Feature-Frequency–Adaptive On-line Training for Fast and Accurate Natural Language Processing

Training speed is an important concern of natural language processing (NLP) systems. Large-scale NLP systems are computationally expensive. In many real-world applications, we further need to optimize high-dimensional model parameters. For example, the state-of-the-art word segmentation system uses more than 40 million features (Sun, Wang, and Li 2012). The heavy NLP models together with high-dimensional parameters lead to a challenging problem on model training, which may require week-level training time even with fast computing machines.

Accuracy is another very important concern of NLP systems. Nevertheless, usually it is quite difficult to build a system that has fast training speed and at the same time has high accuracy. Typically we need to make a tradeoff between speed and accuracy, to trade training speed for higher accuracy or vice versa. In this work, we have tried to overcome this problem: to improve the training speed and the model accuracy at the same time.

There are two major approaches for parameter training: batch and on-line. Standard gradient descent methods are normally batch training methods, in which the gradient computed by using all training instances is used to update the parameters of the model. The batch training methods include, for example, steepest gradient descent, conjugate gradient descent (CG), and quasi-Newton methods like limited-memory BFGS (Nocedal and Wright 1999). The true gradient is usually the sum of the gradients from each individual training instance. Therefore, batch gradient descent requires the training method to go through the entire training set before updating parameters. This is why batch training methods are typically slow.

On-line learning methods can significantly accelerate the training speed compared with batch training methods. A representative on-line training method is the stochastic gradient descent method (SGD) and its extensions (e.g., stochastic meta descent) (Bottou 1998; Vishwanathan et al. 2006). The model parameters are updated more frequently compared with batch training, and fewer passes are needed before convergence. For large-scale data sets, on-line training methods can be much faster than batch training methods.

However, we find that the existing on-line training methods are still not good enough for training large-scale NLP systems—probably because those methods are not well-tailored for NLP systems that have massive features. First, the convergence speed of the existing on-line training methods is not fast enough. Our studies show that the existing on-line training methods typically require more than 50 training passes before empirical convergence, which is still slow. For large-scale NLP systems, the training time per pass is typically long and fast convergence speed is crucial. Second, the accuracy of the existing on-line training methods is not good enough. We want to further improve the training accuracy. We try to deal with the two challenges at the same time. Our goal is to develop a new training method for faster and at the same time more accurate natural language processing.

In this article, we present a new on-line training method, adaptive on-line gradient descent based on feature frequency information (ADF),¹ for very accurate and fast on-line training of NLP systems. Other than the high training accuracy and fast training speed, we further expect that the proposed training method has good theoretical properties. We want to prove that the proposed method is convergent and has a fast convergence rate.

In the proposed ADF training method, we use a learning rate vector in the on-line updating. This learning rate vector is automatically adapted based on feature frequency information in the training data set. Each model parameter has its own learning rate adapted on feature frequency information. This proposal is based on the simple intuition that a feature with higher frequency in the training process should have a learning rate that decays faster. This is because a higher frequency feature is expected to be well optimized with higher confidence. Thus, a higher frequency feature is expected to have a lower learning rate. We systematically formalize this intuition into a theoretically sound training algorithm, ADF.

The main contributions of this work are as follows:

• On the methodology side, we propose a general purpose on-line training method, ADF. The ADF method is significantly more accurate than existing on-line and batch training methods, and has faster training speed. Moreover, theoretical analysis demonstrates that the ADF method is convergent with a fast convergence rate.

• On the application side, for the three well-known tasks, including named entity recognition, word segmentation, and phrase chunking, the proposed simple method achieves equal or even better accuracy than the existing gold-standard systems, which are complicated and use extra resources.

Our main focus is on structured classification models with high dimensional features. For structured classification, the conditional random fields model is widely used. To illustrate that the proposed method is a general-purpose training method not limited to a specific classification task or model, we also evaluate the proposal for non-structured classification tasks like binary classification. For non-structured classification, the maximum entropy model (Berger, Della Pietra, and Della Pietra 1996; Ratnaparkhi 1996) is widely used. Here, we review the conditional random fields model and the related work of on-line training methods.

Other well-known on-line training methods include perceptron training (Freund and Schapire 1999), averaged perceptron training (Collins 2002), more recent development/extensions of stochastic gradient descent (e.g., the second-order stochastic gradient descent training methods like stochastic meta descent) (Vishwanathan et al. 2006; Hsu et al. 2009), and so on. However, the second-order stochastic gradient descent method requires the computation or approximation of the inverse of the Hessian matrix of the objective function, which is typically slow, especially for heavily structured classification models. Usually the convergence speed based on number of training iterations is moderately faster, but the time cost per iteration is slower. Thus the overall time cost is still large.

Compared with the related work on batch and on-line training (Jacobs 1988; Sperduti and Starita 1993; Dredze, Crammer, and Pereira 2008; Duchi, Hazan, and Singer 2010; McMahan and Streeter 2010), our work is fundamentally different. The proposed ADF training method is based on feature frequency adaptation, and to the best of our knowledge there is no prior work on direct feature-frequency–adaptive on-line training. Compared with the confidence-weighted (CW) classification method and its variation AROW (Dredze, Crammer, and Pereira 2008; Crammer, Kulesza, and Dredze 2009), the proposed method is substantially different. While the feature frequency information is implicitly modeled via a complicated Gaussian distribution framework in Dredze, Crammer, and Pereira (2008) and Crammer, Kulesza, and Dredze (2009), the frequency information is explicitly modeled in our proposal via simple learning rate adaptation. Our proposal is more straightforward in capturing feature frequency information, and it has no need to use Gaussian distributions and KL divergence, which are important in the CW and AROW methods. In addition, our proposal is a probabilistic learning method for training probabilistic models such as CRFs, whereas the CW and AROW methods (Dredze, Crammer, and Pereira 2008; Crammer, Kulesza, and Dredze 2009) are non-probabilistic learning methods extended from perceptron-style approaches. Thus, the framework is different. This work is a substantial extension of the conference version (Sun, Wang, and Li 2012). Sun, Wang, and Li (2012) focus on the specific task of word segmentation, whereas this article focuses on the proposed training algorithm.

In traditional on-line optimization methods such as SGD, no distinction is made for different parameters in terms of the learning rate, and this may result in slow convergence of the model training. For example, in the on-line training process, suppose the high frequency feature f₁ and the low frequency feature f₂ are observed in a training sample and their corresponding parameters w₁ and w₂ are to be updated via the same learning rate γ_t. Suppose the high frequency feature f₁ has been updated 100 times and the low frequency feature f₂ has only been updated once. Then, it is possible that the weight w₁ is already well optimized and the learning rate γ_t is too aggressive for updating w₁. Updating the weight w₁ with the learning rate γ_t may make w₁ be far from the well-optimized value, and it will require corrections in the future updates. This causes fluctuations in the on-line training and results in slow convergence speed. On the other hand, it is possible that the weight w₂ is poorly optimized and the same learning rate γ_t is too conservative for updating w₂. This also results in slow convergence speed.

To solve this problem, we propose ADF. In spite of the high accuracy and fast convergence speed, the proposed method is easy to implement. The proposed method with feature-frequency–adaptive learning rates can be seen as a learning method with specific diagonal approximation of the Hessian information based on assumptions of feature frequency information. In this approximation, the diagonal elements of the diagonal matrix correspond to the feature-frequency–adaptive learning rates. According to the aforementioned example and analysis, it assumes that a feature with higher frequency in the training process should have a learning rate that decays faster.

In the proposed ADF method, we try to use more refined learning rates than traditional SGD training. Instead of using a single learning rate (a scalar) for all weights, we extend the learning rate scalar to a learning rate vector, which has the same dimension as the weight vector . The learning rate vector is automatically adapted based on feature frequency information. By doing so, each weight has its own learning rate, and we will show that this can significantly improve the convergence speed of on-line learning.

The learning rate vector is automatically adapted based on feature frequency information in the updating process. Intuitively, a feature with higher frequency in the training process has a learning rate that decays faster. This is because a weight with higher frequency is expected to be more adequately trained, hence a lower learning rate is preferable for fast convergence. We assume that a high frequency feature should have a lower learning rate, and a low frequency feature should have a relatively higher learning rate in the training process. We systematically formalize this idea into a theoretically sound training algorithm. The proposed method with feature-frequency–adaptive learning rates can be seen as a learning method with specific diagonal approximation of the inverse of the Hessian matrix based on feature frequency information.

The ADF training method is efficient because the only additional computation (compared with traditional SGD) is the derivation of the learning rates, which is simple and efficient. As we know, the regularization of SGD can perform efficiently via the optimization based on sparse features (Shalev-Shwartz, Singer, and Srebro 2007). Similarly, the derivation of can also perform efficiently via the optimization based on sparse features. Note that although binary features are common in natural language processing tasks, the ADF algorithm is not limited to binary features and it can be applied to real-valued features.

We want to show that the proposed ADF learning algorithm has good convergence properties. There are two steps in the convergence analysis. First, we show that the ADF update rule is a contraction mapping. Then, we show that the ADF training is asymptotically convergent, and with a fast convergence rate.

To make convergence analysis of the proposed ADF training algorithm, we need to introduce several mathematical definitions. First, we introduce Lipschitz continuity:

Definition 1 (Lipschitz continuity)

A function is Lipschitz continuous with the degree of D if |F(x) − F(y)| ≤ D|x − y| for . can be multi-dimensional space, and |x − y| is the distance between the points x and y.

Based on the definition of Lipschitz continuity, we give the definition of the Lipschitz constant ||F||_Lip as follows:

Further, based on the definition of Lipschitz constant, we give the definition of contraction mapping as follows:

Definition 3 (Contraction mapping)

A function is a contraction mapping if its Lipschitz constant is smaller than 1: ||F||_Lip < 1.

Then, we can show that the traditional SGD update is a contraction mapping.

Lemma 1 (SGD update rule is contraction mapping)

Let γ be a fixed low learning rate in SGD updating. If , the SGD update rule is a contraction mapping in Euclidean space with Lipschitz continuity degree 1 − γ/σ².

The proof can be extended from the related work on convergence analysis of parallel SGD training (Zinkevich et al. 2010). The stochastic training process is a one-following-one dynamic update process. In this dynamic process, if we use the same update rule F, we have and . It is only necessary to prove that the dynamic update is a contraction mapping restricted by this one-following-one dynamic process. That is, for the proposed ADF update rule, it is only necessary to prove it is a dynamic contraction mapping. We formally define dynamic contraction mapping as follows.

Definition 4 (Dynamic contraction mapping)

Given a function , suppose the function is used in a dynamic one-following-one process: x_{t + 1} = F(x_t) and x_{t + 2} = F(x_{t + 1}) for . Then, the function F is a dynamic contraction mapping if ∃D < 1, |x_{t + 2} − x_{t + 1}| ≤ D|x_{t + 1} − x_t| for .

We can see that a contraction mapping is also a dynamic contraction mapping, but a dynamic contraction mapping is not necessarily a contraction mapping. We first show that the ADF update rule with a fixed learning rate vector of different learning rates is a dynamic contraction mapping.

Theorem 1 (ADF update rule with fixed learning rates)

Let be a fixed learning rate vector with different learning rates. Let γ_max be the maximum learning rate in the learning rate vector : . Then if , the ADF update rule is a dynamic contraction mapping in Euclidean space with Lipschitz continuity degree .

The proof is sketched in Section 5.

Further, we need to prove that the ADF update rule with a decaying learning rate vector is a dynamic contraction mapping, because the real ADF algorithm has a decaying learning rate vector. In the decaying case, the condition that can be easily achieved, because continues to decay with an exponential decaying rate. Even if the is initialized with high values of learning rates, after a number of training passes (denoted as T) is guaranteed to be small enough so that and . Without losing generality, our convergence analysis starts from the pass T and we take as in the following analysis. Thus, we can show that the ADF update rule with a decaying learning rate vector is a dynamic contraction mapping:

Theorem 2 (ADF update rule with decaying learning rates)

Let be a learning rate vector in the ADF learning algorithm, which is decaying over the time t and with different decaying rates based on feature frequency information. Let start from a low enough learning rate vector such that , where γ_max is the maximum element in . Then, the ADF update rule with decaying learning rate vector is a dynamic contraction mapping in Euclidean space with Lipschitz continuity degree .

The proof is sketched in Section 5.

Based on the connections between ADF training and contraction mapping, we demonstrate the convergence properties of the ADF training method. First, we prove the convergence of the ADF training.

Theorem 3 (ADF convergence)

ADF training is asymptotically convergent.

The proof is sketched in Section 5.

Theorem 4 (ADF convergence rate)

The proof is sketched in Section 5.

The convergence analysis demonstrates that the proposed method with feature-frequency-adaptive learning rates is convergent and the bound of convergence rate is analyzed. It demonstrates that increasing the values of and β leads to a lower bound of the convergence rate. Because the bound of the convergence rate is just an up-bound rather than the actual convergence rate, we still need to conduct automatic tuning of the hyper-parameters, including and β, for optimal convergence rate in practice. The ADF training method has a fast convergence rate because the feature-frequency-adaptive schema can avoid the fluctuations on updating the weights of high frequency features, and it can avoid the insufficient training on updating the weights of low frequency features. In the following sections, we perform experiments to confirm the fast convergence rate of the proposed method.

Our main focus is on training heavily structured classification models. We evaluate the proposal on three NLP structured classification tasks: biomedical named entity recognition (Bio-NER), Chinese word segmentation, and noun phrase (NP) chunking. For the structured classification tasks, the ADF training is based on the CRF model (Lafferty, McCallum, and Pereira 2001). Further, to demonstrate that the proposed method is not limited to structured classification tasks, we also perform experiments on a non-structured binary classification task: sentiment-based text classification. For the non-structured classification task, the ADF training is based on the maximum entropy model (Berger, Della Pietra, and Della Pietra 1996; Ratnaparkhi 1996).

The biomedical named entity recognition (Bio-NER) task is from the BIONLP-2004 shared task. The task is to recognize five kinds of biomedical named entities, including DNA, RNA, protein, cell line, and cell type, on the MEDLINE biomedical text mining corpus (Kim et al. 2004). A typical approach to this problem is to cast it as a sequential labeling task with the BIO encoding.

This data set consists of 20,546 training samples (from 2,000 MEDLINE article abstracts, with 472,006 word tokens) and 4,260 test samples. The properties of the data are summarized in Table 1. State-of-the-art systems for this task include Settles (2004), Finkel et al. (2004), Okanohara et al. (2006), Hsu et al. (2009), Sun, Matsuzaki, et al. (2009), and Tsuruoka, Tsujii, and Ananiadou (2009).

Following previous studies for this task (Okanohara et al. 2006; Sun, Matsuzaki, et al. 2009), we use word token–based features, part-of-speech (POS) based features, and orthography pattern–based features (prefix, uppercase/lowercase, etc.), as listed in Table 2. With the traditional implementation of CRF systems (e.g., the HCRF package), the edges features usually contain only the information of y_{i − 1} and y_i, and ignore the information of the observation sequence (i.e., ). The major reason for this simple realization of edge features in traditional CRF implementation is to reduce the dimension of features. To improve the model accuracy, we utilize rich edge features following Sun, Wang, and Li (2012), in which local observation information of is combined in edge features just like the implementation of node features. A detailed introduction to rich edge features can be found in Sun, Wang, and Li (2012). Using the feature templates, we extract a high dimensional feature set, which contains 5.3 ×10⁷ features in total. Following prior studies, the evaluation metric for this task is the balanced F-score defined as 2PR/(P + R), where P is precision and R is recall.

Chinese word segmentation aims to automatically segment character sequences into word sequences. Chinese word segmentation is important because it is the first step for most Chinese language information processing systems. Our experiments are based on the Microsoft Research data provided by The Second International Chinese Word Segmentation Bakeoff. In this data set, there are 8.8×10⁴ word-types, 2.4×10⁶ word-tokens, 5×10³ character-types, and 4.1×10⁶ character-tokens. State-of-the-art systems for this task include Tseng et al. (2005), Zhang, Kikui, and Sumita (2006), Zhang and Clark (2007), Gao et al. (2007), Sun, Zhang, et al. (2009), Sun (2010), Zhao et al. (2010), and Zhao and Kit (2011).

The feature engineering follows previous work on word segmentation (Sun, Wang, and Li 2012). Rich edge features are used. For the classification label y_i and the label transition y_{i − 1}y_i on position i, we use the feature templates as follows (Sun, Wang, and Li 2012):

• Character unigrams located at positions i − 2, i − 1, i, i + 1, and i + 2.

• Character bigrams located at positions i − 2, i − 1, i and i + 1.

• Whether x_j and x_{j + 1} are identical, for j = i − 2, …, i + 1.

• Whether x_j and x_{j + 2} are identical, for j = i − 3, …, i + 1.

• The character sequence x_j,i if it matches a word , with the constraint i − 6 < j < i. The item x_j,i represents the character sequence x_j…x_i. represents the unigram-dictionary collected from the training data.

• The character sequence x_i,k if it matches a word , with the constraint i < k < i + 6.

• The word bigram candidate [x_{j,i − 1}, x_i,k] if it hits a word bigram , and satisfies the aforementioned constraints on j and k. represents the word bigram dictionary collected from the training data.

• The word bigram candidate [x_j,i, x_{i + 1,k}] if it hits a word bigram , and satisfies the aforementioned constraints on j and k.

All feature templates are instantiated with values that occurred in training samples. The extracted feature set is large, and there are 2.4×10⁷ features in total. Our evaluation is based on a closed test, and we do not use extra resources. Following prior studies, the evaluation metric for this task is the balanced F-score.

In the phrase chunking task, the non-recursive cores of noun phrases, called base NPs, are identified. The phrase chunking data is extracted from the data of the CoNLL-2000 shallow-parsing shared task (Sang and Buchholz 2000). The training set consists of 8,936 sentences, and the test set consists of 2,012 sentences. We use the feature templates based on word n-grams and part-of-speech n-grams, and feature templates are shown in Table 3. Rich edge features are used. Using the feature templates, we extract 4.8 ×10⁵ features in total. State-of-the-art systems for this task include Kudo and Matsumoto (2001), Collins (2002), McDonald, Crammer, and Pereira (2005), Vishwanathan et al. (2006; Sun et al. (2008), and Tsuruoka, Tsujii, and Ananiadou (2009). Following prior studies, the evaluation metric for this task is the balanced F-score.

To demonstrate that the proposed method is not limited to structured classification, we select a well-known sentiment classification task for evaluating the proposed method on non-structured classification.

Generally, sentiment classification classifies user review text as a positive or negative opinion. This task (Blitzer, Dredze, and Pereira 2007) consists of four subtasks based on user reviews from Amazon.com. Each subtask is a binary sentiment classification task based on a specific topic. We use the maximum entropy model for classification. We use the same lexical features as those used in Blitzer, Dredze, and Pereira (2007), and the total number of features is 9.4 ×10⁵. Following prior work, the evaluation metric is binary classification accuracy.

As for training, we perform gradient descent with the proposed ADF training method. To compare with existing literature, we choose four popular training methods, a representative batch training method, and three representative on-line training methods. The batch training method is the limited-memory BFGS (LBFGS) method (Nocedal and Wright 1999), which is considered to be one of the best optimizers for log-linear models like CRFs. The on-line training methods include the SGD training method, which we introduced in Section 2.2, the structured perceptron (Perc) training method (Freund and Schapire 1999; Collins 2002), and the averaged perceptron (Avg-Perc) training method (Collins 2002). The structured perceptron method and averaged perceptron method are non-probabilistic training methods that have very fast training speed due to the avoidance of the computation on gradients (Sun, Matsuzaki, and Li 2013). All training methods, including ADF, SGD, Perc, Avg-Perc, and LBFGS, use the same set of features.

We also compared the ADF method with the CW method (Dredze, Crammer, and Pereira 2008) and the AROW method (Crammer, Kulesza, and Dredze 2009). The CW and AROW methods are implemented based on the Confidence Weighted Learning Library.² Because the current implementation of the CW and AROW methods do not utilize rich edge features, we removed the rich edge features in our systems to make more fair comparisons. That is, we removed rich edge features in the CRF-ADF setting, and this simplified method is denoted as ADF-noRich. The second-order stochastic gradient descent training methods, including the SMD method (Vishwanathan et al. 2006) and the PSA method (Hsu et al. 2009), are not considered in our experiments because we find those methods are quite slow when running on our data sets with high dimensional features.

We find that the settings of q, α, and β in the ADF training method are not sensitive among specific tasks and can be generally set. We simply set q = n/10 (n is the number of training samples). It means that feature frequency information is updated 10 times per iteration. Via cross-validation only on the training data of different tasks, we find that the following setting is sufficient to produce adequate performance for most of the real-world natural language processing tasks: α around 0.995 and β around 0.6. This indicates that the feature frequency information has similar characteristics across many different natural language processing tasks.

Thus, we simply use the following setting for all tasks: q = n/10, α = 0.995, and β = 0.6. This leaves c (the initial value of the learning rates) as the only hyper-parameter that requires careful tuning. We perform automatic tuning for c based on the training data via 4-fold cross-validation, testing with c = 0.005, 0.01, 0.05, 0.1, respectively, and the optimal c is chosen based on the best accuracy of cross-validation. Via this automatic tuning, we find it is proper to set c = 0.005, 0.1, 0.05, 0.005, for the Bio-NER, word segmentation, phrase chunking, and sentiment classification tasks, respectively.

To reduce overfitting, we use an L₂ Gaussian weight prior (Chen and Rosenfeld 1999) for the ADF, LBFGS, and SGD training methods. We vary the σ with different values (e.g., 1.0, 2.0, and 5.0) for 4-fold cross validation on the training data of different tasks, and finally set σ = 5.0 for all training methods in the Bio-NER task; σ = 5.0 for all training methods in the word segmentation task; σ = 5.0, 1.0, 1.0 for ADF, SGD, and LBFGS in the phrase chunking task; and σ = 1.0 for all training methods in the sentiment classification task. Experiments are performed on a computer with an Intel(R) Xeon(R) 2.0-GHz CPU.

First, we check the experimental results of different methods on their empirical convergence state. Because the perceptron training method (Perc) does not achieve empirical convergence even with a very large number of training passes, we simply report its results based on a large enough number of training passes (e.g., 200 passes). Experimental results are shown in Table 4.

As we can see, the proposed ADF method is more accurate than other training methods, either the on-line ones or the batch one. It is a bit surprising that the ADF method performs even more accurately than the batch training method (LBFGS). We notice that some previous work also found that on-line training methods could have better performance than batch training methods such as LBFGS (Tsuruoka, Tsujii, and Ananiadou 2009; Schaul, Zhang, and LeCun 2012). The ADF training method can achieve better results probably because the feature-frequency–adaptive training schema can produce more balanced training of features with diversified frequencies. Traditional SGD training may over-train high frequency features and at the same time may have insufficient training of low frequency features. The ADF training method can avoid such problems. It will be interesting to perform further analysis in future work.

We also performed significance tests based on t-tests with a significance level of 0.05. Significance tests demonstrate that the ADF method is significantly more accurate than the existing training methods in most of the comparisons, whether on-line or batch. For the Bio-NER task, the differences between ADF and LBFGS, SGD, Perc, and Avg-Perc are significant. For the word segmentation task, the differences between ADF and LBFGS, SGD, Perc, and Avg-Perc are significant. For the phrase chunking task, the differences between ADF and Perc and Avg-Perc are significant; the differences between ADF and LBFGS and SGD are non-significant.

Moreover, as we can see, the proposed method achieves a convergence state with the least number of training passes, and with the least wall-clock time. In general, the ADF method is about one order of magnitude faster than the LBFGS batch training method and several times faster than the existing on-line training methods.

The three tasks are well-known benchmark tasks with standard data sets. There is a large amount of published research on those three tasks. We compare the proposed method with the state-of-the-art systems. The comparisons are shown in Table 5.

As we can see, our system is competitive with the best systems for the Bio-NER, word segmentation, and NP-chunking tasks. Many of the state-of-the-art systems use extra resources (e.g., linguistic knowledge) or complicated systems (e.g., voting over multiple models). Thus, it is impressive that our single model–based system without extra resources achieves good performance. This indicates that the proposed ADF training method can train model parameters with good generality on the test data.

To study the detailed training process and convergence speed, we show the training curves in Figures 2,³–4. Figure 2 focuses on the comparisons between the ADF method and the existing on-line training methods. As we can see, the ADF method converges faster than other on-line training methods in terms of both training passes and wall-clock time. The ADF method has roughly the same training speed per pass compared with traditional SGD training.

Figure 3 (Top Row) focuses on comparing the ADF method with the CW method (Dredze, Crammer, and Pereira 2008) and the AROW method (Crammer, Kulesza, and Dredze 2009). Comparisons are based on similar features. As discussed before, the ADF-noRich method is a simplified system, with rich edge features removed from the CRF-ADF system. As we can see, the proposed ADF method, whether with or without rich edge features, outperforms the CW and AROW methods. Figure 3 (Bottom Row) focuses on the comparisons with different mini-batch (the training samples in each stochastic update) sizes. Representative results with a mini-batch size of 10 are shown. In general, we find larger mini-batch sizes will slow down the convergence speed. Results demonstrate that, compared with the SGD training method, the ADF training method is less sensitive to mini-batch sizes.

Figure 4 focuses on the comparisons between the ADF method and the batch training method LBFGS. As we can see, the ADF method converges at least one order magnitude faster than the LBFGS training in terms of both training passes and wall-clock time. For the LBFGS training, we need to determine the LBFGS memory parameter m, which controls the number of prior gradients used to approximate the Hessian information. A larger value of m will potentially lead to more accurate estimation of the Hessian information, but at the same time will consume significantly more memory. Roughly, the LBFGS training consumes m times more memory than the ADF on-line training method. For most tasks, the default setting of m = 10 is reasonable. We set m = 10 for the word segmentation and phrase chunking tasks, and m = 6 for the Bio-NER task due to the shortage of memory for m > 6 cases in this task.

Many real-world data sets can only observe the training data in one pass. For example, some Web-based on-line data streams can only appear once so that the model parameter learning should be finished in one-pass learning (see Zinkevich et al. 2010). Hence, it is important to test the performance in the one-pass learning scenario.

In the one-pass learning scenario, the feature frequency information is computed “on the fly” during on-line training. As shown in Section 3.1, we only need to have a real-valued vector to record the cumulative feature frequency information, which is updated when observing training instances one by one. Then, the learning rate vector is updated based on the only and there is no need to observe the training instances again. This is the same algorithm introduced in Section 3.1 and no change is required for the one-pass learning scenario. Figure 5 shows the comparisons between the ADF method and baselines on one-pass learning. As we can see, the ADF method consistently outperforms the baselines. This also reflects the fast convergence speed of the ADF training method.

In previous experiments, we showed that the proposed method outperforms existing baselines on structured classification. Nevertheless, we want to show that the ADF method also has good performance on non-structured classification. In addition, this task is based on real-valued features instead of binary features.

Experimental results of different training methods on the convergence state are shown in Table 6. As we can see, the proposed method outperforms all of the on-line and batch baselines in terms of binary classification accuracy. Here again we observe that the ADF and SGD methods outperform the LBFGS baseline.

The training curves are shown in Figure 6. As we can see, the ADF method converges quickly. Because this data set is relatively small and the feature dimension is much smaller than previous tasks, we find the baseline training methods also have fast convergence speed. The comparisons on one-pass learning are shown in Figure 7. Just as for the experiments for structured classification tasks, the ADF method outperforms the baseline methods on one-pass learning, with more than 12.7% error rate reduction.

This section gives proofs of Theorems 1–4.

In this work we tried to simultaneously improve the training speed and model accuracy of natural language processing systems. We proposed the ADF on-line training method, based on the core idea that high frequency features should result in a learning rate that decays faster. We demonstrated that the ADF on-line training method is convergent and has good theoretical properties. Based on empirical experiments, we can state the following conclusions. First, the ADF method achieved the major target of this work: faster training speed and higher accuracy at the same time. Second, the ADF method was robust: It had good performance on several structured and non-structured classification tasks with very different characteristics. Third, the ADF method worked well on both binary features and real-valued features. Fourth, the ADF method outperformed existing methods in a one-pass learning setting. Finally, our method achieved state-of-the-art performance on several well-known benchmark tasks. To the best of our knowledge, our simple method achieved a much better F-score than the existing best reports on the biomedical named entity recognition task.

This work was supported by the National Natural Science Foundation of China (no. 61300063, no. 61370117), the Doctoral Fund of Ministry of Education of China (no. 20130001120004), a Hong Kong Polytechnic University internal grant (4-ZZD5), a Hong Kong RGC Project (no. PolyU 5230/08E), the National High Technology Research and Development Program of China (863 Program, no. 2012AA011101), and the Major National Social Science Fund of China (no. 12&ZD227). This work is a substantial extension of the conference version presented at ACL 2012 (Sun, Wang, and Li 2012).