Conference PaperPDF Available

Making Sense of Subtitles: Sentence Boundary Detection and Speaker Change Detection in Unpunctuated Texts


Abstract and Figures

The rise of deep learning methods has transformed the research area of natural language processing beyond recognition. New benchmark performances are reported on a daily basis ranging from machine translation to questionanswering. Yet, some of the unsolved practical research questions are not in the spotlight and this includes, for example, issues arising at the interface between spoken and written language processing. We identify sentence boundary detection and speaker change detection applied to automatically transcribed texts as two NLP problems that have not yet received much attention but are nevertheless of practical relevance. We frame both problems as binary tagging tasks that can be addressed by fine-tuning a transformer model and we report promising results.
Content may be subject to copyright.
Making Sense of Subtitles: Sentence Boundary
Detection and Speaker Change Detection in
Unpunctuated Texts
Gregor Donabauer
University of Regensburg
Regensburg, Germany
Udo Kruschwitz
University of Regensburg
Regensburg, Germany
David Corney
Full Fact
London, UK
The rise of deep learning methods has transformed the re-
search area of natural language processing beyond recog-
nition. New benchmark performances are reported on a
daily basis ranging from machine translation to question-
answering. Yet, some of the unsolved practical research ques-
tions are not in the spotlight and this includes, for example,
issues arising at the interface between spoken and written
language processing.
We identify sentence boundary detection and speaker
change detection applied to automatically transcribed texts
as two NLP problems that have not yet received much atten-
tion but are nevertheless of practical relevance. We frame
both problems as binary tagging tasks that can be addressed
by ne-tuning a transformer model and we report promising
ACM Reference Format:
Gregor Donabauer, Udo Kruschwitz, and David Corney. 2021. Mak-
ing Sense of Subtitles: Sentence Boundary Detection and Speaker
Change Detection in Unpunctuated Texts. In Companion Proceed-
ings of the Web Conference 2021 (WWW ’21 Companion), April 19–
23, 2021, Ljubljana, Slovenia. ACM, New York, NY, USA, 7 pages.
Text and speech processing are closely related research areas,
yet one still gets the impression that research is conducted in
two separate communities (and if you add video as another
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are not
made or distributed for prot or commercial advantage and that copies bear
this notice and the full citation on the rst page. Copyrights for components
of this work owned by others than ACM must be honored. Abstracting with
credit is permitted. To copy otherwise, or republish, to post on servers or to
redistribute to lists, requires prior specic permission and/or a fee. Request
permissions from
WWW ’21 Companion, April 19–23, 2021, Ljubljana, Slovenia
©2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8313-4/21/04.
mode, then you get another research community). Some
of the interesting problems can therefore be found at the
boundary of the dierent elds. While our research is rmly
rooted in text processing, we see our work as a contribution
to help bridge the gap between work conducted on written
and spoken language.
The immediate motivation for our work comes from the
domain of fact checking. Fact checkers monitor the media to
identify potentially harmful or misleading claims. It is impor-
tant for them to know who said what and when in order to
nd claims worth investigating. To cope with the volume of
potential claims, and the limited time available, fact checkers
are increasingly turning to technology to help, including
]. These tools can help identify claims worth check-
ing, nd repeats of claims that have already been checked
or even assist in the verication process directly. Most such
tools rely on text as input and require the text to be split into
Some media sources, such as ocial parliamentary reports,
are very rich, providing marked-up text showing sentence
and speech boundaries and tag each speaker with a unique
identier. Newspapers and social media usually give some
information about speakers though often implicitly or am-
biguously. In contrast, audio and video feeds – including TV
and radio news broadcasts and videos shared on YouTube
or Facebook – do not usually contain explicit information
about speakers. In some cases, automatic captioning may
be used to generate a transcript, or subtitles may be made
available by broadcasters. But in many cases, using post-hoc
speech-to-text processing is the only way to extract text.
There is thus a need to bridge the gap between large vol-
umes of audio-visual content and the existing text-based
tools that fact checkers use. Our work addresses two aspects
of this gap, namely detecting sentence boundaries in tran-
scripts of speech and detecting when the speaker changes,
such as during an interview or debate.
WWW ’21 Companion, April 19–23, 2021, Ljubljana, Slovenia Donabauer, Kruschwitz and Corney
Figure 1 illustrates the absence of text structure (including
capitalisation and punctuation) as well as conversational
structure, as the result of automatic transcription.1
Figure 1: Auto-generated subtitles on YouTube.
Figure 2 shows the same example dialogue as in Figure 1
but with the full sentence and conversational structure in
place, making it far easier to read and process.
Figure 2: Sample from the DailyDialog dataset.
The problem we address is the restoration of some funda-
mental structure from unpunctuated text data, particularly in
the context of transcribed speech and conversation data. In a
rst step we restore sentence boundary information. Sentences
are generally considered a fundamental information unit of
written text, e.g. [
]. Therefore, this task has been well-
studied, e.g. in the context of automatic speech recognition
]. Frequently, the problem of sentence boundary
detection in unpunctuated text is treated as a tagging task
tackled using IOB sequence labeling [
] as also used in
named entity recognition (NER) [19].
As a subsequent task we want to restore information on
speaker changes based on the previously identied sentences
– as much transcribed data is based on more than a single
speaker (as seen with the earlier example). We therefore want
to detect whether the next sentence was uttered by the same
person or not, an important step in the context of dialogue
data restoration and necessary for further postprocessing in
this area, e.g. [27].
This example is not taken from a fact-checking use case but adopted from
one of the benchmark collections we use in our experimental work.
Given the impressive advances in a variety of NLP tasks
using a transformer-based architecture, e.g. [
], we use this
approach to tackle the problem at hand. More specically,
we treat both steps as sequence tagging tasks using binary
labels by ne-tuning BERT and we compare our work against
strong baselines on previously used benchmarks.
By making all our resources (code and test collections)
available our aim is to provide a solid reference point and a
strong benchmark for future work.
We will briey discuss each of the two problems in turn, i.e.
Sentence Boundary Detection (SBD) and Speaker Change
Detection (SCD).
2.1 Sentence Boundary Detection (SBD)
SBD is an important and well-studied text processing step but
it typically relies on the presence of punctuation within the
input text [
]. Even with such punctuation it can be a dicult
task, e.g. [
], and traditional approaches use a variety
of architectures including CRFs [
] and combinations of
HMMs, maximum likelihood as well as maximum entropy
approaches [
]. With unpunctuated texts (and lack of word-
casing information) it becomes a lot harder as even humans
nd it dicult to determine sentence boundaries in this case
], as illustrated in Figure 1. Song et al
. [22]
simplify the
problem we are addressing by aiming to detect the sentence
boundary within a 5-word chunk – using YouTube subtitle
data. Using LSTMs they report an F1 of 81.43% at predicting
the position of the sample’s sentence boundaries but did
not consider any chunks without sentence boundary. Le
presents a hybrid model (using BiLSTMs and CRFs) originally
used for NER that was evaluated on SBD in the context
of conversational data by preprocessing the CornellMovie-
Dialogue and the DailyDialog datasets to obtain samples that
neither contain sentence boundary punctuation nor word-
casing information (they also predict whether the sentence
is a statement or a question). They report F1-scores of 81.62%
for questions and 91.90% for statements on the CornellMovie-
Dialogue data and 94.66% (questions) and 96.29% (statements)
on DailyDialog. To the best of our knowledge, only Du et
al. [
] present a transformer-based approach to the problem,
but they assume partially punctuated text and word-casing
Hence, Le
and Song et al
. [22]
appear to be the strongest
baselines to compare our approach against.
2.2 Speaker Change Detection (SCD)
Most related work in this area is concerned with audio-based
] with the exception of Meng et al
. [16]
who collected transcribed conversations. The text-data is
Making Sense of Subtitles WWW ’21 Companion, April 19–23, 2021, Ljubljana, Slovenia
pre-processed to lowercase and contains punctuation. They
compare dierent deep learning approaches with the best-
performing being a RNN with LSTM layers, hierarchical con-
text and static attention giving an F1-score of 78.4%. Apart
from this work, there are other approaches that treat the
topic of text-based SCD, though not explicitly, e.g. Serban
and Pineau
, or they aim at assigning specic speaker
ids to sentences [15].
In conclusion, Meng et al
. [16]
appears to be the most plau-
sible baseline to choose. We will also adopt their benchmark
corpus for comparison.
We treat both tasks, SBD and SCD, as sequence labeling tasks.
More specically, we apply IO tagging to label sequences
of tokenized text data adopted from NER [
]. In both cases
two distinct labels are sucient to identify whether a token
marks the sentence boundary or the start of an utterance
by a dierent speaker, respectively. We use a pre-trained
transformer-based language model and ne-tune it on each
of the two tasks. The resulting IO sequence taggers allow
us to deduce sentence boundaries (SBD-TT) and speaker
changes (SCD-TT).
For the experimental setup we opted to ne-tune
(given our input is expected to be in lower-
case, we do not need casing information within the language
model). The model training and evaluation are implemented
using the PyTorch
version of the Python huggingface
formers library. The model’s output is produced utilizing a
dense layer as classication head. Using the argmax opera-
tor, we can deduce labels for resulting vectors in the same
dimension as the label list for each introduced token. The
processes are executed using three Nvidia GeForce RTX 2080
GPUs with an overall memory size of 24GB. Most experi-
ments are executed in 3 epochs, using a batch-size of 16. The
number of epochs is set according to the recommendation of
Devlin et al
. [3]
. Unless specied further down, we refer to
our GitHub repository
for task-specic sequence lengths,
deviations from our parameter settings, all source code, data,
models and additional information.
Where appropriate we apply paired
-tests for signicance
testing (at 𝑝<0.01).
For fair comparison we adopt datasets proposed in prior
work. For SBD we use a Stanford Lectures Dataset reproduced
from Song et al
. [22]
, the DailyDialog Dataset proposed by Li
et al
. [10]
and applied in [
]. In addition we also experiment
with a hybrid set. For SCD we use the dataset introduced by
Meng et al. [16] (we refer to it as MengCorpus).
4.1 Stanford Lectures
Song et al
. [22]
collected the hand-transcribed lecture subti-
tles provided by Stanford University on YouTube using the
text data associated with the lecture series “Natural Language
Processing with Deep Learning” and “Human Behavioral Bi-
ology”. We replicated this process. In addition to that we
identied ve more lecture series Stanford University pro-
vides subtitles for and collected the accompanying text data
(resulting in a corpus about 4 times as big). Details on the
exact lectures, their source as well as the data themselves
can be found on our GitHub repository. For further data pre-
processing, we basically adopt the methods introduced by
Song et al
. [22]
. The punctuated transcripts provide ground
truth information. We transform all text data to lower-case
and tokenize the data using NLTK
. Sentences with fewer
than 7 or more than 70 words are discarded, and any punc-
tuation is removed. Finally, all tokens including sentence
boundary positions are tagged. The text is then split into
chunks of 64 token-tag pairs. Sentence boundary tags can ap-
pear anywhere within those chunks (which is more generic
than the 5-word chunk approach by Song et al. [22]).
The preprocessing steps applied in our work are depicted
in Figure 3.
Figure 3: Data Preprocessing for SBD.
WWW ’21 Companion, April 19–23, 2021, Ljubljana, Slovenia Donabauer, Kruschwitz and Corney
Dataset Train Dev Test
Stanford Lectures 19,285 2,411 2,411
DailyDialog 15,259 1,374 1,405
Hybrid Dataset 34,156 3,848 4,142
MengCorpus 174,702 22,065 21,918
Table 1: Number of samples per dataset
In line with convention, we split the data into training
(80%), development (10%) and test set (10%) [
]. All samples
are saved in CoNLL-2003 format [
] as is frequently used
for tagging tasks like NER.
4.2 DailyDialog
The second dataset used, originally introduced by Li et al
captures daily communication with a wide variety of
daily life’s topics. Since the complete text-data is human
written, it is expected to be less noisy than for example
automatically transcribed conversational data. The dataset
was used for SBD by Le [8] and comes with a 80:10:10 split.
4.3 Hybrid Dataset
To be able to train one single model that can predict sentence
boundaries within conversational data as well as a single
person’s speech data, we create a mixture of the two datasets
introduced above. Given the conversational structure of the
text we cannot simply randomize development and test sets.
Instead we split the data into chunks of 10 sentences each,
which are subsequently shued. They are concatenated and
afterwards split into samples of length 64. Thereby, the struc-
ture of subsequent sentences as well as dialogues should be
preserved. The basic properties of each dataset are listed in
Table 1.
4.4 MengCorpus
For SCD, we use the dataset introduced by Meng et al
. [16]
It is a collection of 3,000 hours of hand-transcribed CNN talk-
shows. The transcripts provide speaker change information
through assigned speaker IDs and comprise approximately
1.5 million utterances. They are split into train, development
and test set by an 80:10:10 ratio. The data are provided in
form of one sentence per line. As before we use NLTK to
perform tokenization and then mark speaker changes accord-
ingly. We split the text into samples of 7 successive sentences
to include as much context as possible and satisfy the max-
imum sequence length of BERT (512 tokens). Hence, the
resulting samples have diering lengths in terms of occur-
ring token-tag pairs. They are saved in CoNLL-2003 format.
Basic properties are included in Table 1.
Accuracy F1-Score
Song et al. [22] 70.84% 81.43%
Le [8] 89.80% 93.07%
SBD-TT 92.49% 93.68%
Table 2: Sentence Boundary Detection applied to Stan-
ford Lectures as described by Song et al. [22]
Data Accuracy F1-Score
Stanford Lectures (big) 97.98% 79.83%
Hybrid Dataset 97.64% 85.31%
Table 3: Sentence Boundary Detection benchmarks
applied to our own datasets
We will rst report the experimental results and then discuss
those further in the next Section. All metrics are calculated
on a token-based level – in line with what had been adopted
in the work we compare our results against.
5.1 Sentence Boundary Detection
Since neither code nor data were available for Song et al
, we simply reproduce their accuracy and F1 measures
in Table 2. However, as described we also replicated the
data collection and processing steps and run 5-fold cross-
validation to compare our approach (SBD-TT) against Le
. The results can also be found in Table 2. A paired
reveals that SBD-TT outperforms Le
in terms of both
accuracy and F1 (signicant at 𝑝<0.01).
We also trained and tested SBD-TT on the DailyDialog
dataset and get an F1 for statements of 97.19% (vs. 96.29%
reported by Le [8]) and for questions: 95.64% (vs. 94.66%).
We observe that our SBD-TT approach for sentence bound-
ary detection outperforms state-of-the-art methods and con-
clude that our vanilla transformer-based approach using
BERT leaves scope for further advances.
As an additional contribution and to foster reproducibility
we also provide benchmarks obtained from the two corpora
we introduced in this paper (and which are available on
our GitHub account): the larger Stanford Lectures Dataset
and the Hybrid Dataset, and these are reported in Table 3.
Note that for these experiments we use sequence lengths of
64 words (unlike the much shorter 5-word chunks used to
compare against baselines in Table 2).
Going back to the discussion of related work, one might
ask why we did not compare our results against those re-
ported by Du et al
. [4]
? That is because they treat a similar
but dierent problem. They have word-casing information
Making Sense of Subtitles WWW ’21 Companion, April 19–23, 2021, Ljubljana, Slovenia
Model Accuracy Precision Recall F1-Score
Random guess 61.8% 26.0% 25.0% 25.4%
Logistic Regression w/ (uni+bi)-gram 80.5% 73.0% 39.0% 50.9%
DNN w/ (uni+bi)-gram 76.6% 54.4% 58.8% 56.5%
CNN w/o context 77.8% 56.8% 58.9% 57.8%
RNN w/o context 83.3% 72.5% 57.1% 63.9%
RNN w/ context (non-hierarchical) 83.7% 72.6% 60.0% 65.7%
RNN w/ context (hierarchical) 85.1% 74.6% 64.6% 69.2%
SCD-TT w/o sliding window evaluation 82.4% 76.2% 72.1% 74.1%
RNN w/ context (hierarchical) + static attention 89.2% 81.5% 75.6% 78.4%
SCD-TT w/ sliding window evaluation 85.4% 80.1% 77.6% 78.8%
Table 4: Results of Speaker Change Detection in comparison to scores reported by Meng et al. [16]
available and more than 90% of the "end of sentence" tokens.
Therefore it was not a suitable comparison.
5.2 Speaker Change Detection
We have two dierent results to report and to compare
against the baseline scores achieved by Meng et al
. [16]
, since
we use two dierent approaches for evaluation. The rst eval-
uation method simply uses 7 successive sentences at a time
and tags those with the speaker change labels. The second
method also uses 7 successive sentences as an input, but only
takes into account the predictions for the middle sentence.
All other sentences are seen as context. This sliding-window
evaluation is executed with a stride of one sentence at a time.
Table 4 presents the results of Meng et al
. [16]
in comparison
to our approach, SCD-TT.
We note that our straightforward ne-tuning approach is
competitive for speaker change detection, and for the sliding-
window-based evaluation we even achieve a 0.4 percentage
point improvement in F1 compared to the best score of Meng
et al
. [16]
. In general, the results show that context is impor-
tant for the model to predict speaker changes. This conrms
the ndings described by Meng et al
. [16]
. While they used 8
sentences of context on each side of the evaluated sentence,
we are limited by the maximum sequence length of 512 to-
kens that can be used as input of our BERT-based model.
Therefore we only used 3 sentences on each side of the eval-
uated sentence as context though were still able to achieve
F1 scores slightly higher than those reported by Meng et al.
There are a number of discussion points emerging from our
experimental setup and the results we obtained.
First of all, why did we only test for statistical signicance
for SBD-TT? The reason is that while we were able to re-
produce the work of Le
, for Song et al
. [22]
we did not
have the code nor the exact data. We requested both from
the authors but did not get a response (hence the compar-
ison against reported results only). Unlike for SBD, where
we achieved a new state-of-the-art performance, in SCD our
results are on par with the best-performing alternative so
we only compare against the performance reported by Meng
et al. [16].
Also, why did we not combine the two methods, to rst
detect sentence boundaries and then detect speaker change
over that result? Our overall aim was to demonstrate the
general suitability of our approach and provide strong bench-
marks for each of the two problems rather than providing
the best possible model that combines both. As such it is
possible to use the models separately where necessary, e.g.,
when processing a single person’s transcript we do not need
the SCD model. Obviously this leaves plenty of room for
future investigations, and the provision of all resources on
GitHub will support this.
Another question arising from our strong performance
against state-of-the-art approaches (keeping in mind that
we are using a relatively straightforward architecture) is
to ask what kind of knowledge does our approach encode
that the other approaches don’t? We would argue that BERT
clearly encodes exactly the type of contextual information
that is needed for the two tasks. This information is captured
implicitly and obtained partly during ne-tuning but also in
pre-training. Note again that our aim was to demonstrate the
general suitability of a transformer-based approach. Using
other BERT-based models as well as better ne-tuning can
result in further improvements.
Finally, one might ask whether a performance that is on
par with other state-of-the-art approaches (as is the case
for speaker change detection) gives us any benet. Well, in
addition to the points just raised we should also point out
that we get better (SBD) or similar (SCD) results for both
tasks with a much simpler model. Given we only apply a very
WWW ’21 Companion, April 19–23, 2021, Ljubljana, Slovenia Donabauer, Kruschwitz and Corney
simple setup (e.g. only using BERT-base) there is potential to
push the eectiveness without losing the overall simplicity.
Coming back to the initial example, Figure 4 demonstrates
the output generated by applying each of our two models.
While the sentence segmentation works perfectly in this
case, we see that speaker change detection (predicted by the
label True) leaves room for improvement: lines 4, 7, 9 and 10
are incorrectly classied.
Figure 4: Restored structure of initial example.
With a bit of delay when compared to image processing,
natural language processing has now also witnessed a par-
adigm shift from traditional statistical approaches to deep
learning architectures. This has resulted in some staggering
performance improvements across a wide range of appli-
cations. However, there are still plenty of open problems
– often based on practical use cases. The rapidly evolving
mix of dierent types of media and new forms of interaction
highlights the fact that at the interface between dierent
communities, such as those working with spoken and those
with written textual data, there are opportunities to make
rapid progress. This can be achieved by adopting paradigms
that have already been shown to push the state of the art
forward elsewhere, most prominently transformer-based ar-
In this paper we identify the detection of sentence bound-
aries and speaker changes in unpunctuated text as problems
of natural language processing that sit at the interface be-
tween spoken and written text, and which have attracted
little interest before now. By making our methods available
to fact checkers, they may nd it as easy to identify and
analyse claims made during televised debates or news in-
terviews as claims made in online textual news sites. This
will help ensure that no matter where harmful or misleading
information is shared, it can also be identied and challenged
rapidly to limit its spread. Beyond the work of fact checkers
we envisage the proposed steps to be also incorporated in
NLP pipelines that will automatically ag up such harmful
or misleading information.6
We should note that the two tasks could be seen as indi-
vidual NLP tasks or combined as a sequence of two steps. In
our work we frame both tasks as an IO tagging problem that
is addressed using ne-tuning of a BERT-based language
The results we report demonstrate that the problems at
hand are yet another pair of examples where the transformer-
based paradigm outperforms existing baselines. There is
much scope to push the eectiveness even further as we
have only experimented with basic models.
To foster further research we also provide a range of cor-
pora and benchmarks that can be used as future reference
This work was supported by the project COURAGE: A So-
cial Media Companion Safeguarding and Educating Students
funded by the Volkswagen Foundation, grant number 95564.
Phoebe Arnold. 2020. The challenges of online fact checking. Technical
Report. Full Fact, London, UK.
S. Chen and P.S. Gopalakrishnan. 1998. Speaker, Environment and
Channel Change Detection and Clustering via the Bayesian Informa-
tion Criterion. In DARPA Broadcast News Transcription and Understand-
ing Workshop. Landsdowne, USA, 127–132.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova.
2019. BERT: Pre-training of Deep Bidirectional Transformers for Lan-
guage Understanding. In Proceedings of the 2019 Conference of the
North American Chapter of the Association for Computational Linguis-
tics: Human Language Technologies, Volume 1 (Long and Short Papers).
Association for Computational Linguistics, Minneapolis, Minnesota,
Jinhua Du, Yan Huang, and Karo Moilanen. 2019. AIG Investments.AI
at the FinSBD Task: Sentence Boundary Detection through Sequence
Labelling and BERT Fine-tuning. In Proceedings of the First Workshop
on Financial Technology and Natural Language Processing. Association
for Computational Linguistics, Macao, China, 81–87. https://www.
Dan Gillick. 2009. Sentence Boundary Detection and the Problem with
the U.S.. In Proceedings of Human Language Technologies: The 2009
Annual Conference of the North American Chapter of the Association for
Computational Linguistics, Companion Volume: Short Papers (NAACL-
Short ’09). Association for Computational Linguistics, USA, 241–244.
C. C. Juin, R. X. J. Wei, L. F. D’Haro, and R. E. Banchs. 2017. Punctuation
prediction using a bidirectional recurrent neural network with part-
of-speech tagging. In TENCON 2017 - 2017 IEEE Region 10 Conference.
IEEE, Penang, Malaysia, 1806–1811.
As part of the COURAGE research project we are exploring ways to help
teenagers manage social media exposure by providing a virtual companion
that would, among other things, automatically identify examples of hate
speech, bullying or other toxic content [17].
Making Sense of Subtitles WWW ’21 Companion, April 19–23, 2021, Ljubljana, Slovenia
Daniel Jurafsky and James Martin. 2020. Speech and Language Pro-
cessing: An Introduction to Natural Language Processing,Computational
Linguistics, and Speech Recognition (third (draft) ed.). https://web.
The Anh Le. 2020. Sequence Labeling Approach to the Task of Sentence
Boundary Detection. In Proceedings of the 4th International Conference
on Machine Learning and Soft Computing (ICMLSC 2020). ACM, New
York, NY, USA, 144–148.
Joan Persily Levinson. 1985. Punctuation and the orthographic sentence:
a linguistic analysis. Doctoral dissertation. City University of New
Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi
Niu. 2017. DailyDialog: A Manually Labelled Multi-turn Dialogue
Dataset. In Proceedings of the Eighth International Joint Conference on
Natural Language Processing (Volume 1: Long Papers). Asian Federation
of Natural Language Processing, Taipei, Taiwan, 986–995. https:
Yang Liu, Andreas Stolcke, Elizabeth Shriberg, and Mary Harper. 2004.
Comparing and Combining Generative and Posterior Probability Mod-
els: Some Advances in Sentence Boundary Detection in Speech. In
Proceedings of the 2004 Conference on Empirical Methods in Natural Lan-
guage Processing. Association for Computational Linguistics, Barcelona,
Spain, 64–71.
Yang Liu, Andreas Stolcke, Elizabeth Shriberg, and Mary Harper. 2005.
Using Conditional Random Fields for Sentence Boundary Detection in
Speech. In Proceedings of the 43rd Annual Meeting on Association for
Computational Linguistics (ACL ’05). Association for Computational
Linguistics, USA, 451–458.
Lie Lu and Hong-Jiang Zhang. 2002. Speaker Change Detection and
Tracking in Real-Time News Broadcasting Analysis. In Proceedings
of the Tenth ACM International Conference on Multimedia (MULTIME-
DIA ’02). Association for Computing Machinery, New York, NY, USA,
Lie Lu and Hong-Jiang Zhang. 2005. Unsupervised Speaker Segmenta-
tion and Tracking in Real-Time Audio Content Analysis. Multimedia
Syst. 10, 4 (April 2005), 332–343.
Kaixin Ma, Catherine Xiao, and Jinho D. Choi. 2017. Text-based Speaker
Identication on Multiparty Dialogues Using Multi-document Convo-
lutional Neural Networks. In Proceedings of the 55th Annual Meeting
of the Association for Computational Linguistics, ACL 2017, Student Re-
search Workshop, Allyson Ettinger, Spandana Gella, Matthieu Labeau,
Cecilia Ovesdotter Alm, Marine Carpuat, and Mark Dredze (Eds.). As-
sociation for Computational Linguistics, Vancouver, Canada, 49–55.
Zhao Meng, Lili Mou, and Zhi Jin. 2017. Hierarchical RNN with Static
Sentence-Level Attention for Text-Based Speaker Change Detection. In
Proceedings of the 2017 ACM on Conference on Information and Knowl-
edge Management (CIKM ’17). ACM, New York, NY, USA, 2203–2206.
Dimitri Ognibene, Davide Taibi, Udo Kruschwitz, Rodrigo Souza
Wilkens, Davinia Hernandez-Leo, Emily Theophilou, Lidia Scifo,
Rene Alejandro Lobo, Francesco Lomonaco, Sabrina Eimler, H. Ul-
rich Hoppe, and Nils Malzahn. 2021. Challenging Social Media Threats
using Collective Well-being Aware Recommendation Algorithms and
an Educational Virtual Companion. arXiv:cs.CY/2102.04211
Sree Hari Krishnan Parthasarathi, Mathew Magimai.-Doss, Daniel
Gatica-Perez, and Hervé Bourlard. 2009. Speaker Change Detection
with Privacy-Preserving Audio Cues. In Proceedings of the 2009 Interna-
tional Conference on Multimodal Interfaces (ICMI-MLMI ’09). ACM, New
York, NY, USA, 343–346.
Lance Ramshaw and Mitch Marcus. 1995. Text Chunking using
Transformation-Based Learning. In Third Workshop on Very Large
Corpora. Association for Computational Linguistics, Cambridge, MA,
USA, 82–94.
George Sanchez. 2019. Sentence Boundary Detection in Legal Text. In
Proceedings of the Natural Legal Language Processing Workshop 2019.
Association for Computational Linguistics, Minneapolis, Minnesota,
Iulian V. Serban and Joelle Pineau. 2015. Text-Based Speaker Iden-
tication For Multi-Participant Open-Domain Dialogue Systems. In
Machine Learning for Spoken Language Understanding and Interaction,
NIPS 2015 Workshop. Montreal, Canada. http://slunips2015.wixsite.
Hye Jeong Song, Hong Ki Kim, Jong Dae Kim, Chan Young Park, and
Yu Seop Kim. 2019. Inter-sentence segmentation of YouTube subtitles
using Long-Short Term Memory (LSTM). Applied Sciences (Switzerland)
9, 7 (2019).
Mark Stevenson and Robert Gaizauskas. 2000. Experiments on sen-
tence boundary detection. In Proceedings of the sixth conference on
Applied natural language processing -. Association for Computational
Linguistics, Morristown, NJ, USA, 84–89.
Ottokar Tilk and Tanel Alumäe. 2015. LSTM for punctuation restora-
tion in speech transcripts. In INTERSPEECH 2015, 16th Annual Con-
ference of the International Speech Communication Association. ISCA,
Dresden, Germany, 683–687.
Ottokar Tilk and Tanel Alumäe. 2016. Bidirectional Recurrent Neural
Network with Attention Mechanism for Punctuation Restoration. In
Interspeech 2016, 17th Annual Conference of the International Speech
Communication Association, Nelson Morgan (Ed.). ISCA, San Francisco,
CA, USA, 3047–3051. 1517
Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to
the CoNLL-2003 shared task. In Proceedings of the seventh conference
on Natural language learning at HLT-NAACL 2003 -, Vol. 4. Association
for Computational Linguistics, Morristown, NJ, USA, 142–147. https:
S.E. Tranter and D.A. Reynolds. 2006. An overview of automatic
speaker diarization systems. IEEE Transactions on Audio, Speech and
Language Processing 14, 5 (sep 2006), 1557–1565.
K. Xu, L. Xie, and K. Yao. 2016. Investigating LSTM for punctuation
prediction. In 2016 10th International Symposium on Chinese Spoken
Language Processing (ISCSLP). IEEE, Tianjin, China, 1–5.
... These tools can help identify claims worth checking, find repeats of claims that have already been checked or even assist in the verification process directly (Nakov et al., 2021). Most such tools rely on text as input and require the text to be split into sentences (Donabauer et al., 2021). ...
... Du et al. (Du et al., 2019) present a transformer-based approach to the problem, but they assume partially punctuated text and wordcasing information. Recently, it was shown that a simple fine-tuned BERT model was able to improve on the state of the art on fully unpunctuated case-folded input data (Donabauer et al., 2021). ...
... The system architecture we use is adopted from our previous work that achieved state-of-the-art performance on a very similar task (Donabauer et al., 2021). That architecture demonstrated the suitability of a BERT-based token classification approach for sentence end prediction in the context of improving text processing pipelines for fact-checking. ...
Conference Paper
Full-text available
This paper describes our approach (UR-mSBD) to address the shared task on Sentence End and Punctuation Prediction in NLG Text (SEPP-NLG) organised as part of SwissText 2021. We participated in Subtask 1 (fully un-punctuated sentences-full stop detection) and submitted a run for every featured language (English, German, French, Italian). Our submissions are based on pre-trained BERT models that have been fine-tuned to the task at hand. We had recently demonstrated, that such an approach achieves state-of-the-art performance when identifying end-of-sentence markers on automatically transcribed texts. The difference to that work is that here we use language-specific BERT models for each featured language. By framing the problem as a binary tagging task using the outlined architecture we are able to achieve competitive results on the official test set across all languages, with Recall , Precision, F1 ranging between 0.91 and 0.96 which makes us joint winners for Recall in two of the languages. The official baselines are beaten by large margins.
Full-text available
Recently, with the development of Speech to Text, which converts voice to text, and machine translation, technologies for simultaneously translating the captions of video into other languages have been developed. Using this, YouTube, a video-sharing site, provides captions in many languages. Currently, the automatic caption system extracts voice data when uploading a video and provides a subtitle file converted into text. This method creates subtitles suitable for the running time. However, when extracting subtitles from video using Speech to Text, it is impossible to accurately translate the sentence because all sentences are generated without periods. Since the generated subtitles are separated by time units rather than sentence units, and are translated, it is very difficult to understand the translation result as a whole. In this paper, we propose a method to divide text into sentences and generate period marks to improve the accuracy of automatic translation of English subtitles. For this study, we use the 27,826 sentence subtitles provided by Stanford University’s courses as data. Since this lecture video provides complete sentence caption data, it can be used as training data by transforming the subtitles into general YouTube-like caption data. We build a model with the training data using the LSTM-RNN (Long-Short Term Memory – Recurrent Neural Networks) and predict the position of the period mark, resulting in prediction accuracy of 70.84%. Our research will provide people with more accurate translations of subtitles. In addition, we expect that language barriers in online education will be more easily broken by achieving more accurate translations of numerous video lectures in English.
Conference Paper
Full-text available
We propose a convolutional neural network model for text-based speaker identification on multiparty dialogues extracted from the TV show, Friends. While most previous works on this task rely heavily on acoustic features, our approach attempts to identify speakers in dialogues using their speech patterns as captured by transcriptions to the TV show. It has been shown that different individual speakers exhibit distinct idiolectal styles. Several convolutional neural network models are developed to discriminate between differing speech patterns. Our results confirm the promise of text-based approaches, with the best performing model showing an accuracy improvement of over 6% upon the baseline CNN model.
Conference Paper
Full-text available
Traditional speaker change detection in dialogues is typically based on audio input. In some scenarios, however, researchers can only obtain text, and do not have access to raw audio signals. Moreover, with the increasing need of deep semantic processing, text-based dialogue understanding is attracting more attention in the community. These raise the problem of text-based speaker change detection. In this paper, we formulate the task as a matching problem of utterances before and after a certain decision point; we propose a hierarchical recurrent neural network (RNN) with static sentence-level attention. Our model comprises three main components: a sentence encoder with a long short term memory (LSTM)-based RNN, a context encoder with another LSTM-RNN, and a static sentence-level attention mechanism, which allows rich information interaction. Experimental results show that neural networks consistently achieve better performance than feature-based approaches, and that our attention-based model significantly outperforms non-attention neural networks.
Conference Paper
Full-text available
The output of automatic speech recognition systems is generally an unpunctuated stream of words which is hard to process for both humans and machines. We present a two-stage recurrent neural network based model using long short-term memory units to restore punctuation in speech transcripts. In the first stage, textual features are learned on a large text corpus. The second stage combines textual features with pause durations and adapts the model to speech domain. Our approach reduces the number of punctuation errors by up to 16.9% when compared to a decision tree that combines hidden-event language model posteriors with inter-word pause information, having largest improvements in period restoration.