 2

+KT M 2B 2 for the M-step.

4Experimental Results

To demonstrate the application of the HMM-CL and HMM-CCL models,we consider the problem of mod-eling precipitation occurrences for a network of rain stations.The data we examine here consists of bi-nary measurements (indicating precipitation or not)recorded each day over a number of years for each of a set of rain stations in a local region.Figure 6shows a network of such stations in Southwestern Australia.The goal is to build models that broadly speaking cap-ture both the temporal and spatial properties of the precipitation data.These models can then be used to simulate realistic rainfall patterns over seasons (e.g.,90-day sequences),as a basis for making seasonal fore-casts (Robertson et al.to appear),and to ?ll in missing rain station reports in the historical record.

Markov chains provide a well-known benchmark for modeling precipitation time-series at individual sta-tions (e.g.,Wilks and Wilby 1999).However,it is non-trivial to couple multiple chains together so that they exhibit realistic spatial correlation in simulated rain-fall patterns.We also compare against the simplest HMM with a conditional independence (CI)assump-tion for the rain stations given the state.This model captures the marginal dependence of the stations to a certain degree since (for example)in a “wet state”the probability for all stations to be wet is higher,and so forth.However,the CI assumption clearly does not fully capture the spatial dependence,motivating the use of models such as HMM-CL and HMM-CCL.In the experiments below we use data from both

Figure 6:Stations in the Southwestern Australia re-gion.Circle radii indicate marginal probabilities of rainfall (>0.3mm )at each location.

Southwestern Australia (30stations,15184-day win-ter seasons beginning May 1)and the Western United States (8stations,3990-day seasons beginning De-cember 1).In ?tting HMMs to this type of precip-itation data the resulting “weather states”are often of direct scienti?c interest from a meteorological view-point.Thus,in evaluating these models,models that can explain the data with fewer states are generally preferable.

We use leave-one-out cross-validation to evaluate the ?t of the models to the data.For evaluation we use two di?erent criteria.We compute the log-likelihood for seasons not in the training data,normalized by the number of binary events in the left-out sets (referred to here as out-of-sample scaled log-likelihood).We also compute the average classi?cation error in predicting observed randomly-selected station readings that are deliberately removed from the training data and then predicted by the model.The models considered are the independent Markov chains model (or “weather gen-erator”model),the chain Chow-Liu forest model,the HMM with conditional independence (HMM-CI),the HMM with Chow-Liu tree emissions (HMM-CL),and the HMM with conditional Chow-Liu tree emissions (HMM-CCL).For HMMs,K is chosen corresponding to the largest out-of-sample scaled log-likelihood for each model—the smallest such K is then used across di?erent HMM types for comparison.

The scatter plots in Figures 7and 8show the out-of-sample scaled log-likelihoods and classi?cation er-rors for the models on the left-out sets.The y-axis is the performance of the HMM-CCL model,and the x-axis represents the performance of the other models (shown with di?erent symbols).Higher implies better

Scaled ll for various models

S c a l e d l l f o r H M M ?C C L K =5

Prediction error for various models

P r e d i c t i o n e r r o r f o r H M M ?C C L K =5

Figure 7:Southwestern Australia data:scatterplots of out-of-sample scaled log-likelihoods (left)and average pre-diction error (right)obtained by leave-one-winter-out cross-validation.The line corresponds to y =x .The inde-pendent chains model is not shown since it is beyond the range of the plot (average ll =?0.6034,average error =0.291).

Scaled ll for various models S c a l e d l l f o r H M M ?C C L K =7

Prediction error for various models

P r e d i c t i o n e r r o r f o r H M M ?C C L K =7

Figure 8:Western U.S.data:Scatterplots of out-of-sample scaled log-likelihoods (left)and average prediction error (right)obtained by leave-one-winter-out cross-validation.The line corresponds to y =x .The independent chains model is not shown since it is beyond the range of the plot (average ll =?0.5204,average error =0.221).

Figure 9:Graphical interpretation of the hidden states for a 5-state HMM-CL trained on Southwestern Australia data (4out of 5shown).Radii of the circles indicate the precipitation probability for each station given the state.Lines between the stations indicate the edges in the graph while di?erent types of lines indicate the strength of mutual information of the edges.

performance for log-likelihood(on the left)and worse for error(on the right).The HMM-CL and HMM-CCL models are systematically better than the CCLF and HMM-CI models,for both score functions,and for both data sets.The HMM-CCL model does relatively better than the HMM-CL model on the U.S.data. This is explained by the fact that the Australian sta-tions are much closer spatially than the U.S.stations, so that for the U.S.the temporal connections that the HMM-CCL adds are more useful than the spatial con-nections that the HMM-CL model is limited to. Examples of the Chow-Liu tree structures learned by the model are shown in Figure9for the5-state HMM-CL model trained on all15years of Southwestern Aus-tralia data.The states learned by the model corre-spond to a variety of wet and dry spatial patterns. The tree structures are consistent with the meteorol-ogy and topography of the region(Hughes et al.1999). Winter rainfall over SW Australia is large-scale and frontal,impacting the southwest corner of the domain ?rst and foremost.Hence,the tendency for correla-tions between stations along the coast during moder-ately wet weather states.Interesting correlation struc-tures are also identi?ed in the north of the domain even during dry conditions.

5Conclusions

We have investigated a number of approaches for mod-eling multivariate discrete-valued time series.In par-ticular we illustrated how Chow-Liu trees could be em-bedded within hidden Markov models to provide im-proved temporal and multivariate dependence model-ing in a tractable and parsimonious manner.We also introduced the conditional Chow-Liu forest model,a natural extension of Chow-Liu trees for modeling con-ditional distributions such as multivariate data with temporal dependencies.Experimental results on real-world precipitation data indicate that these models provide systematic improvements over simpler alter-natives such as assuming conditional independence of the multivariate outputs.There are a number of ex-tensions that were not discussed in this paper but that can clearly be pursued,including(a)using in-formative priors over tree-structures(e.g.,priors on edges based on distance and topography for precipi-tation station models),(b)models for real-valued or mixed data(e.g.,modeling precipitation amounts as well as occurrences),(c)adding input variables to the HMMs(e.g.,to model“forcing”e?ects from atmo-spheric measurements—for initial results see Robert-son et al.(to appear)),and(d)performing systematic experiments comparing these models to more general classes of dynamic Bayesian networks where temporal and multivariate structure is learned directly.Acknowledgements

We would like to thank Stephen Charles of CSIRO, Australia,for providing us with the Western Australia data.This work was supported in part by the De-partment of Energy under grant DE-FG02-02ER63413 and in part by the National Science Foundation un-der Grant No.SCI-0225642as part of the OptIPuter project.

References

F.R.Bach and M.I.Jordan.Thin junction trees.In

T.G.Dietterich,S.Becker,and Z.Ghahramani,editors, Advances in Neural Information Processing Systems14, pages569–576,Cambridge,MA,2002.MIT Press.

L.E.Baum,T.Petrie,G.Soules,and N.Weiss.A max-imization technique occurring in statistical analysis of probabilistic functions of Markov chains.Annals of Mathematical Statistics,41(1):164–171,February1970.

C.K.Chow and C.N.Liu.Approximating discrete proba-

bility distributions with dependence trees.IEEE Trans-actions on Information Theory,IT-14(3):462–467,May 1968.

T.H.Cormen,C.E.Leiserson,and R.L.Rivest.Intro-duction to Algorithms.The MIT Electrical Engineering and Computer Science Series.MIT Press/McGraw Hill, 1990.

A.P.Dempster,https://www.wendangku.net/doc/8316569407.html,ird,and D.

B.Rubin.Max-

imum likelihood from incomplete data via EM algo-rithm.Journal of the Royal Statistical Society Series B-Methodological,39(1):1–38,1977.

J.P.Hughes,P.Guttorp,and S.P.Charles.A non-homogeneous hidden Markov model for precipitation oc-currence.Journal of the Royal Statistical Society Series

C Applied Statistics,48(1):15–30,1999.

S.Kirshner,P.Smyth,and A.W.Robertson.Conditional Chow-Liu tree structures for modeling discrete-valued vector time series.Technical Report04-04,School of Information and Computer Science,University of Cali-fornia,Irvine,2004.

M.Meil?a.An accelerated Chow and Liu algorithm:Fitting tree distributions to high-dimensional sparse data.In

I.Bratko and S.Dzeroski,editors,Proceedings of the

Sixteenth International Conference on Machine Learning (ICML’99),pages249–57.Morgan Kaufmann,1999. M.Meil?a and M.I.Jordan.Learning with mixtures of trees.Journal of Machine Learning Research,1(1):1–48, October2000.

L.R.Rabiner.A tutorial on hidden Markov models and selected applications in speech recognition.Proceedings of IEEE,77(2):257–286,February1989.

A.W.Robertson,S.Kirshner,and P.Smyth.Downscaling

of daily rainfall occurrence over Northeast Brazil using

a hidden Markov model.Journal of Climate,to appear.

D.S.Wilks and R.L.Wilby.The weather generation game:

a review of stochastic weather models.Progress in Phys-

ical Geography,23(3):329–357,1999.

Lemon Tree中英文歌词

Lemon Tree I'm sitting here in a boring room 我坐在这个无聊的房间里 It's just another rainy Sunday afternoon 周日的午后又下着雨 I'm wasting my time I got nothing to do 除了消磨时间我没什么事情可做 I'm hanging around I'm waiting for you 我徘徊我在等待你的到来 But nothing ever happens and I wonder 但什么也没有发生我很诧异 I'm driving around in my car 我开着车四处闲逛 I'm driving too fast I'm driving too far 我开的太快也开得太远 I'd like to change my point of view 我想换一个角度看待一切 I feel so lonely I'm waiting for you 我觉得好寂寞我在等着你 But nothing ever happens and I wonder 但什么也没有发生我很诧异 I wonder how I wonder why 我想知道怎么了我想知道为什么Yesterday you told me about the blue blue sky 昨天你还跟我说天空好蓝好蓝 And all that I can see is just a yellow lemon tree 但我却只看到一颗黄色的柠檬树 I'm turning my head up and down 我晃着脑袋上下观望 I'm turning turning turning turning turning around 我转来转去转来转去转来转去And all that I can see is just another yellow lemon-tree 看到的只是另一颗黄色的柠檬树 I'm sitting here I miss the power 我坐在这里我浑身乏力 I'd like to go out taking a shower 我想出去淋淋雨 But there's a heavy cloud inside my head 但脑海里萦绕着一片乌云 I feel so tired put myself into bed 我觉得好累只想躺在床上 Well, nothing ever happens and I wonder 什么事也没发生我很诧异 Isolation is not good for me 孤独并不适合我 Isolation, I don't want to sit on a lemon tree 我不想孤独地坐在柠檬树上 I'm stepping around in the desert of joy 我漫步在快乐的沙洲上 Baby anyhow I'll get another toy 宝贝无论如何我会找到另一个乐趣And everything will happen and you wonder 一切皆有可能到时候诧异的是你 I wonder how I wonder why 我想知道怎么了我想知道为什么Yesterday you told me about the blue blue sky 昨天你还跟我说天空好蓝好蓝 And all that I can see is just another lemon-tree 但我却只看到一颗黄色的柠檬树 I'm turning my head up and down 我晃着脑袋上下观望 I'm turning turning turning turning turning around 我转来转去转来转去转来转去 And all that I can see is just a yellow lemon tree 看到的只是另一颗黄色的柠檬树

千与千寻Spirited Away(英文剧本)

Spirited Away C: I?ll miss you, Chihiro. Your best friends, Rile F: Chihiro, Chihiro, we?re almost there. M: This really is the middle of nowhere. I?ve got to go to the next town to shop. F: We?ll just have to learn to like it. Look, Chihiro, there?s your new school. Looks really good, doesn?t it? M: It doesn?t look so bad. C: It?s gonna stink. I liked my old school. Oh````Mom, my flowers are dying. M: I told you not to smother them like that. We?ll put them in water when we get to our new home. C: I finally get a bouquet, and it?s a good-bye present. That?s depressing. M: Daddy bought you a rose for your birthday. Don?t you remember? C: Yes, one. Just one rose isn?t a bouquet. M: Hold on to your card. I?m opening the window and quit whining. It?s fun to move to a new place. It?s an adventure. F: Wait, did I take the wrong turn? That can?t be right. M: Look, there?s our house. It?s that blue one on the end. F: Oh, you?re right. I must have missed the turn-off. This road should get us there. M: Honey, don?t take a short cut. You always get us lost. F: Trust me, it?s gonna work. C: What are these stones? They look like little houses. M: They?re shines. Some people think little spirits live there. C: Dad, I think we?re lost. F: We?re fine. I?ll get a full wheel drive. M: Sit down please, sweetie. Slow down. You?re gonna kill us. F: What?s that? M: What?s this old building? F: It looks like an entrance. M: Honey, get back in the car. We are going to be late. Oh for heaven?s sake. F: This building is not old. It?s fake. These stones are just made of plaster. C: The wind?s pulling us in. M: What is it? F: Come one. Let?s go in. I want to see what?s on the other side. C: I?m not going. It gives me the creeps. F: Don?t be such a scaredy cat, Hiro. Let?s just take a look. M: The movers will get to our house before we do. F: It?s alright. They?ve got the keys. They can start without us. M: Alright. Just a quick look. C: Forget it. I?m not going. Come on you guys, le t?s get out of here. F: Come on honey. It will be fun. C: I?m not going. M: Chihiro, just wait in the car then. C: But mom….Wait for me. F: Everybody watch your step. M: Chihiro, don?t cling like that. You make me trip. -What is this place? -Oh, do you hear that? - It sounds like a train. - We must be near a train station. - Come on. Let?s go and check it out. - What are those weird buildings? - I knew it. It?s an abandoned theme park. See? They built the everywhere in early 90s, then the economy went bad and they all went bankrupt. This must be one of them. - Where are you going? You said just a quick look. Now let?s go back. Did you hear that building? It was moaning. - It?s just the wind. Oh, what a beautiful place. We should have brought our lunch. Then we could have a picnic. - Look, they were planning to put a river here. En~~~ you smell that? Something smells delicious. - Yeah, I?m starving. - Maybe this theme park is still in business. Let?s go. - Chihiro, hurry it up. - Wait a minute! - Over there….This way. - How strange! They?re all restaurants. - Where?s everybody? - Aaa… There it is! Hey, I found it. Hi, you got to see this. In here. - Aaa… Look at this! - Hello in there. Does anybody work here? - Come in Hiro. It looks delicious. - Anybody? - Don?t worry honey. We can pay the bill when they get back. - Good plan. Hey that looks great. - I wonder what this is called. Oh… It?s delicious. Chihiro, you have to taste this. - I don?t want any. We?re gonna get in trouble. Let?s just get out of here. - Don?t worry. You?ve got daddy here. He?s got credit cards and cash. - Chihiro, you have to try this. It?s so tender. - Mustard? - Thank you. - Come on, you guys. You can?t - That?s weird. It?s a bath house. There?s the train. - You shouldn?t be here. Get out of here, now! - What? - It?s almost night. Leave before it gets dark. They are lighting the lamps. Get out of here. You?ve gotta get across the river. Go! I?ll distract them. - What?s up with him? - Mom, dad, come on, quit eating, let?s get out of here. Ahh~~~~Mom, dad, where are you?.....Water?!! What?... I?m dreaming! I?m dreaming! Come on, wake up! Wake up! This is a dream. Just a dream. Away, away. Disappear! Aaa~~~~~? I?m see-through! It?s just a bad dream. - Don?t be afraid! I just wantta help you. - No…No! - Open your mouth and eat this. You must eat some food from this world or else you?ll disappear. - No. - Don?t worry. It won?t turn you into a pig. Chew it and swallow. There you go. You?re all better. See for yourself. - I?m OK. - You see. Now come with me. - Where are my mom and dad? They didn?t really turn into pigs, did they? - You can?t see them now, but you will. Don?t move. That bird?s looking for you. You?ve gotta get out of here. - My legs, I can?t stand up. Help! What do I do? - Calm down. Take a deep breath. In the name of the wind and water within thee, unbind her. Get up. - You have to hold your breath when we across the bridge. Even the tiniest breath will break the spell and then everyone will see you. - I?m scared. - Just stay calm. I?m back from my mission. - Welcome back Master Haku. - Take a deep breath. Hold it. Hang on, almost there. - Master Haku, where?ve you been? What? A human? - Let?s go. - They know you?re here. - I?m sorry. I took a breath.

Lemon Tree 歌词-精选学习文档

Lemon Tree一歌由Fool's Garden（傻子的花园）于2019年首唱，2019年，一曲“LEMONTREE”（柠檬树）使这支原本寂寂无名的德国5人乐队一下子红遍欧洲、亚洲，同时也有The brothers four演唱的美国乡村版，在被苏慧伦翻唱后也开始为国人熟知。台湾歌手王若琳在2011年9月23日发行的专辑《为爱做的一切》中翻唱此曲。 歌词 I'm sitting here in the boring room. 我呆坐在这了无生趣的房间。 It's just another rainy Sunday afternoon. 又是一个周日的午后，又是阴雨连绵。 I'm wasting my time, I got nothing to do. 我无事可做，我空耗着时间。 I'm hanging around, I'm waiting for you. 我不安地徘徊，我期待着你的出现。 But nothing ever happens, and I wonder. 但是你终究没有出现，我纳闷。 I'm driving around in my car. 我出去转转，驾着我的车。 I'm driving too fast, I'm driving too far. 我开得太快，我开得太远。 I'd like to change my point of view. 我宁愿转移一下我的注意力。

I feel so lonely, I'm waiting for you. 我仍感到如此孤单，我期待着你的出现。 But nothing ever happens, and I wonder. 但是你至今没有出现，我纳闷。 I wonder how, I wonder why. 我不知所措，我不明所以。 Yesterday you told me about the blue, blue sky. 昨天你给我描绘那蓝色的、蓝色的天空。 And all that I can see is just a yellow lemon tree. 但是我所看见的只有一株黄色的柠檬树。 I'm turning my head up and down. I'm turning, turning, turning, turning, turning around.我转着我的头，上上下下，我转啊、转啊、转啊、转啊、一遍又一遍 And all that I can see is just another lemon tree. 我所看见的只是另一株柠檬树。 Say,da ,da la da la, de la da,da la da la, de la da,da de le de da I'm sitting here, I miss the power. 我呆坐在这，我精疲力尽。 I'd like to go out taking a shower. 我宁愿出去沐浴一下。 But there's a heavy cloud inside my head. 但是这片阴云总在在我头脑中挥之不去。

Lemon Tree中英文歌词教学文稿

L e m o n T r e e中英文 歌词

精品文档 收集于网络，如有侵权请联系管理员删除 Lemon Tree I'm sitting here in a boring room 我坐在这个无聊的房间里 It's just another rainy Sunday afternoon 周日的午后又下着雨 I'm wasting my time I got nothing to do 除了消磨时间 我没什么事情可做 I'm hanging around I'm waiting for you 我徘徊 我在等待你的到来 But nothing ever happens and I wonder 但什么也没有发生 我很诧异 I'm driving around in my car 我开着车四处闲逛 I'm driving too fast I'm driving too far 我开的太快 也开得太远 I'd like to change my point of view 我想换一个角度看待一切 I feel so lonely I'm waiting for you 我觉得好寂寞 我在等着你 But nothing ever happens and I wonder 但什么也没有发生 我很诧异 I wonder how I wonder why 我想知道怎么了 我想知道为什么 Yesterday you told me about the blue blue sky 昨天你还跟我说天空好蓝好蓝 And all that I can see is just a yellow lemon tree 但我却只看到一颗黄色的柠檬树 I'm turning my head up and down 我晃着脑袋 上下观望 I'm turning turning turning turning turning around 我转来转去 转来转去 转来转去 And all that I can see is just another yellow lemon-tree 看到的只是另一颗黄色的柠檬树 I'm sitting here I miss the power 我坐在这里 我浑身乏力 I'd like to go out taking a shower 我想出去淋淋雨 But there's a heavy cloud inside my head 但脑海里萦绕着一片乌云 I feel so tired put myself into bed 我觉得好累 只想躺在床上 Well, nothing ever happens and I wonder 什么事也没发生 我很诧异 Isolation is not good for me 孤独 并不适合我 Isolation, I don't want to sit on a lemon tree 我不想孤独地坐在柠檬树上 I'm stepping around in the desert of joy 我漫步在快乐的沙洲上 Baby anyhow I'll get another toy 宝贝无论如何我会找到另一个乐趣 And everything will happen and you wonder 一切皆有可能 到时候诧异的是你 I wonder how I wonder why 我想知道怎么了 我想知道为什么 Yesterday you told me about the blue blue sky 昨天你还跟我说天空好蓝好蓝 And all that I can see is just another lemon-tree 但我却只看到一颗黄色的柠檬树 I'm turning my head up and down 我晃着脑袋 上下观望 I'm turning turning turning turning turning around 我转来转去 转来转去 转来转去 And all that I can see is just a yellow lemon tree 看到的只是另一颗黄色的柠檬树

lemon_tree的中文对照歌词

lemon tree的中文对照歌词 I'm sitting here in a boring room.我坐在这间空屋子里 It's just another rainy Sunday afternoon。这也只不过是另一个下雨的周日下午。I’m wasting my time, I got nothing to do.除了消磨时间我没什模事情可做。 I’m hanging around, I’m waiting for you。我四处张望，我在等待你的到来。 But nothing ever happens, and i wonder.但是好像什莫事情也未曾发生，我不知道。 I’m driving around in my car.我开着车出去兜风。 I’m driving too fast, I’m driving too far。我把车开得很快，开了很远。 I’d like to change my point of view。我想换种方式生活，换个角度看世界。 I feel so lonely, I’m waiting for you.我感觉到如此孤单，我在等你回来。 but nothing ever happens, and i wonder.但是什莫事也未曾发生，我不明白为什莫会这样。 I wonder how, i wonder why。我不知道怎莫办，我不知道为什莫会这样。yesterday you told me about the blue, blue sky.昨天你还给我讲那蓝蓝的天空会多莫美丽，生活会多莫美好。 and all that i can see is just a yellow lemon tree。但是我看见的只有一株柠檬树。I’m turning my head up and down。我摆动着我的头，上下上下，不停的摆动着我的头。 I’m turning, turning, turning, turning, turning around.我前后左右的看。（我左看右看上看下看） And all that i can see is just another yellow lemon tree.（不管我怎莫看）它也只是一株黄色的柠檬树。 Sing: da, dadadada didada,dadada didada, dadidada. I’m sitting here, and i miss the power.我坐在这里，身上没有了一丝力气。 I’d like to go out taking a shower。我想出去展示一下我自己。 but there's a heavy cloud inside my head.但有一片阴云在我脑中挥之不去。 I feel so tired, and put myself into bed.我感觉如此的疲惫，回到了家里把自己扔到了床上。 while nothing ever happens, and I wonder。也许什莫也没发生过，谁知道呢？Isolation is not good for me 孤独对我不好 Isolation I don't want to sit on the lemon tree 我不想孤独地守着一颗柠檬树。I’m stepping around in a desert of joy.我在快乐的沙漠中踱步。 baby, anyhow I get another toy。宝贝，无论如何我要找到另外一种快乐。 and everything will happen, and you wonder。然后，所有事情都会发生，然而这一切你不会知道。 I’m turning my head up and down。我摆动着我的头，上下上下，不停的摆动着我的头。 I’m turning, turning, turning, turning, turning around.我前后左右的看。（我左看右看上看下看） And all that I can see is just another yellow lemon tree.（不管我怎莫看）它也只是一株黄色的柠檬树。 I wonder how, i wonder why。我不知道怎莫办，我不知道为什莫会这样。

always_with_me《千与千寻》_中文歌词_罗马音及日文歌词

《千与千寻》always with me 中文歌词罗马音及日文歌词呼唤心灵深处的某个地方 总想保持令人心动的梦想 悲伤虽然无法尽数 在它对面一定能与你相逢 每次重蹈覆辙时人总是 仅仅知道碧空之蓝 虽然永无止境的道路看起来总在延续 这双手一定可以拥抱光明 别离时 平静的胸怀 虽然从零开始仍要侧耳倾听 活着的不可思议死去的不可思议 花，风，街道都一样 啦啦啦…… 啦啦啦…… 啦啦啦…… 呼唤心灵深处的某个地方 不论何时与我同在去描绘梦想吧 与其道尽悲伤的数目 不如用相同的双唇轻轻歌唱 走向尘封的回忆中总是 听得到不愿忘记的细语 即使是在 被粉碎的镜子上 也会映出崭新的美景 开始的清晨那宁静的窗口 因为将从零开始渐渐被充实 不再追寻大海的彼端 因为那闪光的东西一直就在这里 在我心中被发现 啦啦啦…… 啦啦啦……

1.yo n de i lu / mu ne no do ko ka o ku de 呼んでいる胸のどこか奥で 2、i tsu mo ko ko lo o do lu / yu me wo mi ta i いつも心踊る梦を见たい 3、ka na shi mi wa / ka zo e ki le na i ke le do 悲しみは数えきれないけれど 4、so no mu ko u de ki to / a na ta ni a e lu その向こうできっとあなたに会える 5、ku li ka e su a ya ma chi no / so no ta bi hi to wa 缲り返すあやまちのそのたびひとは 6、ta da a o i so la no / a o i sa wo shi lu ただ青い空の青さを知る 7、ha te shi na ku / mi chi wa tu zu i te mi e lu ke le do 果てしなく道は続いて见えるけれど 8、ko no li yo u te wa / hi ka li wo i da ke lu この両手は光を抱ける 9、sa yo na la no to ki no /shi zu ka na mu ne さよならのときの静かな胸 10、ze lo ni na lu ka la da ga / mi mi wo su ma se lu ゼロになるからだが耳をすませる 11、i ki te i lu fu shi gi / shi n de i ku fu shi gi 生きている不思议死んでいく不思议 ha na mo ka ze mo ma chi mo minna o na ji 12、ららら……（la la la ……） おおお……（o o o ……） るるる……（lu lu lu ……） 13、yo n de i lu / mu ne no do ko ka o ku de 呼んでいる胸のどこか奥で 14、i tsu mo na n do de mo / yu me wo e ga ko u いつも何度でも梦を描こう

lemontree英文谐音歌词

l e m o n t r e e英文谐音歌 词 The Standardization Office was revised on the afternoon of December 13, 2020

1. I'm sitting here in a boring room. 暗木 C厅嘿一耳音额包ring 容 2. It's just another rainy sunday afternoon. 诶次炸死t 额那the ring泥桑得(dei) 阿夫特农 3. I'm wasting my time, I’ve got nothing to do. 暗木喂死听买叹木哎夫告特那thing 土度 4. I'm hanging around, I'm waiting for you. 暗木憨银额让的暗木喂厅佛有 5. But nothing ever happens, and I wonder. 巴特那thing 唉我嗨喷死安得艾我望的 6. I'm driving around in my car. 暗木抓ving 额让的银卖卡 7. I'm driving too fast, I'm drving too far. 暗木抓ving 吐发死特暗木抓ving 吐发 8. I'd like to change my point of view 艾的来克吐 ching只买抛银特哦夫 V有 9. I feel so lonely, I'm waiting for you. 哎肥呦嗖咙里暗木喂厅佛有 10. But nothing ever happens, and I wonder. 巴特那thing 哎我嗨碰死暗的艾我望的 11. I wonder how, I wonder why. 哎望的好艾我望的外 12. Yesterday you told me about the blue, blue sky. 耶死特得有偷得米额抱特 the 不录不录死盖 13. And all that I can see is just a yellow lemon tree. 暗的奥 that 哎看C 诶Z 炸死特额也楼来蒙吹 14. I'm turning my head up and 'm turning, turning, turning, turning, turning around. 暗木疼宁买害的啊扑暗的荡暗木疼宁疼宁疼宁疼宁疼宁额让的 15. And all that I can see is just another lemon tree. 暗的噢 that 哎看C 诶Z 炸死特额那the 也楼来蒙吹 16. Say,da ,da da da da, de da da, da da da da, de da da,da de de da C一哒哒哒哒哒地哒哒哒哒哒哒地哒哒哒地地哒 17. I'm sitting here,I miss the power. 暗木C厅嘿一耳哎迷死 the 怕我 18. I'd like to go out taking a shower. 哎的来克吐狗奥特忒king额杀我 19. But there's a heavy cloud inside my head. 巴特 Z而死呃嗨V克劳德因赛的买嗨的 20. I feel so tired,put myself into bed. 哎飞儿搜太儿的普特买塞尔福银吐百的

千与千寻的片尾曲歌词

哪位大虾能帮我解释一下千与千寻片尾曲always with me 的歌词的意思？ 歌词： 内心深处的呼唤 我想要走进悸动的梦中 虽然悲伤总是会重演 但我一定能在某处与你相逢 人们总是不停犯错 他们只知道天是蓝的 虽然前路渺茫 但他们的双手仍在寻找光明 离别时平静的心 身体归于虚无时的倾听 莫名的生存，莫名的死去 花，风，城市都是如此 啦啦啦~~~ 内心深处的呼唤 让我们不停的画出梦的色彩 比起回忆心中的悲伤 不如用同样的唇轻声歌唱 即使在封锁的回忆中 仍然还有无法忘记的呢喃 即使在粉碎的镜片中 仍然能映出新的景色 晨色初照下的宁静窗台 还有化为虚无的身体 从此我不会越过大洋去寻找 闪耀的所有都在身边 我将自己去追寻 啦啦啦~~~ 我来帮他解答 满意回答 2009-03-29 12:40 这是我一点不成熟的看法，理解的不到位，见谅

宫崎骏大师总是善于刻画孩子内心对世界的看法和对待事物的心理，大家都看过这部电影吧，刚开始的时候，千寻是一个胆小麻木的孩子，这与家长的教育很有关，家长首先就对千寻不是很重视，从简短的几句对白和在通往神明的隧道口就能够看出来，不过虽然千寻对世事比较麻木，但是毕竟她还是生活在象牙塔里的小学生而已，大人们对她自身的想法不重视导致她对自己的不自信和对除自己之外的人的不相信。到这就不难理解这段歌词了。 首先第一句是说有关她的这段非比寻常的经历，对这段故事的不舍，那个如梦一样的世界跟现实世界一样，值得她去怀念。在即将和那个世界作别时，他和小白的对话中，充分显示了他们互相的不舍。其实这个世界，离别才是永恒的，但他们仍然相信他们可以重逢。 第二段，宫崎骏借千寻的角度道破，其实大人们总是太过自私，总是不顾一切的去追逐自己想要的，却忽略了最重要的。 第三段，千寻刚进入那个世界是胆小的，眼睛里充满茫然，而且不懂礼貌，对救了她的锅炉爷爷都没说一句谢谢，去见汤婆婆连门也没敲。可是后来千寻却能把河神给自己的丸子给白龙和无面人吃，足以看出千寻的变化。在回到人类世界之后，千寻回头不舍得望向给自己很多宝贵东西的时候，眼神是那么坚定。这时的千寻实际上已经从神明世界的小千那里继承了某种精神，她已经不再像以前那样迷茫，相信她在这个故事之后，会有她的另一个人生。 之后的几段我感觉和前面的差不多，就是

lemon tree 英文谐音歌词资料讲解

l e m o n t r e e英文谐 音歌词

1. I'm sitting here in a boring room. 暗木 C厅嘿一耳音额包ring 容 2. It's just another rainy sunday afternoon. 诶次炸死t 额那the ring泥桑得(dei) 阿夫特农 3. I'm wasting my time, I’ve got nothing to do. 暗木喂死听买叹木哎夫告特那thing 土度 4. I'm hanging around, I'm waiting for you. 暗木憨银额让的暗木喂厅佛有 5. But nothing ever happens, and I wonder. 巴特那thing 唉我嗨喷死安得艾我望的 6. I'm driving around in my car. 暗木抓ving 额让的银卖卡 7. I'm driving too fast, I'm drving too far. 暗木抓ving 吐发死特暗木抓ving 吐发 8. I'd like to change my point of view 艾的来克吐 ching只买抛银特哦夫 V有 9. I feel so lonely, I'm waiting for you. 哎肥呦嗖咙里暗木喂厅佛有 10. But nothing ever happens, and I wonder. 巴特那thing 哎我嗨碰死暗的艾我望的 11. I wonder how, I wonder why. 哎望的好艾我望的外 12. Yesterday you told me about the blue, blue sky. 耶死特得有偷得米额抱特 the 不录不录死盖 13. And all that I can see is just a yellow lemon tree. 暗的奥 that 哎看C 诶Z 炸死特额也楼来蒙吹 14. I'm turning my head up and down.I'm turning, turning, turning, turning, turning around. 暗木疼宁买害的啊扑暗的荡暗木疼宁疼宁疼宁疼宁疼宁额让的 15. And all that I can see is just another lemon tree. 暗的噢 that 哎看C 诶Z 炸死特额那the 也楼来蒙吹 16. Say,da ,da da da da, de da da, da da da da, de da da,da de de da C一哒哒哒哒哒地哒哒哒哒哒哒地哒哒哒地地哒 17. I'm sitting here,I miss the power. 暗木C厅嘿一耳哎迷死 the 怕我 18. I'd like to go out taking a shower. 哎的来克吐狗奥特忒king额杀我 19. But there's a heavy cloud inside my head. 巴特 Z而死呃嗨V克劳德因赛的买嗨的 20. I feel so tired,put myself into bed. 哎飞儿搜太儿的普特买塞尔福银吐百的

千与千寻主题曲 歌词

1、yo n de i ru / mu ne no do ko ka o ku de 呼んでいる胸のどこか奥で 2、i tsu mo ko ko ro o do ru / yu me wo mi ta i いつも心踊る梦を见たい 3、ka na shi mi wa / ka zo e ki re na i ke re do 悲しみは数えきれないけれど 4、so no mu ko u de ki to / a na ta ni a e ru その向こうできっとあなたに会える 5、ku ri ka e su a ya ma chi no / so no ta bi hi to wa 缲り返すあやまちのそのたびひとは 6、ta da a o i so ra no / a o i sa wo shi ru ただ青い空の青さを知る 7、ha te shi na ku / mi chi wa tu zu i te mi e ru ke re do 果てしなく道は続いて见えるけれど 8、ko no ri yo u te wa / hi ka ri wo i da ke ru この両手は光を抱ける 9、sa yo na ra no to ki no /shi zu ka na mu ne さよならのときの静かな胸 10、ze ro ni na ru ka ra da ga / mi mi wo su ma se ru ゼロになるからだが耳をすませる 11、i ki te i ru fu shi gi / shi n de i ku fu shi gi 生きている不思议死んでいく不思议

alwayswithme《千与千寻》中文歌词罗马音及日文歌词

《千与千寻》a l w a y s w i t h m e中文歌词罗马音及日文歌词 呼唤心灵深处的某个地方?? 总想保持令人心动的梦想?? 悲伤虽然无法尽数?? 在它对面一定能与你相逢?? 每次重蹈覆辙时人总是?? 仅仅知道碧空之蓝?? 虽然永无止境的道路看起来总在延续?? 这双手一定可以拥抱光明?? 别离时?? 平静的胸怀?? 虽然从零开始仍要侧耳倾听?? 活着的不可思议死去的不可思议?? 花，风，街道都一样?? 啦啦啦……?? 啦啦啦……?? 啦啦啦……?? 呼唤心灵深处的某个地方?? 不论何时与我同在去描绘梦想吧?? 与其道尽悲伤的数目?? 不如用相同的双唇轻轻歌唱?? 走向尘封的回忆中总是?? 听得到不愿忘记的细语?? 即使是在?? 被粉碎的镜子上?? 也会映出崭新的美景?? 开始的清晨那宁静的窗口?? 因为将从零开始渐渐被充实?? 不再追寻大海的彼端?? 因为那闪光的东西一直就在这里?? 在我心中被发现?? 啦啦啦……?? 啦啦啦…… 1.yo n de i lu / mu ne no do ko ka o ku de 呼んでいる胸のどこか奥で 2、i tsu mo ko ko lo o do lu / yu me wo mi ta i いつも心踊る梦を见たい 3、ka na shi mi wa / ka zo e ki le na i ke le do 悲しみは数えきれないけれど

4、so no mu ko u de ki to / a na ta ni a e lu その向こうできっとあなたに会える 5、ku li ka e su a ya ma chi no / so no ta bi hi to wa 缲り返すあやまちのそのたびひとは 6、ta da a o i so la no / a o i sa wo shi lu ただ青い空の青さを知る 7、ha te shi na ku / mi chi wa tu zu i te mi e lu ke le do 果てしなく道は続いて见えるけれど 8、ko no li yo u te wa / hi ka li wo i da ke lu この両手は光を抱ける 9、sa yo na la no to ki no /shi zu ka na mu ne さよならのときの静かな胸 10、ze lo ni na lu ka la da ga / mi mi wo su ma se lu ゼロになるからだが耳をすませる 11、i ki te i lu fu shi gi / shi n de i ku fu shi gi 生きている不思议死んでいく不思议 ha na mo ka ze mo ma chi mo minna o na ji 12、ららら……（la la la ……） おおお……（o o o ……） るるる……（lu lu lu ……） 13、yo n de i lu / mu ne no do ko ka o ku de 呼んでいる胸のどこか奥で 14、i tsu mo na n do de mo / yu me wo e ga ko u いつも何度でも梦を描こう 15、ka na shi mi no ka zu wo / ii tsu ku su yo li 悲しみの数を言い尽くすより 16、o na ji ku chi bi lu de / so to u ta o u 同じくちびるでそっとうたおう 17、To ji te i ku o mo i de no / so no na ka ni i tsu mo 闭じていく思い出のそのなかにいつも