This paper provides evidence that the use of more unlabeled data in semi-supervised learning can improve the performance of Natural Language Processing (NLP) tasks, such as part-of-speech tagging, syntactic chunking, and named entity recognition.
We first propose a simple yet powerful semi-supervised discriminative model appropriate for handling large scale unlabeled data.
Then, we describe experiments performed on widely used test collections, namely, PTB III data, CoNLL'00 and '03 shared task data for the above three NLP tasks, respectively.
We incorporate up to lG-words (one billion tokens) of unlabeled data, which is the largest amount of unlabeled data ever used for these tasks, to investigate the performance improvement.
In addition, our results are superior to the best reported results for all of the above test collections.
1 Introduction
Today, we can easily find a large amount of unlabeled data for many supervised learning applications in Natural Language Processing (NLP).
Therefore, to improve performance, the development of an effective framework for semi-supervised learning (SSL) that uses both labeled and unlabeled data is attractive for both the machine learning and NLP communities.
We expect that such SSL will replace most supervised learning in real world applications.
In this paper, we focus on traditional and important NLP tasks, namely part-of-speech (POS) tagging, syntactic chunking, and named entity recognition (NER).
These are also typical supervised learning applications in NLP, and are referred to as sequential labeling and segmentation problems.
In some cases, these tasks have relatively large
amounts of labeled training data.
In this situation, supervised learning can provide competitive results, and it is difficult to improve them any further by using SSL.
In fact, few papers have succeeded in showing significantly better results than state-of-the-art supervised learning.
Ando and Zhang (2005) reported a substantial performance improvement compared with state-of-the-art supervised learning results for syntactic chunking with the CoNLL'00 shared task data (Tjong Kim Sang and Buchholz, 2000) and NER with the CoNLL'03 shared task data (Tjong Kim Sang and Meulder, 2003).
One remaining question is the behavior of SSL when using as much labeled and unlabeled data as possible.
This paper investigates this question, namely, the use of a large amount of unlabeled data in the presence of (fixed) large labeled data.
To achieve this, it is paramount to make the SSL method scalable with regard to the size ofunlabeled data.
We first propose a scalable model for SSL.
Then, we apply our model to widely used test collections, namely Penn Treebank (PTB) III data (Marcus et al., 1994) for POS tagging, CoNLL'00 shared task data for syntactic chunking, and CoNLL'03 shared task data for NER.
We used up to lG-words (one billion tokens) of unlabeled data to explore the performance improvement with respect to the unla-beled data size.
In addition, we investigate the performance improvement for 'unseen data' from the viewpoint of unlabeled data coverage.
Finally, we compare our results with those provided by the best current systems.
The contributions of this paper are threefold.
First, we present a simple, scalable, but powerful task-independent model for semi-supervised sequential labeling and segmentation.
Second, we report the best current results for the widely used test
collections described above.
Third, we confirm that the use of more unlabeled data in SSL can really lead to further improvements.
2 Conditional Model for SSL
We design our model for SSL as a natural semi-supervised extension of conventional supervised conditional random fields (CRFs) (Lafferty et al., 2001).
As our approach for incorporating unla-beled data, we basically follow the idea proposed in (Suzuki et al., 2007).
2.1 Conventional Supervised CRFs
where Z(x) = J2yeyllcec ^c(yc, x; X) is the partition function.
We generally assume that the potential function is a non-negative real value function.
Therefore, the exponentiated weighted sum over the features of a clique is widely used, so that, ^c(yc, x; X) =exp(X • f c(yc, x)) where f c(yc, x) is a feature vector obtained from the corresponding clique c in G(x, y).
2.2 Semi-supervised Extension for CRFs
a difference in that generative models are directed graphical models while our conditional PM is an undirected.
However, this difference causes no violations when we construct our approach.
= exp(X ■ hc(ycl x)).
where Z'(x)=Y, y^yll cec K&c, x; X', 0).
Hereafter in this paper, we refer to this conditional model as a 'Jointprobability model Embedding style Semi-Supervised Conditional Model', or JESS-CM for short.
where p(X') is a prior probability distribution of X'.
Clearly, JESS-CM shown in Equation 2 has exactly the same form as Equation 1.
With a fixed 0, the log-likelihood, log Pj, can be seen simply as the feature functions of JESS-CM as with f j. Therefore, embedded joint PMs do not violate the global convergence conditions.
As a result, as with supervised CRFs, it is guaranteed that X' has a value that achieves the global maximum of L1(X'\0).
Moreover, we can obtain the same form of gradient as that of supervised CRFs (Sha and Pereira, 2003), that is,
~/2Ep(y\xn;X,&) [h(Y, xn)} +Vlogp(X').
Thus, we can easily optimize Li by using the forward-backward algorithm since this paper solely
focuses on a sequence model and a gradient-based optimization algorithm in the same manner as those used in supervised CRF parameter estimation.
We cannot naturally incorporate unlabeled data into standard discriminative learning methods since the correct outputs y for unlabeled data are unknown.
On the other hand with a generative approach, a well-known way to achieve this incorporation is to use maximum marginal likelihood (MML) parameter estimation, i.e., (Nigam et al., 2000).
Given unlabeled data Du = [xm}M=i, MML estimation in our setting maximizes the marginal distribution of a joint PM over a missing (hidden) variable y, namely, it maximizes £ m logY, yey p(xm, y; 6).
Following this idea, there have been introduced a parameter estimation approach for non-generative approaches that can effectively incorporate unla-beled data (Suzuki et al., 2007).
Here, we refer to it as 'Maximum Discriminant Functions sum' (MDF) parameter estimation.
MDF estimation substitutes p(x, y) with discriminant functions g(x, y).
Therefore, to estimate the parameter 0 of JESS-CM by using MDF estimation, the following objective function is maximized with a fixed X':
where p(0) is a prior probability distribution of 0 .
Since the normalization factor does not affect the determination of y, the discriminant function of JESS-CM shown in Equation 2 is defined as g(x, y; X', 0) = ]Jcec V'c&c, x; X', 0).
With a fixed X', the local maximum of C2(0\Xr) around the initialized value of 0 can be estimated by an iterative computation such as the EM algorithm (Demp-steretal., 1977).
2.3 Scalability: Efficient Training Algorithm
A parameter estimation algorithm of X' and 0 can be obtained by maximizing the objective functions L1(X'\0) and L2(0\X') iteratively and alternately.
Figure 1 summarizes an algorithm for estimating X' and 0 for JESS-CM.
This paper considers a situation where there are many more unlabeled data M than labeled data N, that is, N << M. This means that the calculation cost for unlabeled data is dominant.
Thus, in order to make the overall parameter estimation procedure
maximize £1(A'\@) with fixed 0 — @(t-1) using Vi.
with fixed A' using Vu.
0(*)_ 0(*-i) do lmtil J—]0(t-i)l— < e.
Reestimate A': perform the same procedure as 1.
Figure 1: Parameter estimation algorithm for JESS-CM.
scalable for handling large scale unlabeled data, we only perform one step of MDF estimation for each t as explained on 3. in Figure 1.
In addition, the calculation cost for estimating parameters ofembedded joint PMs (HMMs) is independent of the number of
result, the cost for calculating the JESS-CM parameters, X' and 0, is essentially the same as executing T iterations of the MML estimation for a single HMM using the EM algorithm plus T + 1 time optimizations of the MAP estimation for a conventional supervised CRF if it converged when t = T. In addition, our parameter estimation algorithm can be easily performed in parallel computation.
2.4 Comparison with Hybrid Model
SSL based on a hybrid generative/discriminative approach proposed in (Suzuki et al., 2007) has been defined as a log-linear model that discriminatively combines several discriminative models, pD, and generative models, p^, such that:
With the hybrid model, ifwe use the same labeled training data to estimate both A and T, Yjs will become negligible (zero or nearly zero) since pD is already fitted to the labeled training data while pG are trained by using unlabeled data.
As a solution, a given amount of labeled training data is divided into two distinct sets, i.e., 4/5 for estimating A, and the
remaining 1/5 for estimating T (Suzuki et al., 2007).
Moreover, it is necessary to split features into several sets, and then train several corresponding discriminative models separately and preliminarily.
In contrast, JESS-CM is free from this kind of additional process, and the entire parameter estimation procedure can be performed in a single pass.
Surprisingly, although JESS-CM is a simpler version of the hybrid model in terms of model structure and parameter estimation procedure, JESS-CM provides F-scores of 94.45 and 88.03 for CoNLL'00 and '03 data, respectively, which are 0.15 and 0.83 points higher than those reported in (Suzuki et al., 2007) for the same configurations.
This performance improvement is basically derived from the full benefit of using labeled training data for estimating the parameter of the conditional model while the combination weights, T, of the hybrid model are estimated solely by using 1/5 of the labeled training data.
These facts indicate that JESS-CM has several advantageous characteristics compared with the hybrid model.
3 Experiments
In our experiments, we report POS tagging, syntactic chunking and NER performance incorporating up to lG-words of unlabeled data.
To compare the performance with that of previous studies, we selected widely used test collections.
For our POS tagging experiments, we used the Wall Street Journal in PTB III (Marcus et al., 1994) with the same data split as used in (Shen et al., 2007).
For our syntactic chunking and NER experiments, we used exactly the same training, development and test data as those provided for the shared tasks of CoNLL'00 (Tjong Kim Sang and Buchholz, 2000) and CoNLL'03 (Tjong Kim Sang and Meul-der, 2003), respectively.
The training, development and test data are detailed in Table 11 .
The unlabeled data for our experiments was taken from the Reuters corpus, TIPSTER corpus (LDC93T3C) and the English Gigaword corpus, third edition (LDC2007T07).
As regards the TIP-
1The second-order encoding used in our NER experiments is the same as that described in (Sha and Pereira, 2003) except removing IOB-tag of previous position label.
Training
Development
# oflabels
Data set
Training Development
Table 1: Details of training, development, and test data (labeled data set) used in our experiments
abbr.
(time period)
# of sent.
Gigaword
Table 2: Unlabeled data used in our experiments
STER corpus, we extracted all the Wall Street Journal articles published between 1990 and 1992.
With the English Gigaword corpus, we extracted articles from five news sources published between 1994 and 1996.
The unlabeled data used in this paper is detailed in Table 2.
Note that the total size of the unla-beled data reaches 1G-words (one billion tokens).
We used the same graph structure as the linear chain CRF for JESS-CM.
As regards the design of the feature functions fi, Table 3 shows the feature templates used in our experiments.
In the table, s indicates a focused token position.
Xs-i:s represents the bi-gram of feature X obtained from s — 1 and s positions.
[Xu}B=A indicates that u ranges from A to B. For example, {Xu}U=2s_2 is equal to five feature templates, {Xs-2, Xs-i, Xs,Xs+i,Xs+2}. 'word type' or wtp represents features of a word such as capitalization, the existence of digits, and punctuation as shown in (Sutton et al., 2006) without regular expressions.
Although it is common to use external
Table 3: Feature templates used in our experiments
(a) Influence of n in Dirichlet prior
(b) Changes in performance and convergence property
Figure 2: Typical behavior of tunable parameters
resources such as gazetteers for NER, we used none.
All our features can be automatically extracted from the given training data.
We used first order HMMs for embedded joint PMs since we assume that they have the same graph structure as JESS-CM as described in Section 2.2.
To reduce the required human effort, we simply used the feature templates shown in Table 3 to generate the features of the HMMs.
With our design, one feature template corresponded to one HMM.
This design preserves the feature whereby each HMM emits a single symbol from a single state (or transition).
We can easily ignore overlapping features that appear in a single HMM.
As a result, 47, 39 and 79 distinct HMMs are embedded in the potential functions of JESS-CM for POS tagging, chunking and NER experiments, respectively.
3.4 Tunable Parameters
In our experiments, we selected Gaussian and Dirichlet priors as the prior distributions in L1 and
L2, respectively.
This means that JESS-CM has two tunable parameters, <r2 and rj, in the Gaussian and Dirichlet priors, respectively.
The values of these tunable parameters are chosen by employing a binary line search.
We used the value for the best performance with the development set2.
However, it may be computationally unrealistic to retrain the entire procedure several times using lG-words of unla-beled data.
Therefore, these tunable parameter values are selected using a relatively small amount of unlabeled data (17M-words), and we used the selected values in all our experiments.
The left graph in Figure 2 shows typical r behavior.
The left end is equivalent to optimizing L2 without a prior, and the right end is almost equivalent to considering Pj (xj, y) for all j to be a uniform distribution.
This is why it appears to be bounded by the performance obtained from supervised CRF.
We omitted the influence of <r2 because of space constraints, but its behavior is nearly the same as that of supervised CRF.
Unfortunately, L2(0\X') may have two or more local maxima.
Our parameter estimation procedure does not guarantee to provide either the global optimum or a convergence solution in 0 and X' space.
An example of non-convergence is the oscillation of the estimated 0.
That is, 0 traverses two or more local maxima.
Therefore, we examined its convergence property experimentally.
The right graph in Figure 2 shows a typical convergence property.
Fortunately, in all our experiments, JESS-CM converged in a small number of iterations.
No oscillation is observed here.
4 Results and Discussion
4.1 Impact of Unlabeled Data Size
Table 4 shows the performance of JESS-CM using 1G-words of unlabeled data and the performance gain compared with supervised CRF, which is trained under the same conditions as JESS-CM except that joint PMs are not incorporated.
We emphasize that our model achieved these large improvements solely using unlabeled data as additional resources, without introducing a sophisticated model, deep feature engineering, handling external hand-
2Since CoNLL'00 shared task data has no development set, we divided the labeled training data into two distinct sets, 4/5 for training and the remainder for the development set, and determined the tunable parameters in preliminary experiments.
tunable parameter values: [Log-scale
measures
label accuracy
sent. acc.
entire sent. acc.
eval. data
dev. test
Table 4: Results for POS tagging (PTB III data), syntactic chunking (CoNLL'00 data), and NER (CoNLL'03 data) incorporated with lG-words of unlabeled data, and the performance gain from supervised CRF
(ratio against Unlabeled boning dat_a size)
^ Supervised CRF TTTT performance (test) ■
(b) Syntactic chunking
(ratio against the labeled training data size)
Unlabeled data size (Mega words): [log-scale]
Figure 3: Performance changes with respect to unlabeled data size in JESS-CM
97.
crafted resources, or task dependent human knowledge (except for the feature design).
Our method can greatly reduce the human effort needed to obtain a high performance tagger or chunker.
Figure 3 shows the learning curves of JESS-CM with respect to the size of the unlabeled data, where the x-axis is on the logarithmic scale of the unla-beled data size (Mega-word).
The scale at the top of the graph shows the ratio of the unlabeled data size to the labeled data size.
We observe that a small amount of unlabeled data hardly improved the performance since the supervised CRF results are competitive.
It seems that we require at least dozens of times more unlabeled data than labeled training data to provide a significant performance improvement.
The most important and interesting behavior is that the performance improvements against the unlabeled data size are almost linear on a logarithmic scale within the size of the unlabeled data used in our experiments.
Moreover, there is a possibility that the performance is still unsaturated at the 1G-word unlabeled data point.
This suggests that increasing the unlabeled data in JESS-CM may further improve the performance.
garded as a variant of the MML estimation (see Section 2.2), namely, it is MML estimation with a bias, A(x, y), and smooth factors, XI+j.
MML estimation can be seen as modeling P(x) since it is equivalent to maximizing £m log p(xm) with marginalized hidden variables y, where £yey p(x, y) = P(x).
Generally, more data will lead to a more accurate model of P(x).
With our method, as with modeling P(x) in MML estimation, more unlabeled data is preferable since it may provide more accurate modeling.
This also means that it provides better 'clusters' over the output space since Y is used as hidden states in HMMs.
These are intuitive explanations as to why more unlabeled data in JESS-CM produces better performance.
4.2 Expected Performance for Unseen Data
We try to investigate the impact of unlabeled data on the performance of unseen data.
We divide the test set (or the development set) into two disjoint sets: L.app and L.neg app.
L.app is a set of sentences constructed by words that all appeared in the Labeled training data.
L.-iapp is a set of sentences that have at least one word that does not appear in the Labeled training data.
Table 5 shows the performance with these two sets obtained from both supervised CRF and JESS-CM with 1G-word unlabeled data.
As the supervised CRF results, the performance of the L.-app sets is consistently much lower than that of the cor-
eval. data rates of sentences
supervised CRF (baseline)
(gain from supervised CRF)
Table 5: Comparison with L.-iapp and L.app sets obtained from both supervised CRF and JESS-CM with lG-word unlabeled data evaluated by the entire sentence accuracies, and the ratio of U.app.
unlab. data
#sent.
lG-word unlabeled data
crude company name detector
responding L.app sets.
Moreover, we can observe that the ratios of L.-app are not so small; nearly half (46.1% and 40.4%) in the PTB III data, and more than half (70.7%, 54.3% and 64.3%) in CoNLL'00 and '03 data, respectively.
This indicates that words not appearing in the labeled training data are really harmful for supervised learning.
Although the performance with L. -app sets is still poorer than with L.app sets, the JESS-CM results indicate that the introduction of unlabeled data effectively improves the performance of L. -app sets, even more than that of L.app sets.
These improvements are essentially very important; when a tagger and chunker are actually used, input data can be obtained from anywhere and this may mostly include words that do not appear in the given labeled training data since the labeled training data is limited and difficult to increase.
This means that the improved performance of L.-app can link directly to actual use.
Table 5 also shows the ratios of sentences that are constructed from words that all appeared in the lG-word Unlabeled data used in our experiments (U.app) in the L.-app and L.app.
This indicates that most of the words in the development or test sets are covered by the lG-word unlabeled data.
This may be the main reason for JESS-CM providing large performance gains for both the overall and L.-app set performance of all three tasks.
Table 6 shows the relation between JESS-CM performance and U.app in the NER experiments.
The development data and test data were obtained from
for PTB III data evaluated by label accuracy
15M-word unlabeled data
17M-word unlabeled data
full parser output
[supervised CRF (baseline)]
Table 8: Syntactic chunking results of the previous top systems for CoNLL'00 shared task data (F^=1 score)
30-31 Aug.
1996 and 6-7 Dec.
1996 Reuters news articles, respectively.
We find that temporal proximity leads to better performance.
This aspect can also be explained as U.app.
Basically, the U.app increase leads to improved performance.
The evidence provided by the above experiments implies that increasing the coverage of unlabeled data offers the strong possibility of increasing the expected performance of unseen data.
Thus, it strongly encourages us to use an SSL approach that includes JESS-CM to construct a general tagger and chunker for actual use.
5 Comparison with Previous Top Systems and Related Work
In POS tagging, the previous best performance was reported by (Shen et al., 2007) as summarized in Table 7.
Their method uses a novel sophisticated model that learns both decoding order and labeling, while our model uses a standard first order Markov model.
Despite using such a simple model, our method can provide a better result with the help of unlabeled data.
dev.
additional resources
1G-word unlabeled data
37M-word unlabeled data
own large gazetteers, 2M-word labeled data
27M-word unlabeled data
Table 9: NER results of the previous top systems for CoNLL'03 shared task data evaluated by F^=i score
As shown in Tables 8 and 9, the previous best performance for syntactic chunking and NER was reported by (Ando and Zhang, 2005), and is referred to as 'ASO-semi'.
ASO-semi also incorporates unlabeled data solely as additional information in the same way as JESS-CM.
ASO-semi uses unlabeled data for constructing auxiliary problems that are expected to capture a good feature representation of the target problem.
As regards syntactic chunking, JESS-CM significantly outperformed ASO-semi for the same 15M-word unlabeled data size obtained from the Wall Street Journal in 1991 as described in (Ando and Zhang, 2005).
Unfortunately with NER, JESS-CM is slightly inferior to ASO-semi for the same 27M-word unlabeled data size extracted from the Reuters corpus.
In fact, JESS-CM using 37M-words of unlabeled data provided a comparable result.
We observed that ASO-semi prefers 'nugget extraction' tasks to 'field segmentation' tasks (Grenager et al., 2005).
We cannot provide details here owing to the space limitation.
Intuitively, their word prediction auxiliary problems can capture only a limited number of characteristic behaviors because the auxiliary problems are constructed by a limited number of 'binary' classifiers.
Moreover, we should remember that ASO-semi used the human knowledge that 'named entities mostly consist of nouns or adjectives' during the auxiliary problem construction in their NER experiments.
In contrast, our results require no such additional knowledge or limitation.
In addition, the design and training of auxiliary problems as well as calculating SVD are too costly when the size of the unlabeled data increases.
These facts imply that our SSL framework is rather appropriate for handling large scale unlabeled data.
On the other hand, ASO-semi and JESS-CM have an important common feature.
That is, both meth-
ods discriminatively combine models trained by using unlabeled data in order to create informative feature representation for discriminative learning.
Unlike self/co-training approaches (Blum and Mitchell, 1998), which use estimated labels as 'correct labels', this approach automatically judges the reliability of additional features obtained from unlabeled data in terms of discriminative training.
Ando and Zhang (2007) have also pointed out that this methodology seems to be one key to achieving higher performance in NLP applications.
There is an approach that combines individually and independently trained joint PMs into a discriminative model (Li and McCallum, 2005).
There is an essential difference between this method and JESS-CM.
We categorize their approach as an 'indirect approach' since the outputs of the target task, y, are not considered during the unlabeled data incorporation.
Note that ASO-semi is also an 'indirect approach'.
On the other hand, our approach is a 'direct approach' because the distribution of y obtained from JESS-CM is used as 'seeds' of hidden states during MDF estimation for join PM parameters (see Section 4.1).
In addition, MDF estimation over unlabeled data can effectively incorporate the 'labeled' training data information via a 'bias' since X included in A(x, y) is estimated from labeled training data.
6 Conclusion
We proposed a simple yet powerful semi-supervised conditional model, which we call JESS-CM.
It is applicable to large amounts of unlabeled data, for example, at the giga-word level.
Experimental results obtained by using JESS-CM incorporating 1G-words of unlabeled data have provided the current best performance as regards POS tagging, syntactic chunking, and NER for widely used large test collections such as PTB III, CoNLL'00 and '03 shared task data, respectively.
We also provided evidence that the use of more unlabeled data in SSL can lead to further improvements.
Moreover, our experimental analysis revealed that it may also induce an improvement in the expected performance for unseen data in terms of the unlabeled data coverage.
Our results may encourage the adoption ofthe SSL method for many other real world applications.
