Incremental TextRank - Automatic Keyword Extraction for Text Streams

Rui Portocarrero Sarmento

, M

ario Cordeiro

, Pavel Brazdil

and Jo

ao Gama

LIAAD-INESC TEC, Portugal

PRODEI - Faculty of Engineering, University of Porto, Portugal

Keywords:

Automatic Keyword Extraction, Incremental PageRank, Data Streams, Text Mining, Incremental TextRank.

Abstract:

Text Mining and NLP techniques are a hot topic nowadays. Researchers thrive to develop new and faster

algorithms to cope with larger amounts of data. Particularly, text data analysis has been increasing in interest

due to the growth of social networks media. Given this, the development of new algorithms and/or the upgrade

of existing ones is now a crucial task to deal with text mining problems under this new scenario. In this paper,

we present an update to TextRank, a well-known implementation used to do automatic keyword extraction

from text, adapted to deal with streams of text. In addition, we present results for this implementation and

compare them with the batch version. Major improvements are lowest computation times for the processing of

the same text data, in a streaming environment, both in sliding window and incremental setups. The speedups

obtained in the experimental results are signiﬁcant. Therefore the approach was considered valid and useful

to the research community.

1 INTRODUCTION

Automatic keyword extraction from text is an essen-

tial area of research. It has been used to automatically

summarize documents. From enterprises’ document

production to the analysis of users’ social networks

posts, for example, in Twitter or Facebook, the ap-

propriate summarizing of text brings new possibilities

to better target advertising and recommendations to

users. It is also used to ﬁnd similarities between aut-

hors and improve enterprise departments workﬂow.

The majority of current systems are prepared to be

used statically. Thus, with the explosion of large-scale

document sets, researchers felt a considerable incre-

ase in the need for systems to analyze this evolving

and even more extensive datasets. The improvement

of previous algorithms or the adaptation of these sys-

tems to cope with incoming text is, therefore, impera-

tive nowadays. Although not many researchers were

concerned with the issue of incremental keyword ex-

traction, some of them are starting to implement new

algorithms for this issue (Shin et al., 2014). Incre-

mental algorithms have been developed extensively in

several areas of Data Analysis. Notably, those algo-

rithms that are graph oriented have been proving to be

faster or more efﬁcient when their incremental adap-

tations are compared to their static versions. Our con-

tribution, in this paper, is the transformation of a well-

known algorithm for keyword extraction, TextRank

(Mihalcea and Tarau, 2004). Taking into account that

this algorithm is a graph-based approach to keyword

extraction, we updated it to be used in a streaming set-

ting. We validate the results by comparing the beha-

vior and processing time with the original algorithm.

The paper is organized as follows. Section 2 presents

the related work regarding TextRank and the compo-

nents of its implementation. Section 3 introduces the

algorithm for Incremental TextRank. In Section 4 we

explain how the proposed method was tested and eva-

luated. In Section 5 we present the results and show

the effectiveness of the proposed method. Then, in

Section 6 we discuss the solution and the results. Fi-

nally, the last Section highlights the major contributi-

ons, and discuss further work to enhance the method.

2 RELATED WORK

In this section, related work regarding automatic key-

word extraction systems, particularly those based on

graph approaches is presented. Automatic keyword

extraction is a concept that relies on the process of

selection of words or phrases that, without human in-

tervention, provide a clear idea about the core area of

a document or set of documents. Those implemen-

Sarmento, R., Cordeiro, M., Brazdil, P. and Gama, J.

Incremental TextRank - Automatic Keyword Extraction for Text Streams.

DOI: 10.5220/0006639703630370

In Proceedings of the 20th International Conference on Enterprise Information Systems (ICEIS 2018), pages 363-370

ISBN: 978-989-758-298-1

 2019 by SCITEPRESS – Science and Technology Publications, Lda. All r ights reserved

363

tations can be classiﬁed into four classes according

to Siddiqi and Sharan (2015), Rule-Based Linguis-

tic approaches, Statistical approaches, Machine Le-

arning approaches, and Domain-speciﬁc approaches.

These approaches have their advantages and disad-

vantages when compared to one another. Inside each

class, the approach can also be divided into supervi-

sed and unsupervised methods according to Siddiqi

and Sharan (2015). The use of human tasks or contri-

bution in unsupervised approaches is regarded to be

minimal. Otherwise, by using supervised approaches,

researchers take into account previous annotations by

humans, and their participation in ﬁnding quality ke-

yword annotations from texts is a signiﬁcant concern

in these methods. According to Siddiqi and Sharan

(2015) survey, there are several unsupervised approa-

ches of interest:

Bracewell et al. (2005) extract noun phrases from

a document and then cluster the terms which have the

same noun term. Then, the author’s proposal consists

of the ranking of clusters, taking in account term and

noun phrase frequencies. Finally, only the Top-ranked

clusters are selected as keyphrases. Liu et al. (2009)

proposed to extract keyphrases by utilizing clustering

techniques. According to these authors, they ensure

that the document is semantically covered by these

keyphrases. Rose et al. (2010) described Rapid Auto-

matic Keyword Extraction (RAKE), a method for ex-

tracting keywords from individual documents. RAKE

is based on the observation that keywords frequently

contain multiple words, but they rarely include punc-

tuation or stop words; therefore, any words having

minimal content or information. This new method is

domain and language-independent. Gazendam et al.

(2010) describe the extraction and ranking of key-

words with a restricted vocabulary with the help of

a thesaurus. For ranking words, it uses a weighting

scheme called tf-rr which uses both the term fre-

quency and the number of thesaurus relations realized

between the thesaurus terms found in the speciﬁc do-

cument. This is an approach that does not need any

training from a reference corpus. Finally, we can also

account for two graph-based methods. Mihalcea and

Tarau (2004) proposed TextRank, a new graph-based

ranking model for graphs extracted from texts. Thus,

to rank keywords based on the co-occurrence links be-

tween words, Mihalcea et al. use the concept of Pa-

geRank between words to extract important keyphra-

ses with linked words that have higher weight in the

word graph. Litvak et al. (2011) proposed DegExt, a

graph-based, cross-lingual keyphrase extractor. The

authors proposed a graph representation, based on a

simple graph-based syntactic representation of a do-

cument, and enhanced the model by taking into ac-

count structural document features. In this work, we

choose to improve TextRank, due to its considerable

usefulness in this area, and relatively good results for

keyword extraction. Additionally, by being a graph-

based approach, it is reasonable to think that an in-

cremental approach can provide a signiﬁcant impro-

vement in running time or efﬁciency, in a streaming

context. By considering small graph updates instead

of all the graph in each update, we will try a paralle-

lism with other works in the graphs area, where rese-

archers consider only affected nodes in each update.

The improvements in graph processing efﬁciency are

those authors main contribution.

2.1 TextRank

As previously stated, Mihalcea and Tarau (2004) de-

veloped TextRank as a solution to obtain keywords

automatically from text. Figure 1 shows the workﬂow

of the original TextRank algorithm. The text corpus

processing starts by the pre-processing of the text and

the removal of stop words, numbers and punctuation.

Then, the document goes through a process of anno-

tation where remaining single words are categorized,

for example, like nouns, verbs or adjectives among

others. This method is called Part-of-Speech tagging

(POS tagging). According to the authors, only a few

of these annotated words are essential. The authors

studied which group of words delivered best results,

and they concluded that the best automatic keyphrases

were obtained with nouns and adjectives. Then, with

these ﬁltered words, a graph-based approach is used.

Each word is considered a graph node and the con-

nections of words in this directed graph is determined

by the order they appear in the text. The weight of

these links is obtained by counting the number of ti-

mes these pairs of words occur in the text corpus. The

next phase of the algorithm regards the selection of

the words of high importance. This is done with the

use of the PageRank algorithm by Page et al. (1998).

The words with high PageRank values are selected as

potential keywords. Finally, the keyphrases are obtai-

ned with a post-processing stage. This stage involves

the use of a sliding window evolving through the ini-

tial text to assess the order of words that are contained

in the keyphrases or keywords. This step takes into

account punctuation and other structural features of

the document to retrieve reasonable keyphrases. Tex-

tRank has been widely praised as a consistent method

to automatically retrieve keywords from the text. In-

clusively, it has been used in prototypes of decision

support systems as narrated by Brazdil et al. (2015).

ICEIS 2018 - 20th International Conference on Enterprise Information Systems

364

Figure 1: Original TextRank workﬂow.

Figure 2: Incremental PageRank workﬂow - Addition of a

Node.

2.2 Incremental PageRank

The reader should note that the original TextRank al-

gorithm uses the original PageRank algorithm in one

of its phases. We will show, later in this document,

that a signiﬁcant improvement of TextRank might be

possible by transforming some of its phases algo-

rithms to a streaming approach. Revisions of these

algorithms were already developed as we will see.

Particularly, there have been several attempts to im-

prove the original PageRank algorithm. The purpose

of several of these improvements was to adapt this al-

gorithm for streaming data. Desikan et al. (2005) pro-

vided a solution for the update of nodes’ PageRank

values in evolving graphs. The algorithm explores the

fact that the web evolves incrementally and with small

changes between updates. Thus, according to the aut-

hor’s proposal, we should focus only on the pages that

change their PageRank values due to the addition or

removal of new pages/nodes and the correspondent

change in the affected nodes by this change. Desikan

et al. (2005) continued with this work and stress the

need to ﬁnd the nodes that are affected by the changes

in the graph and ﬁnd three categories of nodes:

1. Nodes that are directly affected by changes in the

graph, e.g., the nodes that are outgoing nodes of

added or removed nodes;

2. Nodes that are indirectly affected. For example,

the neighbors of the nodes in 1, and consequently

their outgoing neighbors also;

3. The nodes not affected by the changes, but that

have their PageRank value updated anyway, due

to a simple graph scale update.

Figure 2 shows a small network with these types

of nodes. In this ﬁgure, changes in the graph are re-

presented by added node 7. In this ﬁgure, we can also

see nodes of type 1 colored dark gray, nodes of type

2 colored light gray, and nodes of type 3 with white

color.

2.3 Space Saving Top-K

The problem of ﬁnding the most frequent items, in

our case, keywords, in a data stream S of size N is the

problem of discovering the elements e

whose relative

frequency f

is higher than a user-speciﬁed support

φN, with 0 ≤ φ ≤ 1 (Gama, 2010). Given the space re-

quirements that exact algorithms addressing this pro-

blem would need (Charikar et al., 2002), several algo-

rithms have been proposed to ﬁnd the top-K frequent

elements, being roughly classiﬁed into counter-based

and sketch-based (Metwally et al., 2005). Counter-

based techniques keep counters for each element in

the monitored set, which is usually a much smaller

than the entire set of elements. When an element is

identiﬁed as not currently being monitored, various

algorithms take different actions to adapt the moni-

tored set accordingly. Sketch-based techniques pro-

vide less rigid guarantees, but they do not monitor

a subset of elements, providing frequency estimators

for the entire set. Simple counter-based algorithms,

such as Sticky Sampling and Lossy Counting, were

proposed in (Manku and Motwani, 2002), which pro-

cess the stream in compressed size. They have the

disadvantage of keeping a lot of irrelevant counters.

Frequent (Demaine et al., 2002) keeps only k coun-

ters for monitoring k elements, incrementing each ele-

ment counter when it is observed, and decrementing

all counters when an unmonitored element is obser-

ved. Zeroed-counted elements are replaced by new

unmonitored elements. This strategy is similar to the

one applied by Space-Saving (Metwally et al., 2005),

which gives guarantees for the top-m most frequent

elements. This was the algorithm we selected to ex-

tract the most frequent keywords in the stream of text.

3 TextRank FOR TEXT STREAMS

In this section, we explain our proposal for two ver-

sions of an incremental TextRank. We propose a

Window-based version and a more complex incre-

mental version. In Figure 1 we showed the sequence

of operations the text goes through, in the original

TextRank. Although some parts of it are eminently

Incremental TextRank - Automatic Keyword Extraction for Text Streams

365

evolving as, for example, the graph construction from

the text words, other parts are not prepared for strea-

ming or incremental approach.

3.1 Window-based Streaming TextRank

The ﬁrst version we developed was based on the con-

cept of using a sliding window to iterate through the

stream of data. Thus, with this version, the cur-

rent snapshot processing is unrelated to the previous

snapshots processing. Therefore, each chunk of data

from the stream is treated independently from snaps-

hot to snapshot. In each iteration of the Window-

based TextRank, we only process the current piece of

data, we pre-process this data, and we do the standard

word selection by considering only nouns and adjecti-

ves. Additionally, when we do the graph construction

we just build it considering the selected words for the

current chunk of data. Thus, the next phase of the

algorithm, i.e., the use of PageRank algorithm to de-

tect strong words in the text also only processes the

words in the current snapshot graph. Again, regar-

ding post-processing, it is done considering the words

in the text available just for the current snapshot. As

previously stated, the top-K phase of this Window-

based streaming version, uses the top-K space saving

implementation that is based on the concept of land-

mark windows. Thus, the top-K list of keywords is the

only information that transits from snapshot to snaps-

hot processing for the streaming data.

3.2 Incremental TextRank

Figure 3 shows our proposal for the upgraded incre-

mental TextRank. We have changed the workﬂow of

the original algorithm to be able to cope with evol-

ving text streams. The input of a new stream of text

starts to be pre-processed in the same way that was

used for the original algorithm. At this stage, as

in the original algorithm, we ﬁlter only the desired

words as, for example, nouns and adjectives. Then,

the workﬂow proceeds with the word graph update,

considering the previous graph from previous upda-

tes, incrementally, with edge weight updates for the

repetition of words sequences, and also the addition

of new nodes/words if new words happen to exist in

the evolving text stream. Then, we use incremental

PageRank algorithm to retrieve the words that have

a higher importance in the text stream. These key-

words are applied yet to a ﬁnal post-processing stage

of the stream to obtain a reasonable combination of

keywords as in the original algorithm. Finally, as

we do not need to accumulate unnecessary keywords

from previous stream updates, and as we want the

proposed system to be sensitive to the occurrence of

new keywords in the text stream, we process the ke-

ywords in a stage where we maintain and update a

top-K list of keywords with the space-saving algo-

rithm previously described. The top-K algorithm ap-

plication, based on landmark window, enables an efﬁ-

cient approach for large-scale text datasets. It focuses

on the relevant keywords and discards less extracted

keywords through time or keywords extracted from

earlier snapshots. The alternative option for sliding

windows (Gama, 2010) would not be appropriate for

the top-K approach, since it may remove relevant key-

words from the extracted top-K keywords. Those key-

words might yet be included in the top-K keyword list

we wish to maintain. In our scenario, top-K represen-

tation of text data streams implies knowing the K key-

words of the simulated data stream from the database

of publications and received after the post-processing

stage of the streaming TextRank algorithms.

4 CASE STUDY AND

EVALUATION

In this section, we present the method to analyze both

window-based and incremental TextRank, regarding

its comparison with the original TextRank. We intro-

duce the reader to the data used for the tests. Finally,

we describe the methodology to perform the compa-

rison tests. We also state and introduce the metrics

used to compare the algorithms regarding obtained

keywords or keyphrases.

4.1 Description of the Data

In this case study, for algorithms efﬁciency measure-

ments, we selected a dataset publicly available for re-

search purposes by Corney et al. (2016). This dataset

has a high amount of information including Reuters

news and also several articles extracted from Blogs.

The data contains articles titles, content, type of pu-

blication (which might include news or Blog posts),

the source institution and also date and time of publis-

hing. The high quality and high organization of this

structured data make it a good source for text mining

or NLP tasks. We selected all Reuters news, from

the 1

to the 30

of September 2015. This corre-

sponds to 400 news articles. Regarding the used text,

we could choose between the news Titles or the news

content to analyze. We choose the news content in

all our studies. For qualitative measurements, we se-

lected R&D publications from Czech Republic resear-

chers. The dataset was publicly available in RDICCR

ICEIS 2018 - 20th International Conference on Enterprise Information Systems

366

Figure 3: Incremental TextRank workﬂow.

(2015) when accessed online, in 2015. This is a com-

plete source of information, and there was informa-

tion for dozens of years starting from around 1985

until 2015. This dataset provides annotated keywords

for each publication. During 2015, while this dataset

was publicly available, we exported the data for se-

veral years, we selected conference papers and book

chapters only. Therefore, the number of publications

was reduced from around 25000 to 5110 publications.

4.2 Methodology

The Window-based TextRank and the incremental

TextRank algorithm were evaluated in an incremental

setup, for both the developed versions. Additionally,

both algorithms were tested and evaluated regarding

Processing Efﬁciency and also quality.

4.2.1 Processing Efﬁciency

Results for both streaming TextRank versions were

obtained from several increments of text news, com-

pared with the original batch algorithm, using the ori-

ginal TextRank (now on called Batch). The propo-

sed TextRank versions, the Window-based implemen-

tation (now on called Window-based) and the incre-

mental (now on called incremental) were tested re-

garding efﬁciency in an incremental setting conﬁgu-

ration. The original batch results served as a baseline.

In this setup, 6 snapshots of news were considered. In

total, the ﬁrst 20 days of September 2015, correspon-

ding to 300 news articles, were passed as input. The

snapshots were built by aggregating publications on a

daily basis, by using the timestamps available in the

dataset. In the Batch algorithm, for every snapshot,

the keywords were extracted having all daily news

since the day 1 to the current day as input. For the

incremental algorithm, in the ﬁrst snapshot it receives

only 15 days of news articles as input. In the follo-

wing snapshots, the algorithm only receives the set of

publications text added to the corpus in that particular

day snapshot (incremental). The empirical evaluation

performed consisted mainly of comparing run times

of each increment (duration of each increment and cu-

mulative execution time) for all versions of TextRank.

Additionally, the size of the corpus (number of unique

words), and the total number of extracted keywords

for each snapshot was also registered. Note that the

batch algorithm will always need to extract keywords

for all the text, while it is expected that the Window-

based version only works with the current snapshot

data, and not regarding previous snapshots informa-

tion except the top-K list of keywords that needs to

be updated between snapshots. Additionally, the in-

cremental algorithm only performs calculations and

keyword extraction for the current snapshot, and par-

ticularly for the words affected. Nonetheless, as pre-

viously explained, the incremental versions takes into

account graph and corpus information from previous

snapshots. In the end, an analysis of the total speed-

up ratio obtained in each of the steps is added.

4.2.2 Qualitative Comparison

Regarding qualitative results, we used the Czech Re-

public R&D Dataset with annotated keywords. For

this purpose, we selected only articles in proceedings,

for 2013. Then, we aggregated the publications by

month, in 12 snapshots. For this amount of publi-

cations, we selected the abstracts for our automatic

keyword extraction tasks. On average, the abstracts

have 240 words. Finally, we calculated the 12 me-

asures of similarity between annotated and extracted

keywords of all the versions of TextRank. We use a

similarity measure to compare the author’s keywords

and the model results for keywords extraction.

This similarity measurement is more appropriated

to test unsupervised keyword extraction algorithms.

For example, imagine if we had the annotated key-

word “car maintenance” for a text. Additionally, we

automatically extracted the keywords “car washing”

and “car cleaning” from the text. In this situation, if

we used Precision, Recall and F measure we would

have no idea that the extraction algorithm is, in fact,

extracting concept keywords that are not that much far

from the annotated keywords. Thus, by decomposing

n-gram keywords we can detect these normal situa-

tions when we are dealing with annotated keywords.

Particularly for this example, we would have a value

of similarity of:

similarity

K,M

common

c(w

) ∗ c(w

)

2 ∗ 2

= 0.25

(1)

where common is the intersection of both groups, i.e.,

the number of words that appear in both the results

and the keywords. c(w

) and c(w

) are the total

Incremental TextRank - Automatic Keyword Extraction for Text Streams

367

number of words the author’s keywords group has and

the method (for all TextRank algorithms) provides, re-

spectively. For all the experimentation and develop-

ment, we used an Intel

 Core

i7-7700HQ CPU

@ 2.80GHz x 8 processors computer with 16 GBytes

of RAM, SSD HDD, and Ubuntu 16.04.3 LTS 64-bit

OS. Three runs per algorithm were performed. The

values presented in the following charts in this paper

are the average values of those three runs.

5 RESULTS

We present the results for all versions of the algorithm

here in this section. All algorithms are compared with

the Batch version of TextRank. The comparisons re-

gard qualitative measurements and efﬁciency.

5.1 Processing Efﬁciency

Figure 4 shows the results for the Reuters News data.

From “#Graph Nodes” and “#Words in Text” we can,

as expected, conclude that the number of TextRank

graph nodes is similar for the batch and incremen-

tal version, and variable for the Window-based ver-

sion of the algorithm. Figure 4 (#Extracted Key-

words) shows that the batch version retrieves a gro-

wing number of keywords while our streaming imple-

mentations retrieve the top-1000 keywords. Figure 4

(Elapsed Time) shows that all algorithms have vari-

able elapsed time for each of the iterations. Nonet-

heless, both streaming algorithms have a clear decre-

asing elapsed time compared to the batch algorithm

which tends to increase elapsed time as the quantity

of the data increases. Otherwise, both streaming ver-

sions tend to achieve much lower values. For exam-

ple, on iteration 4, the Window-based algorithm de-

livers a very low elapsed time, almost instantly pro-

cessing of incoming data. The incremental version

takes around 100 seconds to perform the same in-

coming data processing, and the batch version takes

approximately 900 seconds. Figure 4 (Cumulative

Time) shows a clear advantage of using incremen-

tal or Window-based versions of the algorithm regar-

ding total elapsed time/cumulative time when the data

is growing, and there is a need to analyze all data

since day 1. In the last iteration, both Window-based

and the incremental versions take around 10000 se-

conds to process all the 20 days of data. Otherwise,

if we used the original TextRank algorithm to recur-

rently process all the data and also the incoming text,

it would take almost 50000 seconds to achieve re-

sults. Figure 4 (Speedup Ratio) shows that both stre-

aming algorithms, the incremental and the Window-

based, show growing and signiﬁcant speedups when

we compare cumulative time. At the end of the

stream, i.e., the last iteration, the Window-based ap-

proach provides a speedup of more than 6 times, and

the incremental version achieves a speedup of almost

5 times the speed of processing provided by the origi-

nal TextRank algorithm.

5.2 Qualitative Comparison

Regarding qualitative comparison, the Table 1 shows

the similarity values between extracted keywords and

accumulated annotated keywords as the abstracts text

data evolves. We selected the top-1000 keywords of

the batch algorithm, regarding keyword weight. The-

refore, we compare the similarity values for these

1000 keywords obtained from the batch, Window-

based and incremental algorithms. To measure the

sensitivity to capture recent keywords we repeat the

test, but this time we measure the similarity of ex-

tracted keywords with each snapshot annotated key-

words. This way we can measure the sensibility of

all algorithms to the current keywords in the stream.

Table 2 shows these results. As we expected, in both

tables of results, the incremental version of the algo-

rithm presents higher similarity average. This is even

more pronounced in the second table as expected. We

should also note, as expected in the second table, that

the maximum similarity values for the streaming ver-

sions of the algorithms are obtained in the last snaps-

hots. This is explained by the higher sensibility of

both proposed algorithms to more recent keywords.

6 DISCUSSION

Regarding processing efﬁciency, it is clear that if the

TextRank user is dealing with streaming data, the use

of incremental or Window-based versions of the al-

gorithm is imperative. The Window-based version

presents a more pronounced speedup, and this is ex-

pected as the processed data is variable instead of in-

cremental. Still, regarding cumulative time, the incre-

mental version of the algorithm has a learning curve.

Thus, as the data arrives and the graph construction

and update happens, the algorithm will eventually

need fewer updates of graph structure in the future.

This typically results in a logarithmic type of cumula-

tive time curve. This is the result of less and less nee-

ded addition of nodes, resulting in the need to only

update the weight of edges between nodes, i.e., the

counting of words pairs. Regarding qualitative re-

sults of both streaming proposals, we would like to

note that, if the researcher using our proposed algo-

ICEIS 2018 - 20th International Conference on Enterprise Information Systems

368

0.2

0.4

0.6

0.8

1.2

1.4

1.6

·10

# Keywords

0.2

0.4

0.6

0.8

·10

# Graph Nodes

0.2

0.4

0.6

0.8

1.2

1.4

·10

# Words in Text

Speedup Ratio

0.2

0.4

0.6

0.8

·10

Elapsed Time (s)

·10

Cumulative Time (s)

batch (incremental Text) Window-based (Streaming TextRank) incremental (incremental TextRank)

Figure 4: Efﬁciency results with the Reuters News Dataset (for all algorithms).

Table 1: Qualitative measures of similarity with accumulated annotated keywords.

Algorithm Avg. Sd iter1 iter2 iter3 iter4 iter5 iter6 iter7 iter8 iter9 iter10 iter11 iter12

Batch 0.168 0.108 0.023 0.025 0.032 0.067 0.130 0.254 0.308 0.300 0.260 0.218 0.204 0.193

Window-based 0.114 0.059 0.023 0.045 0.042 0.069 0.125 0.173 0.186 0.207 0.136 0.127 0.134 0.102

Incremental 0.198 0.125 0.011 0.046 0.046 0.124 0.189 0.308 0.350 0.398 0.304 0.217 0.206 0.181

Table 2: Qualitative measures of similarity with current snapshot annotated keywords.

Algorithm Avg. Sd iter1 iter2 iter3 iter4 iter5 iter6 iter7 iter8 iter9 iter10 iter11 iter12

Batch 0.160 0.118 0.023 0.008 0.003 0.026 0.086 0.221 0.280 0.254 0.267 0.250 0.270 0.235

Window-based 0.155 0.103 0.023 0.043 0.016 0.047 0.099 0.183 0.202 0.267 0.211 0.239 0.299 0.237

Incremental 0.254 0.181 0.011 0.032 0.023 0.105 0.133 0.311 0.331 0.480 0.336 0.371 0.425 0.487

rithms is not too preoccupied with the small decre-

ase of quality of extracted keywords from the stream,

then he/she can use the Window-based version of the

algorithm for fast extraction. Nonetheless, if the re-

searcher needs assured quality of extracted keywords,

the incremental version proved to be a better solution

when compared to the batch and the Window-based

approach for keyword extraction. Additionally, the

effect of increasing K in both these top-K approaches

would intuitively increase the quality results as it is

expected that, in Table 1, as the annotated keywords

amount increase we continue to test only 1000 key-

words extracted with the algorithms. Accordingly, in

Table 1, the quality values decrease due to this charac-

teristic, approximately after iteration 8, until the end

of the stream, and for all algorithms.

7 CONCLUSIONS

In this paper, we introduce a Window-based and an

incremental version of the TextRank system. The

Window-based version proved to be similar in qua-

lity and much faster than the original TextRank algo-

rithm, in a streaming context. Through the careful

transformation of the original algorithm to an incre-

mental version, we achieve a more efﬁcient version

and with better quality. Additionally, we introduce

a top-K feature in our improvement, which allows

storing the keywords that are more relevant. As fu-

ture work, we would like to compare this incremental

TextRank system with other systems prepared for text

streams. Additionally, we would like to use this strea-

ming version of the algorithm in a prototype prepared

to extract visual information from afﬁnity networks

of authors, with text mined from their document pro-

duction.

ACKNOWLEDGMENTS

This work is ﬁnanced by the ERDF – European Re-

gional Development Fund through the Operational

Programme for Competitiveness and Internationali-

sation - COMPETE 2020 Programme within project

“POCI-01-0145-FEDER-006961”, and by National

Funds through the FCT – Fundac¸

ao para a Ci

encia e

a Tecnologia (Portuguese Foundation for Science and

Technology) as part of project UID/EEA/50014/2013.

Rui Portocarrero Sarmento also gratefully acknow-

Incremental TextRank - Automatic Keyword Extraction for Text Streams

369

ledges funding from FCT (Portuguese Foundation

for Science and Technology) through a PhD grant

(SFRH/BD/119108/2016)

REFERENCES

Bracewell, D. B., Ren, F., and Kuriowa, S. (2005). Multi-

lingual single document keyword extraction for infor-

mation retrieval. In 2005 International Conference on

Natural Language Processing and Knowledge Engi-

neering, pages 517–522.

Brazdil, P., Trigo, L., Cordeiro, J., Sarmento, R., and Vali-

zadeh, M. (2015). Afﬁnity mining of documents sets

via network analysis, keywords and summaries. Oslo

Studies in Language, 7(1).

Charikar, M., Chen, K., and Farach-Colton, M. (2002). Fin-

ding frequent items in data streams. In Proceedings of

the 29th International Colloquium on Automata, Lan-

guages and Programming, ICALP ’02, pages 693–

703, London, UK, UK. Springer-Verlag.

Corney, D., Albakour, D., Martinez, M., and Moussa, S.

(2016). What do a million news articles look like?

In Proceedings of the First International Workshop

on Recent Trends in News Information Retrieval co-

located with 38th European Conference on Informa-

tion Retrieval (ECIR 2016), Padua, Italy, March 20,

2016., pages 42–47.

Demaine, E. D., L

opez-Ortiz, A., and Munro, J. I. (2002).

Frequency estimation of internet packet streams with

limited space. In Algorithms-ESA 2002, pages 348–

360. Springer.

Desikan, P., Pathak, N., Srivastava, J., and Kumar, V.

(2005). Incremental page rank computation on evol-

ving graphs. In Special Interest Tracks and Posters

of the 14th International Conference on World Wide

Web, WWW ’05, pages 1094–1095, New York, NY,

USA. ACM.

Gama, J. (2010). Knowledge Discovery from Data Streams.

Chapman & Hall/CRC, 1st edition.

Gazendam, L., Wartena, C., and Brussee, R. (2010). The-

saurus based term ranking for keyword extraction. In

Database and Expert Systems Applications, DEXA,

International Workshops, Bilbao, Spain, August 30 -

September 3, 2010, pages 49–53.

Litvak, M., Last, M., Aizenman, H., Gobits, I., and Kan-

del, A. (2011). DegExt — A Language-Independent

Graph-Based Keyphrase Extractor, pages 121–130.

Springer Berlin Heidelberg, Berlin, Heidelberg.

Liu, Z., Li, P., Zheng, Y., and Sun, M. (2009). Cluste-

ring to ﬁnd exemplar terms for keyphrase extraction.

In Proceedings of the 2009 Conference on Empirical

Methods in Natural Language Processing: Volume 1 -

Volume 1, EMNLP ’09, pages 257–266, Stroudsburg,

PA, USA. Association for Computational Linguistics.

Manku, G. S. and Motwani, R. (2002). Approximate fre-

quency counts over data streams. In Proceedings of

the 28th International Conference on Very Large Data

Bases.

Metwally, A., Agrawal, D., and El Abbadi, A. (2005). Ef-

ﬁcient computation of frequent and top-k elements in

data streams. In Proceedings of the 10th Internatio-

nal Conference on Database Theory, ICDT’05, pages

398–412, Berlin, Heidelberg. Springer-Verlag.

Mihalcea, R. and Tarau, P. (2004). TextRank: Bringing or-

der into texts. In Proceedings of EMNLP-04and the

2004 Conference on Empirical Methods in Natural

Language Processing.

Page, L., Brin, S., Motwani, R., and Winograd, T. (1998).

The pagerank citation ranking: Bringing order to the

web. In Proceedings of the 7th International World

Wide Web Conference, pages 161–172, Brisbane, Au-

stralia.

RDICCR (2015). Published data from r&d is of the czech

republic - research, development and innovation coun-

cil of the czech republic (rdiccr). http://www.isvav.cz/.

Accessed: 2015-09-30.

Rose, S., Engel, D., Cramer, N., and Cowley, W. (2010).

Automatic keyword extraction from individual docu-

ments. In Berry, M. W. and Kogan, J., editors, Text

Mining. Applications and Theory, pages 1–20. John

Wiley and Sons, Ltd.

Shin, Y., Ryo, C. Y., and Park, J. (2014). Automatic ex-

traction of persistent topics from social text streams.

World Wide Web, 17(6):1395–1420.

Siddiqi, S. and Sharan, A. (2015). Article: Keyword

and keyphrase extraction techniques: A literature re-

view. International Journal of Computer Applicati-

ons, 109(2):18–23. Full text available.

ICEIS 2018 - 20th International Conference on Enterprise Information Systems

370