=Paper=
{{Paper
|id=Vol-2526/paper2
|storemode=property
|title=Automatic Slide Generation for Scientific Papers
|pdfUrl=https://ceur-ws.org/Vol-2526/paper2.pdf
|volume=Vol-2526
|authors=Athar Sefid,Jian Wu,Prasenjit Mitra,C. Lee Giles
|dblpUrl=https://dblp.org/rec/conf/kcap/Sefid0MG19
}}
==Automatic Slide Generation for Scientific Papers==
<pdf width="1500px">https://ceur-ws.org/Vol-2526/paper2.pdf</pdf>
<pre>
                     Automatic Slide Generation for Scientific Papers
                                   Athar Sefid                                                                 Jian Wu
                       Pennsylvania State University                                                  Old Dominion University
                            azs5955@psu.edu                                                               jwu@cs.odu.edu

                               Prasenjit Mitra                                                              C. Lee Giles
                       Pennsylvania State University                                               Pennsylvania State University
                           pmitra@ist.psu.edu                                                           giles@ist.psu.edu

ABSTRACT                                                                              document. This method has been applied in slide generation tasks,
We describe our approach for automatically generating presenta-                       e.g., [6]. This is because (1) the selected sentences have correct
tion slides for scientific papers using deep neural networks. Such                    grammatical structure and (2) they are scientifically/technically
slides can help authors have a starting point for their slide genera-                 consistent with the original article.
tion process. Extractive summarization techniques are applied to                         The two major components of extractive summarization are
rank and select important sentences from the original document.                       sentence scoring and sentence selection. The goal is to keep
Previous work identified important sentences based only on a lim-                     salient sentences while excluding redundant information. Sentence
ited number of features that were extracted from the position and                     scoring is usually converted to a regression problem. The scores
structure of sentences in the paper. Our method extends previous                      depend on features extracted from the current sentence and its
work by (1) extracting a more comprehensive list of surface fea-                      contextual sentences.
tures, (2) considering semantic or meaning of the sentence, and (3)
using context around the current sentence to rank the sentences.
Once, the sentences are ranked, salient sentences are selected using
Integer Linear Programming (ILP). Our results show the efficacy of
our model for summarization and the slide generation task.

CCS CONCEPTS
• Computing methodologies → Neural networks; Classifica-
tion and regression trees; • Information systems → Document
topic models; Retrieval models and ranking.

KEYWORDS
Natural Language Processing, Text Mining, Slide Generation, Sum-
marization, Deep Learning

1    INTRODUCTION
Scientific results are usually disseminated by publications and re-
search presentations. The latter has been a convention for almost
all scientific domains since it provides an efficient way for others
to grasp the major contribution of a paper. Currently, PowerPoint
automates slide formating and thus significantly reduces the effort
to make professional and useful slides. However, PowerPoint can-
not automatically generate slides based on the content of research
                                                                                                  Figure 1: Slide Generation flow chart
papers. Automatically generating slides would not only save the
presenters’ time in preparing presentations, but it also provides
a way to summarize paper and it can also be used to generate
summaries for papers that do not have publicly available slides.                        The main contributions of this paper are the following.
   The two major approaches for summarization are abstractive and                       (1) To the best of our knowledge, we are the first to use Deep
extractive methods. Abstractive approaches summarize text using                             Neural models to encode sentences and its context as new
words not necessarily appearing in the original document. Extrac-                           features in sentence ranking to build slides.
tive summarization focuses on identifying important constituents
usually at the sentence level and connecting them to generate a                         (2) We combine regression with integer linear programming
text snippet, which is usually significantly shorter than the original                      (ILP) to select salient sentences. Then, noun phrases are
Copyright ©2019 for this paper by its authors. Use permitted under Creative Commons         extracted from the selected sentences to build the first-level
License Attribution 4.0 International (CC BY 4.0).                                          bullet points in slides.
SciKnow ’19, November 19–22, 2019, Marina del Rey, CA                                                                                         Sefid and Wu, et al.


Here, we focus on the textual components. This approach can later        bullet points and to determine their place in the slide. However,
be enhanced with visual effects, e.g., figures and tables, in order to   their model was tested on a limited set of only 175 paper-slide pairs.
make better more complex slides.                                            Compared with previous works, we use a combination of feature
                                                                         based and deep neural network methods to score sentences. Then,
2     RELATED WORK                                                       ILP is applied to build the summary. To transform the summary
There has been much work on text summarization using both ab-            into a slide format, first-level bullet points are generated using key
stractive and extractive methods. However, automatic generation          phrases extracted from the selected sentences.
of slides, which can be seen as a specific form of summarization of
scholarly papers, has not been well studied.                             3     MODEL
                                                                         The summarization system consists of 2 steps. The first is to train a
2.1     Summarization                                                    model to calculate scores used to identify important sentences. The
Unsupervised Models. Identifying important sentences for generat-        second is to extract salient sentences under constraints. The model
ing a limited length summary can be formalized as an optimization        is trained on a corpus of aligned pairs of papers and slides. In the
problem which can be NP-hard. Maximum Marginal Relevance                 training process, we investigate Convolutional Neural Networks
[2] has been used to heuristically select salient sentences while        (CNN), Gated Recurrent Unit (GRU), and Long Short Term Memory
keeping redundancy with the selected summary at the lowest point.        (LSTM) using pre-trained word embeddings (WE) to represent the
Sentence selection could be converted to the Knapsack problem            semantic feature of sentences. This semantic representation is en-
by maximizing the total scores of the selected sentences given the       hanced by combining other surface features of the current sentence
length limit and can be solved by Integer Linear Programming (ILP).      and its contextual features. The high-level architecture is shown in
ILP is an optimization method with linear constraints and objec-         Figure 1.
tive functions [13]. Graph-based summarization methods model a
document as a graph, in which vertices are sentences and edges
are the similarity of vertices. One example is TextRank [14] which
extends Google’s PageRank. TextRank selects important sentences
that have a high degree of overlapping tokens with other sentences.
LexRank [5], a stochastic graph-based method, measures the sen-
tence importance by eigenvector centrality of the sentences in the
graph.
Feature-Based Models. Many traditional extractive summarization
approaches use feature-based models such as term frequency, an
effective feature. Other features include the length and position of
sentences [18]. The Support Vector Regressor was applied to score
                                                                         Figure 2: Architecture to predict salience score of a sentence
and extract salient sentences [9]. Others have modeled a document
                                                                         Si . S j (j , i) are contextual sentences. Three types of embed-
as a sequence of sentences and label each sentence with 1 or 0 in
                                                                         dings are combined as an input to MLP to output the score of
which 1 indicates that the corresponding sentence belongs to the
                                                                         sentence Si . The CNN could be replaced with LSTM or GRU.
summary [19] and a 0 not. Hidden Markov models (HMM) have
been adopted to solve such sequence labeling problems [4].
Deep Neural Network Models. State-of-the-art models now use deep         3.1      Sentence Labeling
neural networks which benefit from both the semantics of the
                                                                         In this section, we describe how to generate the ground truth by
sentences and their structure [7, 8]. Zhou et al. used multiple layers
                                                                         assigning scores to sentences in a paper given a paper-slide pair.
of bidirectional Gated Recurrent Units (GRU) to jointly learn the
                                                                         An academic paper D can be represented as a sequence of sentences
scoring process and the selection of sentences [23]. This work
                                                                         D = {s 1 , s 2 , . . . , sn }. Each sentence is represented by the set of its
applies bi-GRU to make sentence embeddings and then combines
                                                                         unigrams and bigrams, such as si = {t 1 , t 2 , . . . , tk , t 1t 2 , t 2t 3 , . . . , tk −1tk }
sentence vectors with another bi-GRU layer to make document
                                                                         where ti is the i-th token in the sentence. The corresponding pre-
level representations.
                                                                         sentation of the document D can be represented as a sequence of
                                                                         slides P = {p1 , p2 , . . . , pm }. The salience score of each sentence
2.2     Slide Generation                                                 is determined by the highest Jaccard similarity of unigrams and
Previously, a method was proposed to generate slides from docu-          bigrams between a sentence and all slides in the corresponding
ments by identifying important topics and then adding their related      presentation.
sentences to the summary [12]. A tool called PPSGen applies a
Support Vector Regressor for sentence ranking and then ILP to                               ∀si ∈ D, scoresi = max Jaccard(si , p j )                          (1)
select important sentences with a set of sophisticated constraints                                                   p j ∈P
[6]. Another method generates slides by phrases extracted from             ÑJaccard similarity of si and p j sets is defined as Jaccard(si , p j ) =
                                                                         The
papers [22]. The model learns the saliency of each phrase and the        si p j
                                                                           Ð .
hierarchical relationship between a pair of phrases to make the          si p j
Automatic Slide Generation for Scientific Papers                                                     SciKnow ’19, November 19–22, 2019, Marina del Rey, CA


   Having the labeled sentences with their scores, we can start             previous activations instead of replacing the entire activation like
training different models to learn the scoring functions.                   a vanilla RNN.
                                                                               LSTM: The LSTM tracks long term dependencies via input (i t ),
3.2     Sentence Ranking                                                    forget (ft ), and output (ot ) gates; the input gate regulates how much
The goal of the sentence ranking module is to train a model f (si |ϕ)       of the new cell state to keep, the forget gate forgets a portion of
that minimizes the Mean Square Error (MSE) defined as below:                existing memory and the output gate sends important information
                                                                            in the cell state to the next layers of the network.
                             ∀si , 0 ≤ f (si |ϕ) ≤ 1                  (2)      GRU: The GRU operates using a reset gate and an update gate.
                                                                            The reset gate decides how much of the past information to forget,
                                 Õ                                          and the update gate helps the model to determine how much of the
                      MSE =           (f (si |ϕ) − scoresi )2         (3)   past information needs to be passed to the future.
                                 si                                            The input to the recurrent units are pre-trained word vectors
in which ϕ is sentence embedding combined with syntactic and                and the unit outputs are combined to build the sentence semantic
contextual features.                                                        embeddings.
   The sentence ranking model consists of three modules. The first
models the meaning of a sentence using a deep neural network
                                                                            3.2.2 Context Embedding. The context of a sentence is defined as
that embeds the sentence itself into a vector. The second evalu-
                                                                            sentences before and after the current sentence. Intuitively, the
ates the capability of the sentence to summarize its contextual
                                                                            sentence that includes an abstract of its context is more important
sentences. The last extracts a variety of surface features from the
                                                                            and should have a higher chance to be selected for the summary of
sentence structure and its position in the paper. The results from
                                                                            the paper. For instance, some paragraphs are ended by a sentence
all of the three modules are combined to build the input for the
                                                                            that summarizes the whole paragraph. This sentence probably has
final Multi-Layer Perceptron (MLP) with two hidden layers which
                                                                            a high word overlap with other sentences of the paragraph. Contex-
acts as a regressor to predict the final score for a sentence. Figure
                                                                            tual sentence relations have been used to model attention weights
2 depicts the embedding layers for the construction of sentence
                                                                            of sentences and then to identify important sentences [17]. The
representation vectors.
                                                                            context information is embedded as the following vectors.
3.2.1 Semantic Embedding. In this layer, the semantic of a sentence
is encoded into a vector. We compare the performance of both
convolutional and recurrent networks in order to choose the best                               cos(Vsi , Vsi −1 )                cos(Vsi , Vsi +1 )
system for semantic embedding.                                                     pr ev Si = cos(Vsi , Vsi −2 ) , nex t Si = cos(Vsi , Vsi +2 )
                                                                                                                                                    
                                                                                                                                                           (7)
   Convolutional Neural Network (CNN): In this layer, the se-                                  cos(Vs , Vs )                     cos(Vs , Vs )
                                                                                                      i     i −3                         i     i +3 
mantic representation of a sentence si is generated by concatenating
                                                                                                                                  

embeddings of bi-grams. The intuition is that bi-grams preserve
                                                                            3.2.3 Syntactic Embedding. The models suggested in the above
the sequential information of the text. Convolutional layers and
                                                                            focus on representing the meaning of the sentences and their con-
element-wise max-pooling on top of the bigram embeddings cap-
                                                                            text into vectors. Although semantic features are strong metrics for
ture important patterns in the sentence that match similar patterns
                                                                            identification of salient sentences, surface features are also proved
in the reference slides.
                                                                            to be important [6]. The surface features we extracted are tabulated
   The bigram representations are calculated by concatenating pre-
                                                                            in Table 1.
trained word embeddings of tokens in the bigram.
                                                                               The “section” feature is a one-hot vector to represent the section
                            biдram j = [v j , v j+1 ]                 (4)   that sentence belongs to. Certain sections, such as conclusion, can
  In which v j and v j+1 stand for the current and the next word            be more important to indicate sentence importance compared with
vectors in sentence si .                                                    other sections such as acknowledgment. Table 2 shows the average
                                                                            scores of the sentences in a set of pre-defined sections calculated
                  Vbiдr am j = tanh(biдram j ∗ Wc + b)                (5)   from the labeled paper-slide pairs. The table indicates that Intro-
                                                                            duction and Conclusion sections are relatively more important than
   Where filter Wc ∈ R2|v j |∗ |v j | is applied on a vector of two words
                                                                            the other sections.
to make new features. The activation function is the Hyperbolic
                                                                               The second feature is the relative position of the sentence in the
Tangent and b is bias term.
                                                                            section in units of sentences.
                         Vsi =        max         Vbiдr am j          (6)      The other features #NP, #VP, #S, and “height” are obtained from
                                 0<j < |s i |−1
                                                                            the parse tree of the sentence. #NP shows the number of noun
where the max function is an element-wise max pooling and Vsi is            phrases in the sentence, #VP counts verb phrases, #S indicates the
the semantic embedding of sentence si [3].                                  number of sub-sentences and “height” is the height of the sub-tree.
   Recurrent Neural Network (RNN): The convolutional layers                 We used Stanford CoreNLP that outputs tree structures consisting
above can model the sequential patterns up to 2 consecutive tokens.         of constituent noun and verb phrases.
However, RNNs are able to capture the dependencies of tokens in                The stop_word%, #tokens, and “length” are ratios of the stop
long sentences. Both LSTM and GRU have been widely used to                  words among tokens, number of tokens and number of characters
remember long term dependencies by keeping memory/state from                in the sentences, respectively.
SciKnow ’19, November 19–22, 2019, Marina del Rey, CA                                                                          Sefid and Wu, et al.


   The “paper_title_similatiry” is the Jaccard similarity between        document frequency are not informative enough to be added as a
the tokens in the title of the paper and the sentence. The “sec-         bullet point such as “the model”.
tion_title_similarity” is the Jaccard similarity of the section title       Once a set of candidate noun phrases is selected, they are sorted
and the current sentence.                                                based on the average frequency of their tokens. The noun phrases
   The “average_IDF” feature is the average of IDF scores of tokens      and the sentences containing them are added to the slides in order of
of the sentence and “average_TF” takes the average of the TF scores.     their frequency until all of the selected sentences by the model are
                                                                         covered in the slides. The slide generation process is demonstrated
3.3     Sentence Selection                                               in Algorithm 1.
The model above generates a salience score for each sentence in             In order to set the titles of the slides, the section of the first
the paper. Next, we compared a greedy method and the ILP for             sentence in the slide is determined and then the heading of the
sentence selection.                                                      section is borrowed from the paper as the title of the slide. The
                                                                         heading is truncated to the first 5 tokens. We constrain a maximum
3.3.1 Greedy Method. In this intuitive approach, sentences are           of 4 sentences per slide. If a topic has more than 4 related sentences,
ranked based on their salience scores and then are added to the          the slide is split into two distinct ones.
summary while the bigram overlap is less a threshold θ and the
summary size does not exceed a maximum limit. The threshold                Data: sentences
parameter is tuned to be θ = 0.7 as higher or lower amounts result         Result: slides
in worse performance.                                                      ranked_sents = rank(sentences);
3.3.2 Integer Linear Programming (ILP). ILP is an optimization             selected_sents = ILP(ranked_sents);
problem with linear constraints and objective functions. ILP re-           NPs = {} ;      // dictionary of NPs and a list of its
stricts some of the variables to be integers and it is NP-complete.          sentences
We adopted the IBM CPLEX Optimizer1 , which has been used to                for s ∈ selected_sents do
solve ILP problems efficiently.                                                for np ∈ noun_phrases(s) do
   We aim to maximize the following objective function:                            NPs[np].add(s)
                        Õ                                                      end
                   max      li Wi x i , subject to:             (8)        end
                            i ∈N s
                  Õ                                                        freq_rank(NPs);             // rank noun phrases based on
                        li x i < maxLen,     ∀i, x i ∈ {0, 1}      (9)       frequency of tokens
                    i                                                      slides=[];
   where Wi is the sentence scores generated by the deep model             while selected_sents , ∅ do
suggested in the paper. The x i is a binary variable showing whether           for np, sents ∈ N Ps do
sentence i is selected for the summary or not. The li is the number                splides[np] = [];
of characters in sentence i and Eq. (9) sets a constraint to limit the             for s ∈ sents do
size of the summary to a maximum length maxLen. Existence of li                        if s ∈ selected_sents then
in the Eq. (8) helps to penalize short sentences.                                           slides[np].add(s);
                                                                                            selectet_sents \ {s};
3.4     Slide Generation
                                                                                       end
Typical presentation slides include a limited number of bullet points
                                                                                   end
which is often the first-level of the slide structure. These bullet
                                                                               end
points are usually noun phrases or shortened versions of sentences.
Some slides may contain second-level bullet points for further break-      end
downs. Less than 8% of the content of the presentations in the             distribute_sents(slides) ;           // allow a maximum of 4
ground truth corpus is covered in third-level bullets. Therefore, we         sentences per slide
generate slides with up to 2 bullet levels. Slide titles on average                   Algorithm 1: Bullet Point Generation.
contains 4 words and either Level 1 or Level 2 bullets contains on
average 8 words. Each slide is on average made of 36 words in 5
bullets and each level-1 bullet includes 2 second-level bullets.         4 EXPERIMENTS
    The deep neural model first ranks the sentences and then ILP         4.1 Data
helps to select salient sentences by considering the length of the
                                                                         The dataset is a collection of 1200 scientific papers and their corre-
sentences and length limit for the generated presentation. These
                                                                         sponding presentation slides made by their authors. This dataset
selected sentences are considered as the second-level bullets. The
                                                                         is adopted from a previous project called PPSGen [6]. The dataset
first-level bullets are the noun phrases extracted from the sentences.
                                                                         contains conference proceedings in computer science in PDF for-
Some of the noun phrases are removed from the candidate set if
                                                                         mat. To extract the header metadata and content of the paper from
(1) They have more than 10 tokens.(2) They have only 1 token. (3)
                                                                         PDF files, GROBID [11] is used and information is transformed
Their DF is higher than 10. Noun phrases with a high degree of
                                                                         into an XML file with the TEI (Text Encoding Initiative) format, a
1 https://www.ibm.com/products/ilog-cplex-optimization-studio            comprehensive encoding standard for machine-readable texts.
Automatic Slide Generation for Scientific Papers                                                                 SciKnow ’19, November 19–22, 2019, Marina del Rey, CA


                                                                    Table 1: Sentence surface features

                 Feature                         Type               Description
                 section                         one-hot vector     Section (Intro, Related, Model, Result, Conclusion, Acknowledge) of the sentence
                 position                        index              Position of the sentence in the section
                 height                          integer            Height of the Parse tree of the sentence
                 #NP                             integer            Number of noun phrases in the sentence
                 #VP                             integer            Number of verb phrases in the sentence
                 #S                              integer            Number of sub-sentences in the sentence
                 length                          integer            Number of characters in sentence
                 stop_word%                      float              Ratio of the stop words in the sentence
                 #tokens                         integer            Number of tokens in the sentence
                 paper_title_similarity          float              Jaccard similarity of the sentence to the title of the paper
                 section_title_similarity        float              Jaccard similarity of the sentence to the title of the section
                 average_IDF                     float              Average of the IDF scores of tokens
                 average_TF                      float              Average of the frequency of tokens in the paper


    The slides are either in PDF or PPT format. The PPT files are                           where д stands for an n-gram in the reference summary Re f erence,
first converted to PDFs. Then, the text in presentations is extracted                    n is its length, Count match (д) is the number of shared n-grams be-
by the LibreOffice pdftotext tool.                                                       tween the candidate Candidate and the reference summary and
                                                                                         Count(д) is the number of n-grams in reference or candidate sum-
4.2     Model Configuration                                                              mary. The F-measure score(F ) is harmonic mean of Pr and Re.
Stanford CoreNLP 3.9.2 is used to tokenize and lemmatize sentences                          ROUGE-L [10] is similarly defined based on the Longest Common
to the constituent words and also to build the parse trees. The                          Sub-sequence (LCS) and ROUGE-W is Weighted Longest Common
word embeddings are initialized with GloVe 1.2 [16] 50-dimensional                       Sub-sequence developed to differentiate LCSs of different spatial
vectors trained on Wikipedia. The pre-trained GloVe vectors contain                      relations. ROUGE-W gives more weight to the sequences that are
400,000 words and cover 98.9% of our model vocabulary. The rest                          spatially closer to each other [10].
of the word embeddings are randomly initialized using a Gaussian                            Empirically, ROUGE-2 is more suitable for single document sum-
distribution. Drop out layers with probability p = 0.3 are added to                      maries compared with ROUGE-1 [10, 15].
the recurrent model to randomly remove some of the neurons from
the network in order to prevent over-fitting [20]. Adam optimizer                        4.4     Result
with L2-norm regularization and mini-batches of size 100 are used                        We compared our model with the following baselines listed as
to train the model. The learning rate was set to 0.0001 and the model                    follows:
is trained for 50 epochs. We truncate each sentence containing more
than 200 words.                                                                              (1) AvgTFIDF: This model ranks sentences based on the Avg(T F )×
                                                                                                 Ave(IDF ) named AvgTFIDF scores of the tokens in sentence.
4.3     Evaluation Metric                                                                        The model greedily adds the sentences to the summary until
The Recall-Oriented Understudy for Gisting Evaluation (ROUGE)                                    it hits the size limit.
[10] is a standard metric for automatic evaluation of machine-                               (2) TextRank: TextRank is a graph-based algorithm where ver-
generated summaries. ROUGE-N is an n-gram overlap between                                        tices are sentences and the edge weights are the cosine simi-
a candidate summary and a reference summary. The recall-oriented                                 larity of the sentence embeddings [1].
score (Re) and the precision-oriented score (Pr ) are computed as                            (3) MEAD: MEAD is a publicly available platform for multi-
follows:                                                                                         lingual summarization. The platform implements a combi-
                                                                                                 nation of position-based and centroid-based algorithms to
                               д∈Re f e r e nc e Count match (д)
                     Í
                    Re =                                                         (10)            score the sentences [21].
                                   д∈Re f e r e nc e Count (д)
                               Í

                                                                                            The results for the ROUGE scores are shown in Table 3. Our
                             д∈Re f e r e nc e Count match (д)
                           Í
                    Pr =                                                         (11)    model with the CNN neural structure and ILP sentence selector
                                 д∈C and id at e Count (д)
                               Í
                                                                                         achieved the best F scores in terms of ROUGE-2, ROUGE-L, and
                                                                                         ROUGE-W compared with the baselines.
                                                                                            Intuitively, LSTMs should have shown a better performance in
Table 2: Score distribution of sentences in the ground truth corpus.                     representing sentences with long dependencies. However, they are
                                                                                         prone to over-fitting since there are many parameters to be tuned.
                         Related                                       Acknowle-            ILP always outperforms the simple greedy algorithm. The reason
  Section       Intro                 Model     Result   Conclusion
                         Work                                          dgement
                                                                                         is that it is able to recognize if adding multiple shorter and less
  Avg Score     44.29     38.92        39.09    42.42       51.02        31.18           important sentences will result in better ROUGE scores compared
  #sentences    25494     18316        8924     6955        7036          722
                                                                                         with having fewer long and more important sentences.
SciKnow ’19, November 19–22, 2019, Marina del Rey, CA                                                                                                              Sefid and Wu, et al.


                                                                         Table 3: ROUGE Scores

                                                           ROUGE-1                     ROUGE-2                     ROUGE-L                    ROUGE-W
                          model         selector
                                                    Pr%      Re%       F%       Pr%      Re%       F%      Pr%       Re%       F%      Pr%     Re%      F%
                          AvgTFIDF      Greedy     41.94     30.02    32.06    11.50     7.89     8.46     21.37    15.77     16.93    8.45     1.56    2.46
                          AvgTFIDF        ILP      41.45     30.53    32.24    11.31     8.00     8.48     23.17    17.36     18.51    9.21     1.72    2.71
                          TextRank      Greedy     41.04     31.87    32.79    11.09     8.28     8.60     22.61    17.75     18.49    8.77     1.71    2.66
                          MEAD          Greedy     44.58    35.00     35.86    12.49     9.31     9.65     22.84    18.19     18.84    8.90     1.76    2.73
                          GRU           Greedy     40.53     33.63    35.54    11.18     8.86     8.91     20.09    16.80     17.01    7.77     1.66    2.53
                          GRU             ILP      40.28     34.17    33.73    11.06     8.96     8.92     22.55    18.99     19.16    8.84     1.89    2.88
                          LSTM          Greedy     40.04     34.26    33.61    11.60     9.44     9.37     22.20    18.87     18.93    8.71     1.88    2.86
                          LSTM            ILP      40.26     34.94    34.08    11.58     9.46     9.59     22.58    19.46     19.40    8.83     1.94    2.94
                          CNN           Greedy     41.92     33.93    34.28    12.33     9.50     9.71     21.65    17.67     18.10    8.49     1.76    2.73
                          CNN             ILP      41.80     34.74    34.74    12.19     9.70     9.79    23.45     19.46    19.79     9.24    1.94    2.98


5    CONCLUSION AND FUTURE WORK                                                                 [11] Patrice Lopez. 2009. GROBID: Combining Automatic Bibliographic Data Recogni-
                                                                                                     tion and Term Extraction for Scholarship Publications. In Proceedings of the 13th
Although summarization has been well studied, the task of auto-                                      European Conference on Research and Advanced Technology for Digital Libraries
matically generation of slides is relatively new. This paper uses                                    (ECDL’09). 473–474.
                                                                                                [12] Utiyama Masao and Hasida Kôiti. 1999. Automatic slide presentation from
neural network models to encode syntax, semantics, and context                                       semantically annotated documents. In Proceedings of the Workshop on Coreference
of sentences as features to automatically generate slide content.                                    and its Applications. Association for Computational Linguistics, 25–30.
A multilayer perceptron on top of sentence embeddings acts as a                                 [13] Ryan McDonald. 2007. A study of global inference algorithms in multi-document
                                                                                                     summarization. In European Conference on Information Retrieval. 557–564.
regressor that ranks the sentences. Our model then selects salient                              [14] Rada Mihalcea and Paul Tarau. 2004. Textrank: Bringing order into text. In
sentences with the constraint of a limit on summary length. Our                                      Proceedings of the 2004 conference on empirical methods in natural language pro-
method shows higher ROUGE scores compared with the introduced                                        cessing.
                                                                                                [15] Karolina Owczarzak, John M Conroy, Hoa Trang Dang, and Ani Nenkova. 2012.
baselines which we interpret to mean that the slides generated have                                  An assessment of the accuracy of automatic evaluation in summarization. In Pro-
a high overlap with manually made slides.                                                            ceedings of Workshop on Evaluation Metrics and System Comparison for Automatic
                                                                                                     Summarization. Association for Computational Linguistics, 1–9.
   The software implementation of the model is available at GitHub:                             [16] Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove:
https://github.com/atharsefid/Automatic-Slide-Generation.                                            Global vectors for word representation. In Proceedings of the 2014 conference on
                                                                                                     empirical methods in natural language processing (EMNLP). 1532–1543.
                                                                                                [17] Pengjie Ren, Zhumin Chen, Zhaochun Ren, Furu Wei, Jun Ma, and Maarten de
6    ACKNOWLEDGEMENTS                                                                                Rijke. 2017. Leveraging contextual sentence relations for extractive summariza-
                                                                                                     tion using a neural attention model. In Proceedings of the 40th International SIGIR
The National Science Foundation is gratefully acknowledged for                                       Conference on Research and Development in Information Retrieval. 95–104.
                                                                                                [18] Pengjie Ren, Furu Wei, CHEN Zhumin, MA Jun, and Ming Zhou. 2016. A
partial support.                                                                                     redundancy-aware sentence regression framework for extractive summarization.
                                                                                                     In Proceedings of COLING 2016, the 26th International Conference on Computational
                                                                                                     Linguistics: Technical Papers. 33–43.
REFERENCES                                                                                      [19] Dou Shen, Jian-Tao Sun, Hua Li, Qiang Yang, and Zheng Chen. 2007. Document
 [1] Federico Barrios, Federico López, Luis Argerich, and Rosa Wachenchauzer. 2016.                  summarization using conditional random fields.. In IJCAI, Vol. 7. 2862–2867.
     Variations of the Similarity Function of TextRank for Automated Summarization.             [20] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan
     CoRR abs/1602.03606 (2016). arXiv:1602.03606 http://arxiv.org/abs/1602.03606                    Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from
 [2] Jaime G Carbonell and Jade Goldstein. 1998. The Use of MMR and Diversity-Based                  overfitting. The Journal of Machine Learning Research 15, 1 (2014), 1929–1958.
     Reranking for Reodering Documents and Producing Summaries. (1998).                         [21] DR Timothy, T Allison, S Blair-goldensohn, J Blitzer, A Elebi, S Dimitrov, E
 [3] Ronan Collobert, Jason Weston, Léon Bottou, Michael Karlen, Koray Kavukcuoglu,                  Drabek, A Hakim, W Lam, D Liu, et al. 2004. MEAD a platform for multidocu-
     and Pavel Kuksa. 2011. Natural language processing (almost) from scratch.                       ment multilingual text summarization. In International Conference on Language
     Journal of machine learning research 12, Aug (2011), 2493–2537.                                 Resources and Evaluation.
 [4] John M Conroy and Dianne P O’leary. 2001. Text summarization via hidden                    [22] Sida Wang, Xiaojun Wan, and Shikang Du. 2017. Phrase-based presentation
     markov models. In Proceedings of the 24th annual international ACM SIGIR con-                   slides generation for academic papers. In 31st AAAI Conference.
     ference on Research and development in information retrieval. ACM, 406–407.                [23] Qingyu Zhou, Nan Yang, Furu Wei, Shaohan Huang, Ming Zhou, and Tiejun
 [5] Günes Erkan and Dragomir R Radev. 2004. Lexrank: Graph-based lexical centrality                 Zhao. 2018. Neural document summarization by jointly learning to score and
     as salience in text summarization. Journal of artificial intelligence research 22               select sentences. arXiv preprint arXiv:1807.02305 (2018).
     (2004), 457–479.
 [6] Yue Hu and Xiaojun Wan. 2015. PPSGen: Learning-based presentation slides gen-
     eration for academic papers. IEEE transactions on knowledge and data engineering
     27, 4 (2015), 1085–1097.
 [7] Hayato Kobayashi, Masaki Noguchi, and Taichi Yatsuka. 2015. Summarization
     based on embedding distributions. In Proceedings of the 2015 conference on empir-
     ical methods in natural language processing. 1984–1989.
 [8] Chen Li, Xian Qian, and Yang Liu. 2013. Using supervised bigram-based ILP
     for extractive summarization. In Proceedings of the 51st Annual Meeting of the
     Association for Computational Linguistics (Volume 1: Long Papers), Vol. 1. 1004–
     1013.
 [9] Sujian Li, You Ouyang, Wei Wang, and Bin Sun. 2007. Multi-document summa-
     rization using support vector regression. In Proceedings of DUC. Citeseer.
[10] Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries.
     Text Summarization Branches Out (2004).

</pre>