Integrating Prosodic and Lexical Cues for 
Automatic Topic Segmentation 
G6khan Ttir* 
Bilkent University 
Andreas Stolcke t 
SRI International 
Dilek Hakkani-Ttir* 
Bilkent University 
Elizabeth Shriberg t 
SRI International 
We present a probabilistic model that uses both prosodic and lexical cues for the automatic seg- 
mentation of speech into topically coherent units. We propose two methods for combining lexical 
and prosodic information using hidden Markov models and decision trees. Lexical information is 
obtained from a speech recognizer, and prosodic features are extracted automatically from speech 
waveforms. We evaluate our approach on the Broadcast News corpus, using the DARPA-TDT 
evaluation metrics. Results show that the prosodic model alone is competitive with word-based 
segmentation methods. Furthermore, we achieve a significant reduction in error by combining 
the prosodic and word-based knowledge sources. 
1. Introduction 
Topic segmentation is the task of automatically dividing a stream of text or speech into 
topically homogeneous blocks. That is, given a sequence of (written or spoken) words, 
the aim of topic segmentation is to find the boundaries where topics change. Figure 1 
gives an example of a topic change boundary from a broadcast news transcript. Topic 
segmentation is an important task for various language understanding applications, 
such as information extraction and retrieval, and text summarization. In this paper, 
we present our work on automatic detection of topic boundaries from speech input 
using both prosodic and lexical information. 
Other automatic topic segmentation systems have focused on written text and 
have depended mostly on lexical information. This is problematic when segmenting 
speech. First, relying on word identities can propagate automatic speech recognizer 
errors to the topic segmenter. Second, speech lacks typographic cues, as shown in 
Figure 1: there are no headers, paragraphs, sentence punctuation marks, or capitalized 
letters. Speech itself, on the other hand, provides an additional, nonlexical knowledge 
source through its durational, intonational, and energy characteristics, i.e., its prosody. 
Prosodic cues are known to be relevant to discourse structure in spontaneous 
speech (cf. Section 2.3) and can therefore be expected to play a role in indicating topic 
transitions. Furthermore, prosodic cues, by their nature, are relatively unaffected by 
word identity, and should therefore improve the robustness of lexical topic segmenta- 
tion methods based on automatic speech recognition. 
• Department of Computer Engineering, Bilkent University, Ankara, 06533, Turkey. E-mail: {tur, 
hakkani}@cs.bilkent.edu.tr. The research reported here was carried out while the authors were 
International Fellows at SRI International. t Speech Technology and Research Laboratory, SRI International, 333 Ravenswood Ave., Menlo Park, CA 
94025. E-maih {stolcke,ees}@speech.sri.com. 
Computational Linguistics Volume 27, Number 1 
• .. tens of thousands of people are homeless in northern china tonight after a powerful 
earthquake hit an earthquake registering six point two on the richter scale at least forty 
seven people are dead few pictures available from the region but we do know tem- 
peratures there will be very cold tonight minus seven degrees <TOPIC_CHANGE> 
peace talks expected to resume on monday in belfast northern ireland former u. s. sen- 
ator george mitchell is representing u. s. interests in the talks but it is another american 
center senator rather who was the focus of attention in northern ireland today here's 
a. b. c.'s richard gizbert the senator from america's best known irish catholic family 
is in northern ireland today to talk about peace and reconciliation a peace process 
does not mean asking unionists or nationalists to change or discard their identity or 
aspirations... 
Figure 1 
An example of a topic boundary in a broadcast news transcript. 
Topic segmentation research based on prosodic information has generally relied 
on hand-coded cues (with the notable exception of Hirschberg and Nakatani \[1998\]), 
or has not combined prosodic information with lexical cues (Litman and Passon- 
neau \[1995\] is one example where lexical information was combined with hand-coded 
prosodic features for a related task). Therefore, the work presented here is the first 
that combines automatic extraction of both lexical and prosodic information for topic 
segmentation. 
The general framework for combining lexical and prosodic cues for tagging speech 
with various kinds of "hidden" structural information is a further development of 
our earlier work on sentence segmentation and disfluency detection for spontaneous 
speech (Shriberg, Bates, and Stolcke 1997; Stolcke and Shriberg 1996; Stolcke et al. 
1998), conversational dialogue tagging (Stolcke et al. 2000), and information extraction 
from broadcast news (Hakkani-T~ir et al. 1999). 
In the next section, we review previous work on topic segmentation. In Section 3, 
we describe our prosodic and language models as well as methods for combining 
them. Section 4 reports our experimental procedures and results. We close with some 
general discussion (Section 5) and conclusions (Section 6). 
2. Previous Work 
Work on topic segmentation is generally based on two broad classes of cues. On the 
one hand, one can exploit the fact that topics are correlated with topical content-word 
usage, and that global shifts in word usage are indicative of changes in topic. Quite 
independently, discourse cues, or linguistic devices such as discourse markers, cue 
phrases, syntactic constructions, and prosodic signals are employed by speakers (or 
writers) as generic indicators of endings or beginnings of topical segments. Interest- 
ingly, most previous work has explored either one or the other type of cue, but only 
rarely both. In automatic segmentation systems, word usage cues are often captured 
by statistical language modeling and information retrieval techniques. Discourse cues, 
on the other hand, are typically modeled with rule-based approaches or classifiers 
derived by machine learning techniques (such as decision trees). 
2.1 Approaches Based on Word Usage 
Most automatic topic segmentation work based on text sources has explored topical 
word usage cues in one form or other. Kozima (1993) used mutual similarity of words 
in a sequence of text as an indicator of text structure. Reynar (1994) presented a method 
that finds topically similar regions in the text by graphically modeling the distribution 
32 
Ttir, Hakkani-Tiir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
of word repetitions. The method of Hearst (1994, 1997) uses cosine similarity in a word 
vector space as an indicator of topic similarity. 
More recently, the U.S. Defense Advanced Research Projects Agency (DARPA) 
initiated the Topic Detection and Tracking (TDT) program to further the state of the 
art in finding and following new topics in a stream of broadcast news stories. One 
of the tasks in the TDT effort is segmenting a news stream into individual stories. 
Several of the participating systems rely essentially on word usage: Yamron et al. 
(1998) model topics with unigram language models and their sequential structure with 
hidden Markov models (HMMs). Ponte and Croft (1997) extract related word sets for 
topic segments with the information retrieval technique of local context analysis, and 
then compare the expanded word sets. 
2.2 Approaches Based on Discourse and Combined Cues 
Previous work on both text and speech has found that cue phrases or discourse parti- 
cles (items such as now or by the way), as well as other lexical cues, can provide valuable 
indicators of structural units in discourse (Grosz and Sidner 1986; Passonneau and Lit- 
man 1997, among others). 
In the TDT framework, the UMass HMM approach described in Allan et al. (1998) 
uses an HMM that models the initial, middle, and final sentences of a topic segment, 
capitalizing on discourse cue words that indicate beginnings and ends of segments. 
Aligning the HMM to the data amounts to segmenting it. 
Beeferman, Berger, and Lafferty (1999) combined a large set of automatically se- 
lected lexical discourse cues in a maximum entropy model. They also incorporated 
topical word usage into the model by building two statistical language models: one 
static (topic independent) and one that adapts its word predictions based on past 
words. They showed that the log likelihood ratio of the two predictors behaves as an 
indicator of topic boundaries, and can thus be used as an additional feature in the 
exponential model classifier. 
2.3 Approaches Using Prosodic Cues 
Prosodic cues form a subset of discourse cues in speech, reflecting systematic dura- 
tion, pitch, and energy patterns at topic changes and related locations of interest. A 
large literature in linguistics and related fields has shown that topic boundaries (as 
well as similar entities such as paragraph boundaries in read speech, or discourse- 
level boundaries in spontaneous speech) are indicated prosodically in a manner that 
is similar to sentence or utterance boundaries--only stronger. Major shifts in topic 
typically show longer pauses, an extra-high F0 onset or "reset," a higher maximum 
accent peak, greater range in F0 and intensity (Brown, Currie, and Kenworthy 1980; 
Grosz and Hirschberg 1992; Nakajima and Allen 1993; Geluykens and Swerts 1993; 
Ayers 1994; Hirschberg and Nakatani 1996; Nakajima and Tsukada 1997; Swerts 1997) 
and shifts in speaking rate (Brubaker 1972; Koopmans-van geinum and van Donzel 
1996; Hirschberg and Nakatani 1996). Such cues are known to be salient for human 
listeners; in fact, subjects can perceive major discourse boundaries even if the speech 
itself is made unintelligible via spectral filtering (Swerts, Geluykens, and Terken 1992). 
Work in automatic extraction and computational modeling of these characteristics 
has been more limited, with most of the work in computational prosody modeling 
dealing with boundaries at the sentence level or below. However, there have been 
some studies of discourse-level boundaries in a computational framework. They differ 
in various ways, such as type of data (monologue or dialogue, human-human or 
human-computer), type of features (prosodic and lexical versus prosodic only), which 
features are considered available (e.g., utterance boundaries or no boundaries), to 
33 
Computational Linguistics Volume 27, Number 1 
what extent features are automatically extractable and normalizable, and the machine 
learning approach used. Because of these vast difference, the overall results cannot be 
compared directly to each other or to our work, but we describe three of the approaches 
briefly here. 
An early study by Litman and Passonneau (1995) used hand-labeled prosodic 
boundaries and lexical information, but applied machine learning to a training corpus 
and tested on unseen data. The researchers combined pause, duration, and hand-coded 
intonational boundary information with lexical information from cue phrases (such as 
and and so). Additional knowledge sources included complex relations, such as coref- 
erence of noun phrases. Work by Swerts and Ostendorf (1997) used prosodic features 
that in principle could be extracted automatically, such as pitch range, to classify ut- 
terances from human-computer task-oriented dialogue into two categories: initial or 
noninitial in the discourse segment. The approach used CART-style decision trees to 
model the prosodic features, as well as various lexical features that, in principle, could 
also be estimated automatically. In this case, utterances were presegmented, so the task 
was to classify segments rather than find boundaries in continuous speech; some of the 
features included, such as type of boundary tone, may not be easy to extract robustly 
across speaking styles. Finally, Hirschberg and Nakatani (1998) proposed a prosody- 
only front end for tasks such as audio browsing and playback, which could segment 
continuous audio input into meaningful information units. They used automatically 
extracted pitch, energy, and "other" features (such as the cross-correlation value used 
by the pitch tracker in determining the estimate of F0) as inputs to CART-style trees, 
and aimed to predict major discourse-level boundaries. They found various effects of 
frame window length and speakers, but concluded overall that prosodic cues could 
be useful for audio browsing applications. 
3. The Approach 
Topic segmentation in the paradigm used in this study and others (Allan et al. 1998) 
proceeds in two phases. In the first phase, the input is divided into contiguous strings 
of words assumed to belong to the same topic. We refer to this step as chopping. For ex- 
ample, in textual input, the natural units for chopping are sentences (as can be inferred 
from punctuation and capitalization), since we can assume that topics do not change in 
mid sentence. 1 For continuous speech input, the choice of chopping criteria is less obvi- 
ous; we compare several possibilities in our experimental evaluation. Here, for simplic- 
ity, we will use "sentence" to refer to units of chopping, regardless of the criterion used. 
In the second phase, the sentences are further grouped into contiguous stretches 
belonging to one topic, i.e., the sentence boundaries are classified into topic bound- 
aries and nontopic boundaries. 2 Topic segmentation is thus reduced to a boundary 
classification problem. We will use B to denote the string of binary boundary classi- 
fications. Furthermore, our two knowledge sources are the (chopped) word sequence 
W and the stream of prosodic features F. Our approach aims to find the segmentation 
B with highest probability given the information in W and F 
argmax P( BI W, F ) (1) 
B 
using statistical modeling techniques. 
1 Similarly, it is sometimes assumed for topic segmentation purposes that topics change only at 
paragraph boundaries (Hearst 1997). 
2 We do not consider the problem of detecting recurring, discontinuous instances of the same topic, a 
task known as topic tracking in the TDT paradigm (Doddington 1998). 
34 
Ttir, Hakkani-Ttir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
In the following subsections, we first describe the prosodic model of the depen- 
dency between prosody F and topic segmentation B; then, the language model relating 
words W and B; and finally, two approaches for combining the models. 
3.1 Prosodic Modeling 
The job of the prosodic model is to estimate the posterior probability (or, alternatively, 
likelihood) of a topic change at a given word boundary, based on prosodic features ex- 
tracted from the data. For the prosodic model to be effective, one must devise suitable, 
automatically extractable features. Feature values extracted from a corpus can then be 
used in training probability estimators and to select a parsimonious subset of features 
for modeling purposes. We discuss each of these steps in turn in the following sections. 
3.1.1 Features. We started with a large collection of features capturing two major 
aspects of speech prosody, similar to our previous work (Shriberg, Bates, and Stolcke 
1997): 
Duration features: duration of pauses, duration of final vowels and final 
rhymes, and versions of these features normalized both for phone 
durations and speaker statistics. 3 
Pitch features: fundamental frequency (F0) patterns preceding and 
following the boundary, F0 patterns across the boundary, and pitch 
range relative to the speaker's baseline. We processed the raw F0 
estimates (obtained with ESPS signal processing software from Entropic 
Research Laboratory \[1993\]), with robustness-enhancing techniques 
developed by S6nmez et al. (1998). 
We did not use amplitude- or energy-based features since exploratory work showed 
these to be much less reliable than duration and pitch and largely redundant given the 
above features. One reason for omitting energy features is that, unlike duration and 
pitch, energy-related measurements vary with channel characteristics. Since channel 
properties vary widely in broadcast news, features based on energy measures can 
correlate with shows, speakers, and so forth, rather than with the structural locations 
in which we were interested. 
We included features that, based on the descriptive literature, should reflect breaks 
in the temporal and intonational contour. We developed versions of such features that 
could be defined at each interword boundary, and that could be extracted by com- 
pletely automatic means (no human labeling). Furthermore, the features were designed 
to be as independent of word identities as possible, for robustness to imperfect recog- 
nizer output. A brief characterization of the informative features for the segmentation 
task is given with our results in Section 4.6. Since the focus here is on computational 
modeling we refer the reader to a companion paper (Shriberg et al. 2000) for a detailed 
description of the acoustic processing and prosodic feature extraction. 
3.1.2 Decision Trees. Any of a number of probabilistic classifiers (such as neural net- 
works, exponential models, or naive Bayes networks) could be used as posterior prob- 
ability estimators. As in past prosodic modeling work (Shriberg, Bates, and Stolcke 
1997), we chose CART-style decision trees (Breiman et al. 1984), as implemented by 
3 The rhyme is the part of a syllable that comprises the nuclear phone (typically a vowel) and any following phones. This is the part of the syllable most typically affected by lengthening. 
35 
Computational Linguistics Volume 27, Number 1 
the IND package (Buntine and Caruana 1992), because of their ability to model feature 
interactions, to deal with missing features, and to handle large amounts of training 
data. The foremost reason for our preference for decision trees, however, is that the 
learned models can be inspected and diagnosed by human investigators. This ability 
is crucial for understanding what features are used and how, and for debugging the 
feature extraction process itself. 4 
Let Fi be the features extracted from a window around the ith potential topic 
boundary (chopping boundary), and let Bi be the boundary type (boundary/no-bound- 
ary) at that position. We trained decision trees to predict the ith boundary type, i.e., to 
estimate P(\]3ilFi, W). The decision is only weakly conditioned on the word sequence 
W, insofar as some of the prosodic features depend on the phonetic alignment of the 
word models (which we will denote with Wt). We can thus expect the prosodic model 
estimates to be robust to recognition errors. The decision tree paradigm also allows us 
to add, and automatically select, other (nonprosodic) features that might be relevant 
to the task. 
3.1.3 Feature Selection. The greedy nature of the decision tree learning algorithm 
implies that larger initial feature sets can give worse results than smaller subsets. Fur- 
thermore, it is desirable to remove redundant features for computational efficiency 
and to simplify the interpretation of results. For this purpose we developed an itera- 
rive feature selection "wrapper" algorithm (John, Kohavi, and Pfleger 1994) that finds 
useful, task-specific feature subsets. The algorithm combines elements of a brute-force 
search with previously determined heuristics about good groupings of features. The 
algorithm proceeds in two phases: In the first phase, the number of features is reduced 
by leaving out one feature at a time during tree construction. A feature whose removal 
increases performance is marked as to be avoided. The second phase then starts with 
the reduced feature set and performs a beam search over all possible subsets to max- 
imize tree performance. 
We used entropy reduction in the overall tree (after cross-validation pruning) as a 
metric for comparing alternative feature subsets. Entropy reduction is the difference in 
entropy between the prior class distribution and the posterior distribution estimated 
by the tree, as measured on a held-out set; it is a more fine-grained metric than 
classification accuracy, and is also more relevant to the model combination approach 
described later. 
3.1.4 Training Data. To train the prosodic model, we automatically aligned and ex- 
tracted features from 70 hours (about 700,000 words) of the Linguistic Data Consortium 
(LDC) 1997 Broadcast News (BN) corpus. Topic boundary information determined by 
human labelers was extracted from the SGML markup that accompanies the word 
transcripts of this corpus. The word transcripts were aligned automatically with the 
acoustic waveforms to obtain pause and duration information, using the SRI Broadcast 
News recognizer (Sankar et al. 1998). 
3.2 Lexical Modeling 
Lexical information in our topic segmenter is captured by statistical language models 
(LMs) embedded in an HMM. The approach is an extension of the topic segmenter 
4 Interpreting large trees can be a daunting task. However, the decision questions near the tree root are 
usually interpretable, or, when nonsensical, usually indicate problems with the data. Furthermore, as 
explained in Section 4.6, we have developed simple statistics that give an overview of feature usage throughout the tree. 
36 
Tfir, Hakkani-Tfir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
Figure 2 
Structure of the basic HMM developed by Dragon for the TDT Pilot Project. The labels on the 
arrows indicate the transition probabilities. TSP represents the topic switch penalty. 
developed by Dragon Systems for the TDT2 effort (Yamron et al. 1998), which was 
based purely on topical word distributions. We extend it to also capture lexical and 
(as described in Section 3.3) prosodic discourse cues. 
3.2.1 Model Structure. The overall structure of the model is that of an HMM (Rabiner 
and Juang 1986) in which the states correspond to topic clusters Tj, and the obser- 
vations are sentences (or chopped units) W1 ..... WN. The resulting HMM, depicted 
in Figure 2, forms a complete graph, allowing for transitions between any two topic 
clusters. Note that it is not necessary that the topic clusters correspond exactly to the 
actual topics to be located; for segmentation purposes, it is sufficient that two adjacent 
actual topics are unlikely to be mapped to the same induced cluster. The observation 
likelihoods for the HMM states, P(WilTj), represent the probability of generating a 
given sentence Wi in a particular topic cluster Tj. 
We automatically constructed 100 topic cluster LMs, using the multipass k-means 
algorithm described in Yamron et al. (1998). Since the HMM emissions are meant to 
model the topical usage of words, but not topic-specific syntactic structures, the LMs 
37 
Computational Linguistics Volume 27, Number 1 
consist of unigram distributions that exclude stopwords (high-frequency function and 
closed-class words). To account for unobserved words, we interpolate the topic-cluster- 
specific LMs with the global unigram LM obtained from the entire training data. The 
observation likelihoods of the HMM states are then computed from these smoothed 
unigram LMs. 
All HMM transitions within the same topic cluster are given probability one, 
whereas all transitions between topics are set to a global topic switch penalty (TSP) 
that is optimized on held-out training data. The TSP parameter allows trading off 
between false alarms and misses. Once the HMM is trained, we use the Viterbi al- 
gorithm (Viterbi 1967; Rabiner and Juang 1986) to search for the best state sequence 
and corresponding segmentation. Note that the transition probabilities in the model 
are not normalized to sum to one; this is convenient and permissible since the out- 
put of the Viterbi algorithm depends only on the relative weight of the transition 
weights. 
We augmented the Dragon segmenter with additional states and transitions to 
also capture lexical discourse cues. In particular, we wanted to model the initial and 
final sentences in each topic segment, as these often contain formulaic phrases and 
keywords used by broadcast speakers (From Washington, this is .... And now ... ). We 
added two additional states, BEGIN and END, to the HMM (Figure 3) to model these 
sentences. Likelihoods for the BEGIN and END states are obtained as the unigram 
language model probabilities of the initial and final sentences, respectively, of the 
topic segments in the training data. Note that a single BEGIN and END state are 
shared for all topics. Best results were obtained by making traversal of these states 
optional in the HMM topology, presumably because some initial and final sentences 
are better modeled by the topic-specific LMs. 
The resulting model thus effectively combines the Dragon and UMass HMM topic 
segmentation approaches described in Allan et al. (1998). In preliminary experiments, 
we observed a 5% relative reduction in segmentation error with initial and final states 
over the baseline HMM topology of Figure 2. Therefore, all results reported later use an 
HMM topology with initial and final states. Note that, since the topic-initial and topic- 
final states are optional, our training of the model is suboptimal. Instead of labeling all 
topic-initial and topic-final training sentences as data for the corresponding state, we 
would expect further improvements by training the HMM in unsupervised fashion 
using the Baum-Welch algorithm (Baum et al. 1970; Rabiner and Juang 1986). 
3.2.2 Training Data. Topic unigram language models were trained from the pooled 
TDT Pilot and TDT2 training data (Cieri et al. 1999), covering transcriptions of broad- 
cast news from January 1992 through June 1994 and from January 1998 through Febru- 
ary 1998, respectively. These corpora are similar in style, but do not overlap with the 
1997 LDC BN corpus from which we selected our prosodic training data and the eval- 
uaton test set. For training the language models, we removed stories with fewer than 
300 and more than 3,000 words, leaving 19,916 stories with an average length of 538 
words (including stopwords). 
3.3 Model Combination 
We are now in a position to describe how lexical and prosodic information can be 
combined for topic segmentation. As discussed before, the LMs in the HMM capture 
topical word usage as well as lexical discourse cues at topic transitions, whereas a 
decision tree models prosodic discourse cues. We expect that these knowledge sources 
are largely independent, so their combination should yield significantly improved 
performance. 
38 
T~r, Hakkani-Ttir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
TSI 
TSP 
1 
• 
TSP 
TSP 
TSP 
TSP 
1 
TSP 
TSP 
Figure 3 
Structure of an HMM with topic BEGIN and END states. TSP represents the topic switch 
penalty. 
Below we present two approaches for building a combined statistical model that 
performs topic segmentation using all available knowledge sources. For both ap- 
proaches it is convenient to associate a "boundary" pseudotoken with each potential 
topic boundary (i.e., with each sentence boundary). Correspondingly, we introduce into 
the HMM new states that emit these boundary tokens. No other states emit boundary 
tokens; therefore each sentence boundary must align with one of the boundary states 
in the HMM. As shown in Figure 4, there are two boundary states for each topic 
cluster, one representing a topic transition and the other representing a topic-internal 
transition between sentences. Unless otherwise noted, the observation likelihoods for 
the boundary states are set to unity. 
The addition of boundary states allows us to compute the model's prediction of 
topic changes as follows: Let B1,.. •, Bc denote the topic boundary states and, similarly, 
let N1,..., Nc denote the nontopic boundary states, where C is the number of topic 
clusters. Using the forward-backward algorithm for HMMs (Rabiner and Juang 1986), 
we can compute P(qi = BflW) and P(qi = NjlW), the posterior probabilities that one 
of these states is occupied at boundary i. The model's prediction of a topic boundary 
39 
Computational Linguistics Volume 27, Number i 
TSP 
TSP 
TSP 
Figure 4 Structure of the final HMM with fictitious boundary states used for combining language and 
prosodic models. In the figure, states B1, B2, ..., B100 represent the presence of a topic 
boundary, whereas states N1, N2,..., N100 represent topic-internal sentence boundaries. TSP 
is the topic switch penalty. 
is simply the sum over the corresponding state posteriors: 
c 
PHMM(Bi ~- yes\]W) = ~P(qi = BjlW) (2) 
j=l 
c 
PHMM(Bi = nolW ) = ~_~P(qi = NjlW) 
j=l 
= 1 - PHMM(Bi = yes\[W) (3) 
40 
Tiir, Hakkani-T(ir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
3.3.1 Model Combination in the Decision Tree. Decision trees allow the training of 
a single classifier that takes both lexical and prosodic features as input, provided we 
can compactly encode the lexical information for the decision tree. We compute the 
posterior probability PHMM(Bi = yeslW) as shown above, to summarize the HMM's 
belief in a topic boundary based on all available lexical information W. The posterior 
value is then used as an additional input feature to the prosodic decision tree, which is 
trained in the usual manner. During testing, we declare a topic boundary whenever the 
tree's overall posterior estimate PDT(BilFi, W) exceeds some threshold. The threshold 
may be varied to trade off false alarms for miss errors, or to optimize an overall cost 
function. 
Using HMM posteriors as decision tree features is similar in spirit to the knowl- 
edge source combination approaches used by Beeferman, Berger, and Lafferty (1999) 
and Reynar (1999), who also used the output of a topical word usage model as in- 
put to an overall classifier. In previous work (Stolcke et al. 1998) we used the present 
approach as one of the knowledge source combination strategies for sentence and 
disfluency detection in spontaneous speech. 
3.3.2 Model Combination in the HMM. An alternative approach to knowledge source 
combination uses the HMM as the top-level model. In this approach, the prosodic 
decision tree is used to estimate likelihoods for the boundary states of the HMM, thus 
integrating the prosodic evidence into the HMM's segmentation decisions. 
More formally, let Q =- (rl, ql ..... ri, qi,...,rN, qN) be a state sequence through 
the HMM. The model is constructed such that the states ri representing topic (or 
BEGIN/END) clusters alternate with the states qi representing boundary decisions. 
As in the baseline model, the likelihoods of the topic cluster states Tj account for the 
lexical observations: 
P(Wi\]ri ~- Tj) = P(WilTj) (4) 
as estimated by the unigram LMs. Now, in addition, we let the likelihood of the 
boundary state at position i reflect the prosodic observation Fi. Recall that, like Wi, Fi 
refers to complete sentence units; specifically, Fi denotes the prosodic features of the 
ith boundary between such units. 
P(Fi\]qi = Bj, W) = P(Fi\]Bi = yes, W) 
P(Fi\[qi = Nj, W) = P(FiIBi = no, W) j for all j = 1 .... , C (5) 
Using this construction, the product of all state likelihoods will give the overall like- 
lihood, accounting for both lexical and prosodic observations: 
N N 
1-\[ P(Wilri) II P(Filqi, W) = P(W, FIQ ) (6) 
i=1 i=1 
Applying the Viterbi algorithm to the HMM will thus return the most likely segmen- 
tation conditioned on both words and prosody, which is our goal. 
Although decomposing the likelihoods as shown allows prosodic observations to 
be conditioned on the words W, we use only the phonetic alignment information Wt 
from the word sequence W in our prosodic models, ignoring the word identities, so 
as to make them more robust to recognition errors. 
The likelihoods P(FilBi, Wt) for the boundary states can now be obtained from the 
prosodic decision tree. Note that the decision tree estimates posteriors PDT(Bil\]2i, Wt). 
41 
Computational Linguistics Volume 27, Number 1 
These can be converted to likelihoods using Bayes rule as in 
P(Fi\[Bi, Wt) = P(FilWt)PDT (BilEG Wt) 
P(BiIWt) (7) 
The term P(FilWt) is a constant for all decisions Bi and can thus be ignored when 
applying the Viterbi algorithm. Next, we approximate P(BilWt) ,~ P(Bi), justified by the 
fact that the Wt contains information about start and end times of phones and words, 
but not directly about word identities. Instead of explicitly dividing the posteriors, 
1 we prefer to downsample the training set to make P(Bi = yes) = P(Bi = no) = ~. 
A beneficial side effect of this approach is that the decision tree models the lower- 
frequency events (topic boundaries) in greater detail than if presented with the raw, 
highly skewed class distribution. 
As is often the case when combining probabilistic models of different types, it is 
advantageous to weight the contributions of the language models and the prosodic 
trees relative to each other. We do so by introducing a tunable model combination 
weight (MCW), and by using PDT(FilBi, Wt) MCW as the effective prosodic likelihoods. 
The value of MCW is optimized on held-out data. 
4. Experiments and Results 
To evaluate our topic segmentation models, we carried out experiments in the TDT 
paradigm. We first describe our test data and the evaluation metrics used to compare 
model performance, then give the results we obtained with individual knowledge 
sources, followed by the results of the combined models. 
4.1 Test Data 
We evaluated our system on three hours (6 shows, about 53,000 words) of the 1997 
LDC BN corpus. The threshold for the model combination in the decision tree and 
the topic switch penalty were optimized on the larger development training set of 
104 shows, which includes the prosodic model training data. The MCW for the model 
combination in the HMM was optimized using a smaller held-out set of 10 shows of 
about 85,000 words total size, separate from the prosodic model training data. 
We used two test conditions: forced alignments using the true words, and recog- 
nized words as obtained by a simplified version of the SRI Broadcast News recognizer 
(Sankar et al. 1998), with a word error rate of 30.5%. 
Our aim in these experiments was to use fully automatic recognition and pro- 
cessing wherever possible. For practical reasons, we departed from this strategy in 
two areas. First, for word recognition, we used the acoustic waveform segmentations 
provided with the corpus (which also included the location of non_news material, such 
as commercials and music). Since current BN recognition systems perform this seg- 
mentation automatically with very good accuracy and with only a few percentage 
points penalty in word error rate (Sankar et al. 1998), we felt the added complication 
in experimental setup and evaluation was not justified. 
Second, for prosodic modeling, we used information from the corpus markup 
concerning speaker changes and the identity of frequent speakers (e.g., news anchors). 
Automatic speaker segmentation and labeling is possible, although not without errors 
(Przybocki and Martin 1999). Our use of speaker labels was motivated by the fact 
that meaningful prosodic features may require careful normalization by speaker, and 
unreliable speaker information would have made the analysis of prosodic feature 
usage much less meaningful. 
42 
Tfir, Hakkani-Ttir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
4.2 Evaluation Metrics 
We have adopted the evaluation paradigm used by the TDT2--Topic Detection and 
Tracking Phase 2 (Doddington 1998) program, allowing fair comparisons of various 
approaches both within this study and with respect to other recent work. Segmentation 
accuracy was measured using TDT evaluation software from NIST, which implements 
a variant of an evaluation metric suggested by Beeferman, Berger, and Lafferty (1999). 
The TDT segmentation metric is different from those used in most previous topic 
segmentation work, and therefore merits some discussion. It is designed to work on 
data streams without any potential topic boundaries, such as paragraph or sentence 
boundaries, being given a priori. It also gives proper partial credit to segmentation 
decisions that are close to actual boundaries; for example, placing a boundary one 
word from an actual boundary is considered a lesser error than if the hypothesized 
boundary is off by, say, 100 words. 
The evaluation metric reflects the probability that two positions in the corpus 
probed at random and separated by a distance of k words are correctly classified as 
belonging to the same story or not. If the two words belong to the same topic segment, 
but are erroneously claimed to be in different topic segments by the segmenter, then 
this will increase the system's false alarm probability. Conversely, if the two words are 
in different topic segments, but are erroneously marked to be in the same segment, 
this will contribute to the miss probability. The false alarm and miss rates are defined 
as averages over all possible probe positions with distance k. 
Formally, miss and false alarm rates are computed as 5 
PMis, = Y~s ~N--~lk d~yp (i, i + k) x (1 - dS~¢f (i, i + k)) (8) 
E E Ns-kl'' (i,i+k)) s i=1 ,l-d~¢i 
v"N~-k (1 • s • • PFalseAla~m -- Y'~s A..M=I \ -- dShyp( l" i + k)) x d~ef(Z, ~ + k) (9) 
v-,N~-k as (i i + k) Es Z.~i=l ~ref \ " 
where the summation is over all broadcast shows s and word positions i in the test 
corpus and where d  sli,,={i 
if words i and j in show s are deemed by sys to 
be within the same story 
otherwise 
Here sys can be ref to denote the reference (correct) segmentation, or hyp to denote the 
segmenter's decision. 
An analogous metric is defined for audio sources, where segmentation decisions 
(same or different topic) are probed at a time-based distance A: 
-T~ -A -s - PMiss = GsJt=o ahyp(t,t+A) x (1-d~¢f(t,t+A))dt (10) 
~' ~G-A(1 s Jr=0 -dScd(t,t+A))dt 
f -a(. PFa,,¢A,a .... = Gsat=O ,l -d~yp(t,t+A)) x d~¢/(t,t+A)dt (11) 
-G-A -s - ~s J,=o a~z(t,t+n)dt 
5 The definitions are those from Doddington (1998), but have been simplified and edited for clarity. 
43 
Computational Linguistics Volume 27, Number 1 
Table 1 
Segmentation error rates for various chopping criteria, using true words of the larger 
development data set. 
Chopping Criterion P Miss P FalseA larm C Seg 
FIXED 0.5688 0.0639 0.2153 
TURN 0.6737 0.0436 0.2326 
SENTENCE 0.5469 0.0557 0.2030 
PAUSE 0.5111 0.0688 0.2002 
where the integration is over the entire duration of all stories of the shows in the test 
corpus, and where 
if times tl and t2 in show s are deemed by sys to 
be within the same story 
otherwise 
We used the same parameters as used in the official TDT2 evaluation: k = 50 
and A = 15 seconds. Furthermore, again following NIST's evaluation procedure, we 
combine miss and false alarm rates into a single segmentation cost metric 
Cseg : CMiss X PMiss X P~eg + CFalseAlarm X PFalseAlarm × (1 - P~¢9) (12) 
where the CMis~ = 1 is the cost of a miss, CFalseAlarm : 1 is the cost of a false alarm, 
and Pseg = 0.3 is the a priori probability of a segment being within an interval of k 
words or A seconds on the TDT2 training corpus. 6 
4.3 Chopping 
Unlike written text, the output of the automatic speech recognizer contains no sentence 
boundaries. Therefore, chopping text into (pseudo)sentences is a nontrivial problem 
when processing speech. Some presegmentation into roughly sentence-length units is 
necessary since otherwise the observations associated with HMM states would com- 
prise too few words to give robust likelihoods of topic choice, causing poor perfor- 
mance. 
We investigated chopping criteria based on a fixed number of words (FIXED), at 
speaker changes (TURN), at pauses (PAUSE), and, for reference, at actual sentence 
boundaries (SENTENCE) obtained from the transcripts. Table 1 gives the error rates 
for the four conditions, using the true word transcripts of the larger development 
data set. For the PAUSE condition, we empirically determined an optimal minimum 
pause duration threshold to use. Specifically, we considered pauses exceeding 0.575 
of a second as potential topic boundaries in this (and all later) experiments. For the 
FIXED condition, a block length of 10 words was found to work best. 
We conclude that a simple prosodic feature, pause duration, is an excellent criterion 
for the chopping step, giving comparable or better performance than standard sentence 
boundaries. Therefore, we used pause duration as the chopping criterion in all further 
experiments. 
6 Another parameter in the NIST evaluation is the deferral period, i.e., the amount of look-ahead before 
a segmentation decision is made. In all our experiments, we allowed unlimited deferral, effectively 
until the end of the news show being processed. 
44 
Tfir, Hakkani-Tiir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
Table 2 
Summary of error rates with the language model only (LM), the prosody model only (PM), the 
combined decision tree (CM-DT), and the combined HMM (CM-HMM). (a) shows word-based 
error metrics, 00) shows time-based error metrics. In both cases a "chance" classifier that labels 
all potential boundaries as nontopic would achieve 0.3 weighted segmentation cost. 
(a) Error Rates on Forced Alignments Error Rates on Recognized Words 
Model PMiss PFalseAlarm Cseg PMiss PFalseAlarm Cseg 
Chance 1.0 0.0 0.3 1.0 0.0 0.3 
LM 0.4847 0.0630 0.1895 0.4978 0.0577 0.1897 
PM 0.4130 0.0596 0.1657 0.4125 0.0705 0.1731 
CM-DT 0.4677 0.0260 0.1585 0.4891 0.0146 0.1569 
CM-HMM 0.3339 0.0536 0.1377 0.3748 0.0450 0.1438 
(b) Error Rates on Forced Alignments Error Rates on Recognized Words 
Model PMiss PFalseAlarm Cseg PMiss PFalseAlarm Cseg 
Chance 1.0 0.0 0.3 1.0 0.0 0.3 
LM 0.5260 0.0490 0.1921 0.5361 0.0415 0.1899 
PM 0.3503 0.0892 0.1675 0.3846 0.0737 0.1669 
CM-DT 0.5136 0.0210 0.1688 0.5426 0.0125 0.1715 
CM-HMM 0.3426 0.0496 0.1375 0.3746 0.0475 0.1456 
4.4 Source-Specific Model Tuning 
As mentioned earlier, the segmentation models contain global parameters (the topic 
transition penalty of the HMM and the posterior threshold for the combined decision 
tree) to trade false alarms for miss errors. Optimal settings for these parameters depend 
on characteristics of the source, in particular on the relative frequency of topic changes. 
Since broadcast news programs come from identified sources, it is useful and legitimate 
to optimize these parameters for each show type. 7 We therefore optimized the global 
parameter for each model to minimize the segmentation cost on the training corpus 
(after training all other model parameters in a source-independent fashion). 
Compared to a baseline using source-independent global TSP and threshold, the 
source-dependent models showed between 5% and 10% relative error reduction. All 
results reported below use the source-dependent approach. 
4.5 Segmentation Results 
Table 2 shows the results for both individual knowledge sources (words and prosody), 
as well as for the combined models (decision tree and HMM). It is worth noting 
that the prosody-only results were obtained by running the combined HMM without 
language model likelihoods; this approach gave better performance than using the 
prosodic decision trees directly as classifiers. 
Both word- and time-based metrics are given; they exhibit generally very similar 
results. Another dimension of the evaluation is the use of correct word transcripts 
(forced alignments) versus automatically recognized words. Again, results along this 
dimension are very similar, with some exceptions noted below. 
Comparing the individual knowledge sources, we observe that prosody alone does 
somewhat better than the word-based HMM alone. The types of errors made differ 
7 Shows in the 1997 BN corpus come from eight sources: ABC World News Tonight, CNN Headline 
News, CNN Early Prime, PRI The World, CNN Prime News, CNN The World Today, C-SPAN Public 
Policy, and C-SPAN Washington Journal. Six of these occurred in the test set. 
45 
Computational Linguistics Volume 27, Number 1 
consistently: the prosodic model has a higher false alarm rate, while the word-LMs 
have more miss errors. The prosodic model shows more false alarms because regular 
sentence boundaries often show characteristics similar to those of topic boundaries. It 
also suggests that both models could be combined by letting the prosodic model select 
candidate topic boundaries that would then be filtered using lexical information. 
The combined models generally improve on the individual knowledge sources, s 
In the word-based evaluation, the combined decision tree (DT) reduced overall seg- 
mentation cost by 19% over the language model on true words (17% on recognized 
words). The combined HMM gave even better results: 27% and 24% improvement in 
the error rate over the language model for true and recognized words, respectively. 
Looking again at the breakdown of errors, we can see that the two model combina- 
tion approaches work quite differently: the combined DT has about the same miss rate 
as the LM, but a lower false alarms rate. The combined HMM, by contrast, combines 
a miss rate as low as (or lower than) that of the prosodic model with the lower false 
alarm rate of the LM, suggesting that the functions of the two knowledge sources are 
complementary, as discussed above. Furthermore, the different error patterns of the 
two combination approaches suggest that further error reductions could be achieved 
by combining the two hybrid models. 9 
The trade-off between false alarms and miss probabilities is shown in more de- 
tail in Figure 5, which plots the two error metrics against each other. Note that the 
false alarm rate does not reach one because the segmenter is constrained by the chop- 
ping algorithm: the pause criterion prevents the segmenter from hypothesizing topic 
boundaries everywhere. 
4.6 Decision Tree for the Prosody-Only Model 
Feature subset selection was run with an initial set of 73 potential features, which the 
algorithm reduced to a set of 7 nonredundant features helpful for the topic segmen- 
tation task. The full decision tree learned is shown in Figure 6. We can identify four 
different kinds of features used in the tree, listed below. For each feature type, we give 
the feature names found in the tree and the relative feature usage, an approximate 
measure of feature importance (Shriberg, Bates, and Stolcke 1997). Relative feature 
usage is computed as the relative frequency with which features of a given type are 
queried in the tree, over a held-out test set. 
. 
. 
Pause duration (PhU_DUR, 42.7% usage). This feature is the duration of 
the nonspeech interval occurring at the boundary. The importance of 
pause duration is underestimated here because, as explained earlier, 
pause durations are already used during the chopping process, so that 
the decision tree is applied only to boundaries exceeding a certain 
duration. Separate experiments using boundaries below our chopping 
threshold show that the tree also distinguishes shorter pause durations 
for segmentation decisions. 
FO differences across the boundary (FOK_LRd~EAN_KBASELN and 
FOK_WRD_DIFF_MNMI~_NG, 35.9% usage). These features compare the mean 
8 The exception is the time-based evaluation of the combined decision tree. We found that the posterior 
probability threshold optimized on the training set works poorly on the test set for this model 
architecture and the time-based evaluation. The threshold that is optimal on the test set achieves 
Csea = 0.1651. Section 4.7 gives a possible explanation for this result. 9 Such a combination of combined models was suggested by one of the reviewers; we hope to pursue it 
in future research. 
46 
Ti.ir, Hakkani-Ttir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
.~..~ 
,.Q 
< 
r.,g3 
0.6 
0.4 
0.2 
0.0 
* * LM 
\[\] \[\] PM 
a ~ CM-DT 
: CM-HMM 
0.2 0.4 0.6 0.8 1.0 
Miss Probability 
Figure 5 
False alarm versus miss probabilities (word-based metrics) for automatic topic segmentation 
from known words (forced alignments). The segmenters used were a words-only HMM (LM), 
a prosody-only HMM (PM), a combined decision tree (CM-DT), and a combined HMM 
(CM-HMM). 
. 
F0 of the word preceding the boundary (measured from voiced regions 
within that word) to either the speaker's estimated baseline F0 
(FOK_LR_MEAN_KBASELN) or to the mean F0 of the word following the 
boundary (FOK_WRD_DIFF_.MNMN_N). Both features were computed based on 
a log-normal scaling of F0. Other measures (such as minimum or 
maximum F0 in the word or preceding window) as well as other 
normalizations (based on F0 toplines, or non-log-based scalings) were 
included in the initial feature set, but were not selected in the 
best-performing tree. The baseline feature captures a pitch range effect, 
and is useful at boundaries where the speaker changes (since range here 
is compared only within-speaker). The second feature captures the 
relative size of the pitch change at the boundary, but of course is not 
meaningful at speaker boundaries. 
Turn features (TURN_F and TURN_TIME, 14.6% usage). These features 
reflect the change of speakers. TURN_F indicates whether a speaker 
47 
Computational Linguistics Volume 27, Number 1 
/ 
z 
v 
~a 
z 
z z 
v 
i 
O 
48 
Tar, Hakkani-Ti~r, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
. 
change occurred at the boundary, while TURN_TIME measures the time 
passed since the start of the current turn. 
Gender (GEN, 6.8% usage). This feature indicates the speaker gender 
right before a potential boundary. 
Inspection of the tree reveals that the purely prosodic features (pause duration and 
F0 differences) are used as the prosody literature suggests. The longer the observed 
pause, the more likely a boundary corresponds to a topic change. Also, the closer a 
speaker comes to his or her F0 baseline, or the larger the difference to the F0 following 
a boundary, the more likely a topic change occurs. These features thus correspond 
to the well-known phenomena of boundary tones and pitch reset that are generally 
associated with sentence boundaries (Vaissi6re 1983). We found these indicators of 
sentences boundaries to be particularly pronounced at topic boundaries. 
While turn and gender features are not prosodic features per se, they do interact 
closely with them since prosodic measurements must be informed by and carefully 
normalized for speaker identity and gender, and it is therefore natural to include them 
in a prosodic classifier. 1° Not surprisingly, we find that turn boundaries are positively 
correlated with topic boundaries, and that topic changes become more likely the longer 
a turn has been going on. 
Interestingly, speaker gender is used by the decision tree for several reasons. One 
reason is stylistic differences between males and females in the use of F0 at topic 
boundaries. This is true even after proper normalization, e.g., equatIng the gender- 
specific nontopic boundary distributions. In addition, we found that nontopic pauses 
(i.e., chopping boundaries) are more likely to occur in male speech. It could be that 
male speakers in BN are assigned longer topic segments on average, or that male 
speakers are more prone to pausing in general, or that male speakers dominate the 
spontaneous speech portions, where pausing is naturally more frequent. The details 
of this gender effect await further study. 
4.7 Decision Tree for the Combined Model 
Figure 7 depicts the decision tree that combines the HMM language model topic deci- 
sions with prosodic features (see Section 3.3.1). Again, we list the features used with 
their relative feature usages. 
1. Language model posterior (POST_TOPIC, 49.3% usage). This is the 
posterior probability P(Bi = yeslW) computed from the HMM. 
2. Pause duration (PAU_DUR, 49.3% usage). This feature is the same as 
described for the prosody-only model. 
3. FO differences across the boundary (FOK_WRD_DIFF_HILO_N and 
FOK_LR_MEAN_KBASELN, 1.4% usage). These features are similar to those 
found for the prosody-only tree. The only difference is that for the first 
feature, the comparison of FO values across the boundary is done by 
taking the maximum FO of the previous word and the minimum FO of 
the following word, rather than the mean for both cases. 
10 For example, the features that measure F0 differences across boundaries do not make sense if the speaker changes at the boundary. Accordingly, we made such features undefined for the decision tree 
at turn boundaries. 
49 
Computational Linguistics Volume 27, Number 1 
POST_TOPIC >= -0.083984 
r PAU_DUR >= 1058.3 PAU_DUR <82.5 ~PAU_DUR >= 82.5 
KBASELN < 0.1977~R MEAN KBASELN >= 0.19777 
~_WRD_DIFF_HILO_N <-0.02498~RD DIFF HILO N >=-0.024989 
Figure 7 
The decision tree of the combination model. 
The decision tree found for the combined task is smaller and uses fewer features 
than the one trained with prosodic features only, for two reasons. First, the LM poste- 
rior feature is found to be highly informative, superseding the selection of many of the 
low-frequency features previously found. Furthermore, as explained in Section 3.3.2, 
the prosody-only tree was trained on a downsampled dataset that equalizes the priors 
for topic and nontopic boundaries, as required for integration into the HMM. A wel- 
come side effect of this procedure is that it forces the tree to model the less frequent 
class (topic boundaries) in much greater detail than if the tree were trained on the raw 
class distribution, as is the case here. 
Because of its small size, the tree in Figure 7 is particularly easy to interpret. The 
top-level split is based on the LM posterior. The right branch handles cases where 
words are highly indicative of a topic boundary. However, for short pauses, the tree 
queries further prosodic features to prevent false alarms. Specifically, short pauses 
must be accompanied both by an F0 close to the speaker's baseline and by a large 
F0 reset to be deemed topic boundaries. Conversely, if the LM posteriors are low (left 
top-level branch), but the pause is very long, the tree still outputs a topic boundary. 
4.8 Comparison of Model Combination Approaches 
Results indicate that the model combination approach using an HMM as the top-level 
model works better than the combined decision tree. While this result deserves more 
investigation, we can offer some preliminary insights. 
We found it difficult to set the posterior probability thresholds for the combined 
decision tree in a robust way. As shown by the CM-DT curve in Figure 5, there is a 
large jump in the false alarm/miss trade-off for the combined tree, in contrast to the 
combined HMM approach, which controls the trade-off by a changing topic switch 
penalty. This occurs because posterior probabilities from the decision tree do not vary 
smoothly; rather, they vary in steps corresponding to the leaves of the tree. The dis- 
50 
Tfir, Hakkani-Tiir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
Table 3 
Segmentation error rates with the language model only (LM), the combined HMM using all 
prosodic features (CM-HMM-all), the combined HMM using only pause duration and turn 
features (CM-HMM-pause-turn), and using only pause duration, turn, and gender features 
(CM-HMM-pause-turn-gender). 
Model C~ex 
LM 0.1895 
CM-HMM-pause-turn 0.1519 
CM-HMM-pause-turn-gender 0.1511 
CM-HMM-all 0.1377 
continuous character of the thresholded variable makes it hard to estimate a threshold 
on the training data that performs robustly on the test data. This could account for 
the poor result on the time-based metrics for the combined tree (where the threshold 
optimized on the training data was far from optimal on the test set; see footnote 8). 
The same phenomenon is reflected in the fact that the prosody-only tree gave better 
results when embedded in an HMM without LM likelihoods than when used by itself 
with a posterior threshold. 
4.9 Contributions of Different Feature Types 
We saw in Section 4.6 that pause duration is by far the single most important feature in 
the prosodic decision tree. Furthermore, speaker changes are queried almost as often 
as the F0-related features. Pause durations can be obtained using standard speech rec- 
ognizers, and are in fact used by many current TDT systems (see Section 4.10). Speaker 
changes are not prosodic features per se, and would be detected independently from 
the prosodic features proper. To determine if prosodic measurements beyond pause 
and speaker information improve topic segmentation accuracy, we tested systems that 
consisted of the HMM with the usual topic LMs, plus a decision tree that had ac- 
cess only to various subsets of pause- and speaker-related features, without using 
any of the F0-based features. Decision tree and HMM were combined as described in 
Section 3.3.2. 
Table 3 shows the results of the system using only topic language models (LM) 
as well as combined systems using all prosodic features (CM-HMM-all), only pause 
duration and turn features (CM-HMM-pause-turn), and using only pause duration, 
turn, and gender features (CM-HMM-pause-turn-gender). These results show that by 
using only pause duration, turn, and gender features, it is indeed possible to obtain 
better results (20% reduced segmentation cost) than with the lexical model alone, with 
gender making only a minor contribution. However, we also see that a substantial 
further improvement (9% relative) is obtained by adding F0 features to the prosodic 
model. 
4.10 Results Compared to Other Approaches 
Because our work focused on the use of prosodic information and required detailed 
linguistic annotations (such as sentence punctuation, turn boundaries, and speaker 
labels), we used data from the LDC 1997 BN corpus to form the training set for the 
prosodic models and the (separate) test set used for evaluation. This choice was crucial 
for the research, but unfortunately complicates a quantitative comparison of our results 
to other TDT segmentation systems. The recent TDT2 evaluation used a different set 
of broadcast news data that postdated the material we used, and was generated by 
a different speech recognizer (although with a similar word error rate) (Cieri et al. 
51 
Computational Linguistics Volume 27, Number 1 
Table 4 
Word-based segmentation error rates for different corpora. Note that a hand-transcribed 
(forced alignment) version of the TDT2 test set was not available. 
Error Rates on Forced Alignments Error Rates on Recognized Words 
Test Set PMis~ PFalseAlarm Cseg PMiss PFalseAlarrn Cseg 
TDT2 NA NA NA 0.5509 0.0694 0.2139 
BN'97 0.4685 0.0817 0.1978 0.5128 0.0683 0.2017 
1999). Nevertheless we have attempted to calibrate our results with respect to these 
TDT2 results, n We have not tried to compare our results to research outside the TDT 
evaluation framework. In fact, other evaluation methodologies differ too much to allow 
meaningful quantitative comparisons across publications. 
We wanted to ensure that the TDT2 evaluation test set was comparable in seg- 
mentation difficulty to our test set drawn from the 1997 BN corpus, and that the TDT2 
metrics behaved similarly on both sets. To this end, we ran an early version of our 
words-only segmenter on both test sets. As shown in Table 4, not only are the results 
on recognized words quite close, but the optimal false alarm/miss trade-off is similar 
as well, indicating that the two corpora have roughly similar topic granularities. 
While the full prosodic component of our topic segmenter was not applied to the 
TDT2 test corpus, we can compare the performance of a simplified version of SRI's 
segmenter to other evaluation systems (Fiscus et al. 1999). The two best-performing 
systems in the evaluation were those of CMU (Beeferman, Berger, and Lafferty 1999) 
with Cse9 = 0.1463, and Dragon (Yamron et al. 1998; van Mulbregt et al. 1999) with 
Cse9 = 0.1579. The SRI system achieved Cs~g = 0.1895. All systems in the evaluation, 
including ours, used only information from words and pause durations determined 
by a speech recognizer. 
A good reference to calibrate our performance is the Dragon system, from which 
we borrowed the lexical HMM segmentation framework. Dragon made adjustments 
in its lexical modeling that account for the improvements relative to the basic HMM 
structure on which our system is based. As described by van Mulbregt et al. (1999), 
a significant segmentation error reduction was obtained from optimizing the number 
of topic clusters (kept fixed at 100 in our system). Second, Dragon introduced more 
supervision into the model training by building separate LMs for segments that had 
been hand-labeled as not related to news (such as sports and commercials) in the TDT2 
training corpus, which also resulted in substantial improvements. Finally, Dragon used 
some of the TDT2 training data for tuning the model to the specifics of the TDT2 
corpus. 
In summary, the performance of our combined lexical-prosodic system with Cs¢9 = 
0.1438 is competitive with the best word-based systems reported to date. More impor- 
tantly, since we found the prosodic and lexical knowledge sources to complement each 
other, and since Dragon's improvements for TDT2 were confined to a better modeling 
of the lexical information, we would expect that adding these improvements to our 
combined segmenter would lead to a significant improvement in the state of the art. 
11 Since our study was conducted, a third round of TDT benchmarks (TDT3) has taken place (NIST 1999). 
However, for TDT3, the topic segmentation evaluation metric was modified and the most recent results 
are thus not directly comparable with those from TDT2 or the present study, 
52 
Tiir, Hakkani-Ti.ir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
5. Discussion 
Results so far indicate that prosodic information provides an excellent source of in- 
formation for automatic topic segmentation, both by itself and in conjunction with 
lexical information. Pause duration, a simple prosodic feature that is readily available 
as a by-product of speech recognition, proved highly effective in the initial chopping 
phase, and was the most important feature used by prosodic decision trees. Additional, 
pitch-based prosodic features are also effective as features in the decision tree. 
The results obtained with recognized words (at 30% word error rate) did not differ 
greatly from those obtained with correct word transcripts. No significant degradation 
was found with the words-only segmentation model, while the best combined model 
exhibited about a 5% error increase with recognized words. The lack of degradation 
on the words-only model may be partly due to the fact that the recognizer generally 
outputs fewer words than contained in the correct transcripts, biasing the segmenter 
toward a lower false alarm rate. Still, part of the appeal of prosodic segmentation is 
that it is inherently robust to recognition errors. This characteristic makes it even more 
attractive for use in domains with higher error rates due to poor acoustic conditions or 
more conversational speaking styles. It is especially encouraging that the prosody-only 
segmenter achieved competitive performance. 
It was fairly straightforward to modify the original Dragon HMM segmenter (Yam- 
ron et al. 1998), which is based purely on topical word usage, to incorporate discourse 
cues, both lexical and prosodic. The addition of these discourse cues proved highly 
effective, especially in the case of prosody. The alternative knowledge source combi- 
nation approach, using HMM posterior probabilities as decision tree inputs, was also 
effective, although less so than the HMM-based approach. Note that the HMM-based 
integration, as implemented here, makes more stringent assumptions about the in- 
dependence of lexical and prosodic cues. The combined decision tree, on the other 
hand, has some ability to model dependencies between lexical and prosodic cues. The 
fact that the HMM-based combination approach gave the best results is thus indirect 
evidence that lexical and prosodic knowledge sources are indeed largely independent. 
Apart from the question of probabilistic independence, it seems that lexical and 
prosodic models are also complementary in the errors they make. This is manifested 
in the different distributions of miss and false alarm errors discussed in Section 4.5. 
It is also easy to find examples where the two models make complementary errors. 
Figure 8 shows two topic boundaries that are missed by one model but not the other. 
Several aspects of our model are preliminary or suboptimal in nature and can be 
improved. Even when testing on recognized words, we used parameters optimized 
on forced alignments. This is suboptimal but convenient, since it avoids the need to 
run word recognition on the relatively large training set. Since results on recognized 
words are very similar to those on true words, we can conclude that not much was lost 
with this expedient. Also, we have not yet optimized the chopping stage relative to 
the combined model (only relative to the words-only segmenter). The use of prosodic 
features other than pause duration for chopping should further improve the overall 
performance. 
The improvement obtained with source-dependent topic switch penalties and 
posterior thresholds suggests that more comprehensive source-dependent modeling 
would be beneficial. In particular, both prosodic and lexical discourse cues are likely 
to be somewhat source specific (e.g., because of different show formats and different 
speakers). Given enough training data, it is straightforward to train source-dependent 
models. 
53 
Computational Linguistics Volume 27, Number 1 
(a) •.. we have a severe thunderstorm watch two severe thunderstorm watches 
and a tornado watch in effect the tornado watch in effect back here in eastern 
colorado the two severe thunderstorm watches here indiana over into ohio 
those obviously associated with this line which is already been producing 
some hail i'll be back in a moment we'll take a look at our forecast weather 
map see if we can cool it off in the east will be very cold tonight minus seven 
degrees <TOPIC_CHANGE> 
LM probability: 0.018713 
PM probability: 0.937276 
karen just walked in was in the computer and found out for me that national 
airport in washington d. c. did hit one hundred degrees today it's a record 
high for them it's going to be uh hot again tomorrow but it will begin to 
cool off the que question is what time of day is this cold front going to move 
by your house if you want to know how warm it's going to be tomorrow 
comes through early in the day won't be that hot at all midday it'll still be 
into the nineties but not as hot as it was today comes through late in the day 
you'll still be in the upper nineties but some relief is on the way ... 
(b) ... you know the if if the president has been unfaithful to his wife and at 
this point you know i simply don't know any of the facts other than the 
bits and pieces that we hear and they're simply allegations at this point but 
being unfaithful to your wife isn't necessarily a crime lying in an affidavit is 
a crime inducing someone to lie in an affidavit is a crime but that occurred 
after this apparent taping so i'll tell you there are going to be extremely 
thorny legal issues that will have to be sorted out white house spokesman 
mike mccurry says the administration will cooperate in starr's investigation 
<TOPIC_CHANGE> 
LM probability: 1•000000 
PM probability: 0.134409 
cubans have been waiting for this day for a long time after months of plan- 
ning and preparation pope john paul the second will make his first visit to 
the island nation this afternoon it is the first pilgrimage ever by a pope to 
cuba judy fortin joins us now from havana with more .... 
Figure 8 
Examples of true topic boundaries where lexical and prosodic models make opposite 
decisions• (a) The prosodic model correctly predicts a topic change, the LM does not. (b) The 
LM predicts a topic change, the prosodic model does not. 
6. Conclusion 
We have presented a probabilistic approach to topic segmentation of speech, combining 
both lexical and prosodic cues. Topical word usage and lexical discourse cues are 
represented by language models embedded in an HMM. Prosodic discourse cues, 
such as pause durations and pitch resets, are modeled by a decision tree based on 
automatically extracted acoustic features and alignments. Lexical and prosodic features 
can be combined either in the HMM or in the decision tree framework. 
Our topic segmentation model was evaluated on broadcast news speech, and 
found to give competitive performance (around 14% error according to the weighted 
TDT2 segmentation cost metric). Notably, the segmentation accuracy of the prosodic 
54 
Tier, Hakkani-T(ir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues 
model alone is competitive with a word-based segmenter, and a combined prosodic/ 
lexical HMM achieves a substantial error reduction over the individual knowledge 
sources. 
Acknowledgments 
We thank Becky Bates, Madelaine PlauchG 
Ze'ev Rivlin, Ananth Sankar, and Kemal 
S6nmez for invaluable assistance in 
preparing the data for this study. The paper 
was greatly improved as a result of 
comments by Andy Kehler, Madelaine 
Plauch6, and the anonymous reviewers. 
This research was supported by DARPA 
and NSF under NSF grant IRI-9619921 and 
DARPA contract no. N66001-97-C-8544. The 
views herein are those of the authors and 
should not be interpreted as representing 
the policies of the funding agencies. 

References 
Allan, J., J. Carbonell, G. Doddington, 
J. Yamron, and Y. Yang. 1998. Topic 
detection and tracking pilot study: Final 
report. In Proceedings of DARPA Broadcast 
News Transcription and Understanding 
Workshop, pages 194-218, Lansdowne, VA, 
February. Morgan Kaufmann. 
Ayers, Gayle M. 1994. Discourse functions 
of pitch range in spontaneous and read 
speech. In Working Papers in Linguistics 
No. 44. Ohio State University, pages 1-49. 
Baum, Leonard E., Ted Petrie, George 
Soules, and Norman Weiss. 1970. A 
maximization technique occurring in the 
statistical analysis of probabilistic 
functions in Markov chains. The Annals of 
Mathematical Statistics, 41(1):164-171. 
Beeferman, Doug, Adam Berger, and John 
Lafferty. 1999. Statistical models for text 
segmentation. Machine Learning, 
34(1-3):177-210. Special Issue on Natural 
Language Learning. 
Breiman, L., J. H. Friedman, R. A. Olshen, 
and C. J. Stone. 1984. Classification and 
Regression Trees. Wadsworth and Brooks, 
Pacific Grove, CA. 
Brown, G., K. L. Currie, and J. Kenworthy. 
1980. Questions of Intonation. University 
Park Press, Baltimore. 
Brubaker, R. S. 1972. Rate and pause 
characteristics of oral reading. Journal of 
Psycholinguistic Research, 1:141-147. 
Buntine, Wray and Rich Caruana, 1992. 
Introduction to IND Version 2.1 and Recursive 
Partitioning. NASA Ames Research 
Center, Moffett Field, CA, December. 
Cieri, Chris, David Graft, Mark Liberman, 
Nii Martey, and Stephanie Strassell. 1999. 
The TDT-2 text and speech corpus. In 
Proceedings of DARPA Broadcast News 
Workshop, pages 57-60, Herndon, VA, 
February. Morgan Kaufmann. 
Doddington, George. 1998. The Topic 
Detection and Tracking Phase 2 (TDT2) 
evaluation plan. In Proceedings of DARPA 
Broadcast News Transcription and 
Understanding Workshop, pages 223-229, 
Lansdowne, VA, February. Morgan 
Kaufmann. Revised version available 
from http://www.nist.gov/speech/ 
tests/tdt/tdt98/. 
Entropic Research Laboratory, 1993. ESPS 
Version 5.0 Programs Manual. Washington, 
D.C. August. 
Fiscus, Jon, George Doddington, John 
Garofolo, and Alvin Martin. 1999. NIST's 
1998 Topic Detection and Tracking 
evaluation (TDT2). In Proceedings of 
DARPA Broadcast News Workshop, 
pages 19-24, Herndon, VA, February. 
Morgan Kaufmann. 
Geluykens, R. and M. Swerts. 1993. Local 
and global prosodic cues to discourse 
organization in dialogues. In Working 
Papers 41, Proceedings of ESCA Workshop on 
Prosody, pages 108-111, Lurid, Sweden. 
Grosz, B. and J. Hirschberg. 1992. Some 
intonational characteristics of discourse 
structure. In John J. Ohala, Terrance M. 
Nearey, Bruce L. Derwing, Megan M. 
Hodge, and Grace E. Wiebe, editors, 
Proceedings of the International Conference on 
Spoken Language Processing, volume 1, 
pages 429432, Banff, Canada, October. 
Grosz, B. and C. Sidner. 1986. Attention, 
intention, and the structure of discourse. 
Computational Linguistics, 12(3):175-204. 
Hakkani-Tiir, Dilek, GOkhan Tiir, Andreas 
Stolcke, and Elizabeth Shriberg. 1999. 
Combining words and prosody for 
information extraction from speech. In 
Proceedings of the 6th European Conference on 
Speech Communication and Technology, 
volume 5, pages 1991-1994, Budapest, 
September. 
Hearst, Marti A. 1994. Multi-paragraph 
segmentation of expository text. In 
Proceedings of the 32nd Annual Meeting, 
pages 9-16, New Mexico State University, 
Las Cruces, NM, June. Association for 
Computational Linguistics. 
Hearst, Marti A. 1997. TextTiling: 
Segmenting text info multi-paragraph 
subtopic passages. Computational 
Linguistics, 23(1):33-64. 
Hirschberg, Julia and Christine Nakatani. 
1996. A prosodic analysis of discourse 
segments in direction-giving monologues. 
In Proceedings of the 34th Annual Meeting, 
pages 286-293, Santa Cruz, CA, June. 
Association for Computational 
Linguistics. 
Hirschberg, Julia and Christine Nakatani. 
1998. Acoustic indicators of topic 
segmentation. In Robert H. Mannell and 
Jordi Robert-Ribes, editors, Proceedings of 
the International Conference on Spoken 
Language Processing, pages 976-979, 
Sydney, December. Australian Speech 
Science and Technology Association. 
John, George H., Ron Kohavi, and Karl 
Pfleger. 1994. Irrelevant features and the 
subset selection problem. In William W. 
Cohen and Haym Hirsh, editors, Machine 
Learning: Proceedings of the llth International 
Conference, pages 121-129, San Francisco. 
Morgan Kaufmann. 
Koopmans-van Beinum, Florien J. and 
Monique E. van Donzel. 1996. 
Relationship between discourse structure 
and dynamic speech rate. In H. Timothy 
Bunnell and William Idsardi, editors, 
Proceedings of the International Conference on 
Spoken Language Processing, volume 3, 
pages 1724-1727, Philadelphia, October. 
Kozima, H. 1993. Text segmentation based 
on similarity between words. In 
Proceedings of the 31st Annual Meeting, 
pages 286-288, Ohio State University, 
Columbus, Ohio, June. Association for 
Computational Linguistics. 
Litman, Diane J. and Rebecca J. Passonneau. 
1995. Combining multiple knowledge 
sources for discourse segmentation. In 
Proceedings of the 33rd Annual Meeting, 
pages 108-115, MIT, Cambridge, MA, 
June. Association for Computational 
Linguistics. 
Nakajima, Shin'ya and J. F. Allen. 1993. A 
study on prosody and discourse structure 
in cooperative dialogues. Phonetica, 
50:197-210. 
Nakajima, Shin'ya and Hajime Tsukada. 
1997. Prosodic features of utterances in 
task-oriented dialogues. In Yoshinori 
Sagisaka, Nick Campbell, and Norio 
Higuchi, editors, Computing Prosody: 
Computational Models for Processing 
Spontaneous Speech. Springer, New York, 
chapter 7, pages 81-94. 
NIST. 1999. 1999 Topic Detection and 
Tracking Evaluation Project (TDT-3) 
Evaluation Project. Speech Group, 
National Institute for Standards and 
Technology, Gaithersburg, MD. http:// 
www.nist.gov/speech/tests/tdt/tdt99/. 
Passonneau, Rebecca J. and Diane J. Litman. 
1997. Discourse segmentation by human 
and automated means. Computational 
Linguistics, 23(1):103-139. 
Ponte, J. M. and W. B. Croft. 1997. Text 
segmentation by topic. In Proceedings of the 
First European Conference on Research and 
Advanced Technology for Digital Libraries, 
pages 120-129, Pisa, Italy. 
Przybocki, M. A. and A. F. Martin. 1999. 
The 1999 NIST speaker recognition 
evaluation, using summed two-channel 
telephone data for speaker detection and 
speaker tracking. In Proceedings of the 6th 
European Conference on Speech 
Communication and Technology, volume 5, 
pages 2215-2218, Budapest, September. 
Rabiner, L. R. and B. H. Juang. 1986. An 
introduction to hidden Markov models. 
IEEE ASSP Magazine, 3(1):4-16, January. 
Reynar, Jeffrey C. 1994. An automatic 
method of finding topic boundaries. In 
Proceedings of the 32nd Annual Meeting, 
pages 331-333, New Mexico State 
University, Las Cruces, NM, June. 
Association for Computational 
Linguistics. 
Reynar, Jeffrey C. 1999. Statistical models 
for topic segmentation. In Proceedings of 
the 37th Annual Meeting, pages 357-364, 
University of Maryland, College Park, 
MD, June. Association for Computational 
Linguistics. 
Sankar, Ananth, Fuliang Weng, Ze'ev Rivlin, 
Andreas Stolcke, and Ramana Rao Gadde. 
1998. The development of SRI's 1997 
Broadcast News transcription system. In 
Proceedings DARPA Broadcast News 
Transcription and Understanding Workshop, 
pages 91-96, Lansdowne, VA, February. 
Morgan Kaufmann. 
Shriberg, Elizabeth, Rebecca Bates, and 
Andreas Stolcke. 1997. A prosody-only 
decision-tree model for disfluency 
detection. In G. Kokkinakis, N. Fakotakis, 
and E. Dermatas, editors, Proceedings of the 
5th European Conference on Speech 
Communication and Technology, volume 5, 
pages 2383-2386, Rhodes, Greece, 
September. 
Shriberg, Elizabeth, Andreas Stolcke, Dilek 
Hakkani-Ti~r, and G6khan Tiir. 2000. 
Prosody-based automatic segmentation of 
speech into sentences and topics. Speech 
Communication, 32(1-2), pages 127-154. 
Special Issue on Accessing Information in 
Spoken Audio. 
S6nmez, Kemal, Elizabeth Shriberg, Larry 
Heck, and Mitchel Weintraub. 1998. 
Modeling dynamic prosodic variation for 
speaker verification. In Robert H. Mannell 
and Jordi Robert°Ribes, editors, 
Proceedings of the International Conference on 
Spoken Language Processing, volume 7, 
pages 3189-3192, Sydney, December. 
Australian Speech Science and 
Technology Association. 
Stolcke, Andreas, Klaus Ries, Noah Coccaro, 
Elizabeth Shriberg, Dan Jurafsky, Paul 
Taylor, Rachel Martin, Carol Van 
Ess-Dykema, and Marie Meteer. 2000. 
Dialogue act modeling for automatic 
tagging and recognition of conversational 
speech. Computational Linguistics, 
26(3):339-373. 
Stolcke, Andreas and Elizabeth Shriberg. 
1996. Automatic linguistic segmentation 
of conversational speech. In H. Timothy 
Bunnell and William Idsardi, editors, 
Proceedings of the International Conference on 
Spoken Language Processing, volume 2, 
pages 1005-1008, Philadelphia, October. 
Stolcke, Andreas, Elizabeth Shriberg, 
Rebecca Bates, Mari Ostendorf, Dilek 
Hakkani, Madelaine PlauchG GOkhan 
Tier, and Yu Lu. 1998. Automatic 
detection of sentence boundaries and 
disfluencies based on recognized words. 
In Robert H. Mannell and Jordi 
Robert-Ribes, editors, Proceedings of the 
International Conference on Spoken Language 
Processing, volume 5, pages 2247-2250, 
Sydney, December. Australian Speech 
Science and Technology Association. 
Swerts, M. 1997. Prosodic features at 
discourse boundaries of different 
strength. Journal of the Acoustical Society of 
America, 101:514-521. 
Swerts, M., R. Geluykens, and J. Terken. 
1992. Prosodic correlates of discourse 
units in spontaneous speech. In John J. 
Ohala, Terrance M. Nearey, Bruce L. 
Derwing, Megan M. Hodge, and Grace E. 
Wiebe, editors, Proceedings of the 
International Conference on Spoken Language 
Processing, volume 1, pages 421-424, 
Banff, Canada, October. 
Swerts, M. and M. Ostendorf. 1997. Prosodic 
and lexical indications of discourse 
structure in human-machine interactions. 
Speech Communication, 22(1):25--41. 
Vaissi6re, Jacqueline. 1983. 
Language-independent prosodic features. 
In A. Cutler and D. R. Ladd, editors, 
Prosody: Models and Measurements. 
Springer, Berlin, chapter 5, pages 53-66. 
van Mulbregt, P., I. Carp, L. Gillick, S. Lowe, 
and J. Yamron. 1999. Segmentation of 
automatically transcribed broadcast news 
text. In Proceedings of DARPA Broadcast 
News Workshop, pages 77-80, Herndon, 
VA, February. Morgan Kaufmann. 
Viterbi, A. 1967. Error bounds for 
convolutional codes and an 
asymptotically optimum decoding 
algorithm. IEEE Transactions on Information 
Theory, 13:260-269. 
Yamron, J. P., I. Carp, L. Gillick, S. Lowe, 
and P. van Mulbregt. 1998. A hidden 
Markov model approach to text 
segmentation and event tracking. In 
Proceedings of the IEEE Conference on 
Acoustics, Speech, and Signal Processing, 
volume 1, pages 333-336, Seattle, WA, 
May. 
