Signal Processing - July 2017 - 119

probability, the probability margin between the two most likely
class labels, or the entropy of prediction [94]. In the context of
speech recognition, word posterior probabilities or the HMMstate entropy are frequently used as confidence measures [79],
[81]. When using a nonprobability model (e.g., an SVM), similar measures can be constructed from discriminant functions.
Considering the SVM as an example, pseudoprobabilistic
values can be transformed from the output distances from the
SVM hyperplane (see [17] for more details). The effectiveness
of this approach has been extensively assessed for emotion recognition from speech [83].
Despite the reported performance improvement, many studies have found that uncertainty-based AL is inclined toward
selecting noise and garbage data (i.e., outliers from the main
data distribution) for human labeling. This issue occurs even
more frequently when using AL to annotate data collected in
the wild, i.e., not under controlled laboratory conditions, where
environmental noises severely distort the speech, and many
unexpected words are potentially uttered. Labeling these outliers is usually difficult and time consuming [95]. Furthermore,
these data often offer little information on the overall system
performance [17], [95]. A straightforward solution to address
this outlier problem is to raise the threshold of a confidence
score. For example, the authors of [17] used a median uncertainty strategy instead of the least certainty one for actively
selecting spontaneously emotional utterances, which delivered
a positive performance improvement.
Sampling by uncertainty and density (SUD) is a more
sophisticated method that was introduced for ASR in [96]. In
this approach, unlabeled instances that are both near the decision boundary and very close to other examples are assumed to
be more important than those that are isolated (i.e., likely to be
outliers). Hence, SUD considers not only the most informative
data in terms of uncertainty but also the most representative
data in terms of density. That is, those data predicted with least
certainty and distributed in a low-density area are ignored.
A similar idea was proposed in [80], where the global criterion was used in ASR to maximize the expected lattice entropy reduction over all nontranscribed data. Specifically, it first
measures the entropy among the lattices generated by decoding unlabeled utterances. It then estimates the expected entropy reduction over the whole data set for each given utterance,
and selects the utterances that should deliver the highest entropy reduction for human labeling. After that, the transcribed
utterances can be weighted according to the number of similar
utterances in the whole data set to achieve better performance
for speech recognition. This algorithm is also analogous to the
error-rate reduction strategy introduced in [95].

Query by committee
This strategy uses a committee (group) of weak models (learners), denoted by H = {i 1, f, i k}, to select unlabeled data by
the principle of maximal disagreement among these models
[97]. Mathematically, this can be expressed as:
xl = argmax Q d (x; H).
x!U

(12)

The two key problems in committee-based approaches are 1)
constructing a committee H that represents competing
hypotheses and 2) defining a disagreement measurement Q d .
To alleviate the first problem, the models are usually built by
employing multiple different classifiers (e.g., HMMs, SVMs,
and RNNs) with the same training data, or by splitting the
training data or features into partitions for training several different versions of the same type of classifier, or by a combination thereof. For the second problem, the commonly used
disagreement measures are vote entropy and Kullback-Leibler
divergence (see [94] for more details). In speech recognition,
this strategy has been applied to both acoustic and language
models, resulting in a significant data annotation reduction
while achieving the same word accuracy [82].

Meta query strategies
One often deals with imbalance across classes of interest in
the data. As an example, for emotion recognition, the emotional speech of interest usually appears sparsely within a data
set, while the less interesting nonemotional speech often
appears at a much higher frequency. In this scenario, an initial
coarse model can be used to first decide which data are of
interest by distinguishing between neutral and emotional
speech. A subsequent finer model can be then used to recognize different emotions or respective other classes in other
tasks in the selected emotional speech data. An example of
such an approach is the sparse-tracking query strategy [83]. It
tracks only sparse (emotional) instances, via iterative retraining and labeling, using a novelty detection paradigm.
One issue when analyzing subjective speaker states and
traits (e.g., emotion and personality) is the requirement of multiple annotations per sample to obtain a reliable gold standard,
which linearly increases the annotation workload. Recently,
dynamic active query strategies have been shown to be successful in overcoming this issue [84]. These approaches, e.g.,
sequentially query human annotators to label a specific instance
up to the achievement of a predefined agreement level (i.e., a
certain number of votes for a specific class). The general idea
is to learn and exploit the varying reliability of raters to discern
whom to best trust and when. The results presented indicate
that this approach can contribute to a meaningful reduction of
annotation effort [84].

Cooperative learning
As discussed previously, SSL techniques can perform annotation work from machines with a bare minimum of human
intervention. However, the performance of SSL is hampered
by the issue of potential error accumulation [94]. Alternatively,
AL techniques have the potential to achieve higher accuracy
with fewer training labels by actively selecting the data it can
learn the most from. However, AL still requires a considerable
amount of human intervention.
To take advantage of the best of both approaches, it is plausible to jointly conduct AL and SSL in a unified CL framework [17]. A general CL flowchart is illustrated in Figure 3. CL
allows the sharing of the labeling effort between human and

IEEE SIGNAL PROCESSING MAGAZINE

July 2017

119

Table of Contents for the Digital Edition of Signal Processing - July 2017