SEMANTICS GRAPH MINING FOR TOPIC DISCOVERY AND WORD ASSOCIATIONS

International Journal of Data Mining & Knowledge Management Process (IJDKP)
Vol.11, No.2/3/4, July 2021
DOI:10.5121/ijdkp.2021.11401 1
SEMANTICS GRAPH MINING FOR TOPIC
DISCOVERY AND WORD ASSOCIATIONS
Alex Romanova
Melenar, LLC, McLean, VA, USA
ABSTRACT
Big Data creates many challenges for data mining experts, in particular in getting meanings of text data.
It is beneficial for text mining to build a bridge between word embedding process and graph capacity to
connect the dots and represent complex correlations between entities. In this study we examine processes
of building a semantic graph model to determine word associations and discover document topics. We
introduce a novel Word2Vec2Graph model that is built on top of Word2Vec word embedding model. We
demonstrate how this model can be used to analyze long documents, get unexpected word associations and
uncover document topics. To validate topic discovery method we transfer words to vectors and vectors to
images and use CNN deep learning image classification.
KEYWORDS
Graph Mining, Semantics, Topics Discovery, Word Associations, Deep Learning, Transfer Learning, CNN
Image Classification.
1. INTRODUCTION
Big Data creates many challenges for data experts, in particular on text data mining: nowadays
data volumes are growing exponentially. For organizations that are daily getting huge amounts of
unstructured text data, analyzing this data is too difficult and time consuming task to do manually.
Automation of topic discovery and word associations can solve document analysis problems as
well as support other NLP problems such as search, text mining, and documents summarization.
For topic discovery most common traditional approaches are topic modelings and topic
classifications. Topic classifications as supervised machine learning techniques require topic
knowledge before starting the analysis. Topic modelings as unsupervised machine learning
techniques such as K-means clustering, Latent Semantic Indexing, Latent Dirichlet Allocation
can infer patterns without defining topic tags on training data beforehand [1]. In this study we
will introduce method of finding document topics through semantic graph clusters.
Word embedding methods such as Word2Vec [2] are capable of capturing context of a word in a
document, semantic and syntactic similarity, and therefore solving many complicated NLP
problems such as finding semantically related pairs of words. Based on Word2Vecsemantic
similarity between two words is generally taken as cosine similarity of word vectors. However
word associations, unlike cosine similarities, are expected to be asymmetric [3]. In semantic
graph model that we introduce in this study we will be able to find not just directed pairs of
associated words but also lines of associated words of any size.

Vol.11, No.2/3/4, July 2021
2
Word embedding models are conceptually based on sequential, logical thinking but they are
missing capabilities to ‘connect the dots’, i.e. determine connections between entities.
Understanding word relationships within documents is very important for topic discovery process
and graph techniques can help to feel this gap.
In this article we will introduce a semantic graph model Word2Vec2Graph. This model combines
word embedding and graph methods to gain the benefits of both. Based on this model we will
analyze text documents, find unexpected word association lines and uncover document topics.
Document topics observed as semantic graph clusters will not only uncover keyword sets,but they
will also show relationships between words within topics.
By looking at semantic graph mining techniques from geometrical view, we can see the following
benefits:
• Traditional text mining techniques are based on bags of words with no dependencies between
the words. This can be considered as zero-dimensional data mining.
• Semantic graph pairs of words can be used to find word neighbors, paths between the words or
lines of associated words. It can be considered as one-dimensional data mining.
• Graph clusters determine community relationships within semantic groups and can be
considered as multi-dimensional data mining. As Aristotle said: “The whole is greater than the
sum of its parts”.
Figure 1. Finding text topics through a Word2Vec2Graph model and
validating topics via CNN classification
As a language for building the Word2Vec2Graph model we use Spark — a powerful open source
analytic engine [4] with libraries for SQL (DataFrames), graphs (GraphFrames), machine
learning, and NLP [1]. Until recently there were no single processing framework that was able to
solve several very different analytical problems in one place. Spark is the first framework for data
mining and graph mining right out of the box.
Finding text document topics within semantic graph can be done using various community
detection algorithms. In this paper we will use a simple community detection method — graph
connected components — subgraphs where any two nodes are connected by paths, and no nodes
are connected to nodes from other subgraphs.

Vol.11, No.2/3/4, July 2021
3
To validate topic correctness by method independent on semantic graph topic discovery, we will
transform word vectors to images and use Convolutional Neural Network (CNN) image
classification technique. Please see Figure 1 that shows the data flow diagram for the process of
finding and validating document topics.
In this paper we propose a new, graph-based methodology, which has the following original
contributions:
• Introduced a novel Word2Vec2Graph model that combines analytic thinking and holistic
thinking functionalities in semantic graph.
• Established an ability of the Word2Vec2Graph model to analyze long documents, find
unexpected associated word lines, and discover document topics.
• Proposed CNN transfer learning image classification method for topic
validation.In the pages that follow, we will show:
• Studies related to semantic graph building methods and algorithms of semantics graph mining.
• Process of building Word2Vec2Graph model by training Word2Vec model, getting collocated
pairs of words from data corpus and building a graph based on word pairs as edges and vector
cosine similarities as edge weights.
• Topic discovery method by calculating connected components and using top PageRank words
within components as topic class words.
• Topic correctness validation method by transfer learning CNN image classification.
2. RELATED WORK
There are various methods of building semantics graphs. Some of these methods are based on
more traditional deep syntactic text analysis like RDF triples (subject–predicate–object) [5], other
methods are based on unsupervised key phrase extractions and identifying statistically significant
words [6] or on structuring asynchronous text streams [7].
Words association techniques are useful to extract word meaning from text data. Is some studies
word associations are used to identify themes within sets of texts and are calculated based on
statistical significancy of words within text subsets [8]. In other studies word associations are
revealed through Word2Vec semantic similarity [9].
Recently because of enormous progress of word embedding methods such as Word2Vec [2] some
methods of building semantic graphs are based on word embeddings. For example,
WordGraph2Vec method [10] is a semantic graph built on top of Word2Vec model that enriches
text by adding target words for a specific context word in a sliding window.
Our Word2Vec2Graph model is similar to the WordGraph2Vec model [10] as in both models
semantic graphs are built on top of Word2Vec. However in our semantic graph model we use
pairs of words located next to each other in the document and mapping these words to vectors
through Word2Vec model. For pairs of words we are calculating cosine similarities between
words and building a direct graph on pairs of words as edges and vector cosine similarities as
edge weights. This allows us to find not only highly connected groups of words but also to find
unexpected words associations.

Vol.11, No.2/3/4, July 2021
4
In recent years, some studies are trying to integrate semantic graph structures with topic
modeling. These models apply different methods of combining text with semantics graphs.
Some studies integrate topic mining and time synchronization into a unified model [7] or combine
semantic graphs with the textual information for topic modeling to estimate the probabilities of
topics for documents [11].
Other studies are looking for topics through semantic graphs built on semantic relatedness
between entities and concepts based on Wikipedia metadata[12]. For community detection of
modern networks diverse methods are being used for sparse or dense graphs [13]. In this paper to
find topics we are concentrated on sparse graphs and using a simple community detection method
graph connected components.
CNN image classification techniques are very useful for image classification [14, 15]. In this
study we used CNN image classification as independent method for topic validation.
Transformation vectors to images was done with Gramian Angular Fields (GAF) methods,
similar to techniques described in studies [16, 17].
3. METHODS
For this study we implemented the following methods:
• Retrained Word2Vec model on data corpus of interest.
• Built direct semantic graph using collocated word pairs as graph edges.
• Determined associated word lines.
• Created and analyzed graph clusters.
• Converted embedded vectors to GAF images and used CNN image classification to
validategraph clustering accuracy.
For building and mining semantic graph we used Spark framework. Spark code is described in
several posts of our technical blog [18].
3.1. Build Semantic Graph
In this paper we introduce a novel Word2Vec2Graph model —semantic graph model that
combines word embeddings and graph functionalities.Word2Vec2Graphmodel will give us new
insights like top words in text file — pageRank, word topics — connected components, word
neighbors — 'find' function.
To build Word2Vec2Graph model and find document topics we will use Spark framework:
Machine Learning and DataFrame libraries for Word2Vec model training and GraphFrame
library for graphs.Spark Scala code is described in several posts of our technical blog [18].
3.1.1. Train Word2Vec Model
There are different practices of using Word2Vec model for word embedding: using pre-trained
model or training model on domain specific corpus. Based on our observations, for topic finding
and word association computation Word2Vec models trained on domain specific corpus work
much better than pre-trained generic models. This observation corresponds with a study [19] that
shows that domain specific training corpuses work with less ambiguity than general corpuses for
these problems.

Vol.11, No.2/3/4, July 2021
5
To prove the difference, we trained two Word2Vec models. The first model was trained on generic
corpus (News) and the second model was trained on combination of generic corpus and data about
Stress extracted from Wikipedia (News + Wiki). In Table 1 you can see the differences of
synonyms to words ‘Stress’ and ‘Rain’. As the word ‘Stress’ belongs to Stress corpus, the
synonyms on these models are very different, but for a neutral word ‘Rain’ synonyms taken from
these models are very similar.
Table 1. Examples of synonyms based on word2vec model corpuses: 'News' is word2vec model trained
ongeneric corpus and 'News + Wiki' is word2vec model trained on combination of generic corpus and
'Stress'related corpus.
Stress Rain
News News + Wiki News News + Wiki
risk obesity snow snow
adversely adverse winds rains
clots systemic rains winds
anxiety averse fog mph
traumatic risk inches storm
persistent detect storm storms
problems infection gusts inches
One of the goals of semantic graph mining is to understand meanings of entity relationships.
Based on these circumstances, to build Word2Vec2Graph model we will train Word2Vec models
on domain specific datacorpora. Spark code for training and analyzing Word2Vec model can be
found in our blog post [20].
3.1.2. Build Word2Vec2Graph Model
To build Word2Vec2Graph model we will do the following steps:
• We will look at pairs of words located next to each other in the document. To extract such
pairsof words {word1, word2} we will use Spark Ngram(2) function.
• For words from word pairs we will get word vectors from Word2Vec model, i.e. for {word1,
word2} pair we will map word1 to [word1, vector1] and word2 to [word2, vector2].
• Then we will calculate cosine similarities for wordpairs, i.e. for {word1, word2} pair we
willcalculate cosine between [vector1] and [vector2].
• Finally, we will build a direct graph on word pairs with words as nodes, word pairs as edges
and cosine similarities as edge weights.
Spark code for steps of building Word2Vec2Graph model can be found in our technical blog post
[21].

Vol.11, No.2/3/4, July 2021
6
3.2. Semantic Graph Mining
By comparing semantic graph mining methods with traditional text mining from geometrical
view, we can see that traditional ‘bag of words’ method represents zero dimensional text mining,
graph connections represent one dimensional text mining, and graph clusters represent multi-
dimensional text mining.
In this study we will demonstrate:
• One-dimensional text mining techniques such as word neighbors, unexpected word
associations and lines between the words.
• Multi-dimensional text mining techniques such graph clustering.
3.2.1. Lines between the Words
On Word2Vec2Graph model finding word neighbors can be done through Spark GraphFrame
motif ‘find’ function. Using ‘find’ function for finding word neighbors is better understandable
than more traditional way of finding word neighbors via Word2Vec model based on word
synonyms.
graph.find("(a) - [ab] -> (b)”)
Elegance of this style can be illustrated when looking for second degree neighbors, i.e. ’friends of
friends’. Coding it via Word2Vec model would require self join word synonyms.
def foaf(graph: GraphFrame, node: String) = {
graph.find("(a) - [ab] -> (b); (b) - [bc] -> (c)").
filter($”a.id"=!=$"c.id").
filter($”a.id"===node))}
The Spark GraphFrame motif ‘find’ method is conceptually similar to {subject - predicate ->
object} and is better understandable than code for self-joining tabular data [22].
In addition to finding word neighbors this method is applicable to analysis of word to word
associations. One of the ways to examine word to word connections is using Shortest Path
GraphFrame function. The method we propose in this study will find word paths in direct graph
using number of words as parameters.
Here is how to get single words between 'startWord' and‘endWord' :
val path=graph.
find("(a) - [] -> (b); (b) - [] -> (c)”).
filter($”a.id”===startWord &&$”c.id”===endWord)
and here is how to get any two words between 'startWord' and 'endWord':
val path=graph.
find("(a) - [] -> (b); (b) - [] -> (c); (c) - [] -> (d)”).
filter($”a.id”===startWord &&$”c.id”===endWord)
To find a predefined number ofwords (wordCount) in [startWord, endWord] line:
def formLine(wordCount:Int): String = {

Vol.11, No.2/3/4, July 2021
7
var line=new String
for (i <- 1 to wordCount-1)
line+=“(x"+lit(i)+")-[]-
>(x"+lit(i+1)+");"
line.substring(0, line.length() - 1)}
Examples of formLine function:
formLine(3)
(x1)-[]->(x2);(x2)-[]->(x3)
formLine(4)
(x1)-[]->(x2);(x2)-[]->(x3);(x3)-[]->(x4)
formLine(5)
(x1)-[]->(x2);(x2)-[]->(x3);(x3)-[]->(x4);(x4)-[]->(x5)
To get a predefined number ofwords (wordCount) in [startWord, endWord] line:
def findForm(graph: GraphFrame, startWord:
String,endWord: String, wordCount: Int):
DataFrame ={ val
path=graph.find(formLine(wordCount))}
Detail explanation and Spark codeis published in our tech blog [23].
3.2.2. Graph Clusters
Finding graph clusters is very challenging process. In social network graphs this is called
‘community detection’. In this study we are using the simplest method — graph connected
components. Graph connected components are subgraphs where every two nodes have path
between them and none of the nodes is related to nodes outside of this subgraph.
In dense graph connected component practice the largest connected component usually contains
large amount of graph nodes and therefore connected component method is not useful for
community detection in dense graphs. On the contrary, community detection via this method
works well for sparse graphs. Based on this we will propose two ways to come up with semantic
graph clusters through connected components:
• Create sparse graphs based on small range threshold of word-vector cosine similarities.
• Create graphs with no limitations on word-vector cosine similarities but calculate
connectedcomponents on small range edge weights.
In Experiments section of this paper we will demonstrate the first way for text topic discovery
and the second way to observe unexpected word associations. The first method follows standard
Spark GraphFrame Connected Component function and Spark code for second method is
published in our blog post [24].
3.3. CNN Image Classification
Word2Vec2Graph model is built on top of word embedding model where word vectors are
transformed to graphs. Word vectors can also be transformed to images and CNN image
classification can be used as independent validation method.

Vol.11, No.2/3/4, July 2021
8
3.3.1. Transform Vectors to Images
As a method of vector to image translation in this study we used Gramian Angular Field (GAF)
— a polar coordinate transformation based techniques [16, 17].This transformation method works
well for images classification and data visualization. We were inspired by practice suggested on
fast.ai forum by Ignacio Oguiza as a method of encoding time series as images and using fast.ai
library for CNN image classifications.
3.3.2. Train CNN Image Classification Model
For this study we used fast.ai CNN transfer learning image classification. To deal with
comparatively small set of training data, instead of training the model from scratch, we followed
ResNet-50 transfer learning: loaded the results of model trained on images from the ImageNet
database and fine tuned it with data of interest [25, 26].Python code for transforming vectors to
GAF images and fine tuning ResNet-50 is described in fast.ai forum [27].
4. EXPERIMENTS
4.1. Source Data
For this study we used two domain-specific data corpuses: one data corpus about Creativity and
Aha Moments and another data corpus about Psychoanalysis.
"Psychoanalysis" data was used to recognize unexpected word associations and "Creativity and
Aha Moments” data corpus was used for text topic discovery and for calculating word association
lines.
4.2. Word Associations
Word2Vec2Graph technique of finding text topics is conceptually similar to Free Association
that is a practice in psychoanalytic therapy. We will show some examples that prove this analogy.
As a text file we will use text data about Psychoanalysis extracted from multiple articles in
Wikipedia.
In Free Association practice, a therapist asks a person in therapy to freely share thoughts, words,
and anything else that comes to mind. In traditional free association, a person in therapy is
encouraged to verbalize or write all thoughts that come to mind. Free association is not a linear
thought pattern. Rather, a person might produce an incoherent stream of words, such as dog, red,
mother, and scoot. They may also jump randomly from one memory or emotion to another. The
idea is that free association reveals associations and connections that might otherwise go
uncovered. People in therapy may then reveal repressed memories and emotions [28].
Word associations plays another important role in text mining: lines of associated words show
meanings of word to word connections to better understand themes of the document.
One of the problems of word association calculation through embedding space like Word2Vec is
that word associations are expected to be asymmetric but cosine similarities are symmetric [9].

Vol.11, No.2/3/4, July 2021
9
To solve this problem we will build Word2Vec2Graph model as direct graph on pairs of words
located next to each other in the text.
Figure 2. Free associations between words: semantics subgraphs on pairs of words with a) high cosine
similarities; b) low cosine similarities.
Adjusted word pairs in the stream of words are expected to be similar and if so in
Word2Vec2Graph model these pairs would have high cosine similarities. To uncover unexpected
associated word pairs we will look at graph edges with low weights.
4.2.1. Low Weight Graph Clusters
In ‘Graph Clusters’ part of ‘Semantic Graph Mining’ section of this paper we propose two ways
of building graph clusters. To find unexpected word associations we will use a method of building
graph with no weight limitations and calculating connected components with weight threshold
parameters. Then we will compare high edge weight clusters with low edge weight clusters.
Figure 2 shows examples of graph clusters built on connected components (a) with cosine
similarities higher than 0.6 and (b) with low cosine similarities, between 0.1 and 0.2. Words in
hight cosine similarity cluster look semantically closer to each other than words in low cosine
similarity cluster.
4.2.2. Associated Word Lines
Another Word2Vec2Graph method of finding word associations is described in ‘Lines between
the Words’ part of ‘Semantic Graph Mining’ section. As the data corpus for experiment we will
use text data about ‘Creativity and Aha Moments’.
Looking at word associations in Figure 2 we can see then word pairs with high cosine similarities
are well known similar words - they are called "synonyms" in Word2Vec model. Onthe contrary,
pairs of words with low cosine similarity represent unexpected associations and therefore they
are more interesting to look at.
To find word associations in text data we will do the following:
• Train Word2Vec model on ‘Creativity and Aha Moments’ data corpus.
• Extract collocated pairs of words.

Vol.11, No.2/3/4, July 2021
10
• Map these words to pairs of vectors.
• From pairs of vectors select pairs with cosine similarities less than 0.33 and build direct graph.
• Calculate word association lines between words.
Here are examples of word association movements between the words ‘brain’ and ‘insight’ that
demonstrate how to get from ‘brain’ to ‘insight’ based on ‘Creativity and Aha Moments’ data
corpus. One word between:
brain -> right -> insight
brain -> activity -> insight
brain -> moments -> insight
Two words between the words ‘brain’ and ‘insight’:
brain -> require -> spontaneous -> insight
brain -> thought -> called -> insight
brain -> dominant -> problem -> insight
brain -> response -> associated -> insight
Three words between the words ‘brain’ and ‘insight’:
brain -> require -> neural-> activity -> insight
brain -> dominant -> problem-> called -> insight
brain -> functions -> creator-> ideas -> insight
brain -> thinking -> likely-> solve -> insight
4.3. Uncover and Validate Document Topics
Finding text document topics within semantic graph can be done using various community
detection algorithms. In this paper to detect document topics we will examine units of semantic
graph that are separated from each other — graph connected components. Within each of these
components we will find the most highly connected word using graph PageRank function.
To validate topic correctness by method independent on semantic graph topic discovery, we will
transform word vectors to images and use Convolutional Neural Network image classification
technique. Please see Figure 1 that shows the data flow diagram for the process of finding and
validating document topics.
4.3.1. Uncover Document Topics
For topic discovery we will use the first method described in ‘Graph Clusters’ part of ‘Semantic
Graph Mining’ section of this paper: we will create a sparse graph based on high threshold of
word-vector cosine similarities. As data source we will use a document that consists of data
about Creativity and Aha Moments that was manually extracted from several Wikipedia articles.
We will do the following steps:
• Retrain Word2Vec model on Creativity and Aha Moments data corpus.
• Extract collocated pairs of words and calculate cosine similarities based on Word2Vec model.
• Build Word2Vec2Graph model on pairs on vectors with cosine similarities higher than 0.8.
• Calculate graph clusters using Connected Components function from Spark GraphFrame
library.

Vol.11, No.2/3/4, July 2021
11
• Calculate graph PageRank scores by Spark PageRank function.
• For each connected component find the word with highest PageRank score and use this word
asa topic class word.
• Map words to vectors and label vectors with topic class words.
• Transform vectors to images for CNN classification.
Spark code for topic finding and vector labelings can be found in our blog post [29].
4.3.2. Validate Topics
To validate topic correctness we will apply CNN image classification method. Vectors from
uncovered topics will be converted to images with topic class words labels. Based on CNN image
classification we will compare topics with image classes. This validation method does not fully
prove topic modeling technique because clusters will have some noise: if two words are getting
into the same image cluster it does not mean that they are highly connected. But if two words are
in different image clusters they obviously do not belong to the same topic.
To convert vectors to images we will use Gramian Angular Field (GASF) — a polar coordinate
transformation. The method was suggested by Ignacio Oguiza as a method of encoding time
series as images for CNN transfer learning classification based on fast.ai library[16, 26].To
convert arrays to images and classify images we used open source code created by Ignacio
Oguiza[30].
As usual, many graph connected components have very small sizes. For that reason for topics
validation we used only connected components with size bigger than 12 nodes. Our image
classification model achieved accuracy about 91 percent.
4.3.3. Topic Examples
Examples of topics of “Creativity and Aha Moments” data corpus are displayed in Figure 3. For
each topic as a center of graph representation we used a topic class word and calculated a list of
two degree neighbors (‘friends of friends’) around topics class words. For example, here are two
degree neighbors for the class word ‘symptoms’:
• symptoms -> brain; brain -> cells
• symptoms -> disorders; disorders -> cognitive
To find two degree neighbors we used Spark GraphFrame ‘motif’ technique [31]and transformed
the results to DOT language[32]. For graph visualization we used Gephi tool [33]. Spark code
for graph visualization can be found in our blog post [29].

Vol.11, No.2/3/4, July 2021
12
Figure 3. Subgraph topic examples: top PageRank words of topics: a) "integrated"; b) "decrease"; c)
"funny"; d) "symptoms".
Topic visualization demonstrates an additional strength of using semantics graphs to uncover
document topics: graph clusters that not only reveal sets of keywords in topics, but also
demonstrate word relationships within topics.
5. CONCLUSION AND FUTURE WORK
In this paper we introduced a novel semantic graph model Word2Vec2Graphthat combines
analytic thinking and holistic thinking functionalities. We demonstrated an ability of the
Word2Vec2Graph model to analyze long documents, find unexpected word associations,
calculated word association lines, and discover document topics. Document topics that are
calculated as graph clusters not only reveal sets of topic keywords, but also show word
relationships within topics. For topic validation we suggested independent method CNN transfer
learning image classification.
In the future we are planning to do the following:
• Use more advanced word embedding models, like BERT, in particularly, examine phrase
embedding process. Evaluate new Spark NLP library [1] that allows to fine tune various word
embedding models and combine them with graph and machine learning models in Spark.
• Apply Word2Vec2Graph model to NLP problems that benefit from graph capacity to examine
relationships between objects, such as entity disambiguation, semantic similarity, question
answering, and others.
• Experiment with mapping words to vectors and vectors to images and classifying words and
sequences of words through CNN image classification methods.

Vol.11, No.2/3/4, July 2021
13
6. BROADER IMPACT
In this study for text data exploration we used a hybrid of independent techniques - semantic graph
mining and deep learning image classification. Both of these techniques are implemented by
transforming text to embedded vectors and transforming vectors to images for CNN image
classification and transforming vectors to graphs for graph mining.
The combination of graph and CNN image classification practices can also be used for other data
mining scenarios. In this study we started data exploration with graph mining and used CNN
image classification as validation method. On the contrary data investigation can start with CNN
image classification and use graph mining to uncover patterns on lower levels of granularity [34].
In addition to text data, both techniques can be applied to a variety of embeddable entities such
aswords, documents, images, videos, and many other [35].
REFERENCES
[1] Alex Thomas (2020) Natural Language Processing with Spark NLP, O'Reilly Media, Inc.
[2] T Mikolov & I Sutskever & K Chen & GS Corrado & J Dean, (2013) “Distributed representations
of words and phrases and their compositionality”, Neural information processing systems.
[3] Andrew Cattle and Xiaojuan Ma, (2017) “Predicting Word Association Strengths”, 2017Proceedings
of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1283–1288.
[4] Bill Chambers &Matei Zaharia (2018) Spark: The Definitive Guide: Big Data Processing Made
Simple, O'Reilly Media, Inc.
[5] Jurij Leskovec & Marko Grobelnik & Natasa Milic-Frayling, (2004). "Learning Substructures of
Document Semantic Graphs for Document Summarization", LinkKDD 2004
[6] Juan Martinez-Romo & Lourdes Araujo & Andres Duque Fernandez, (2016). "SemGraph:
Extracting Keyphrases Following a Novel Semantic Graph-Based Approach", Journal of the
Association for Information Science and Technology, 67(1):71–82, 2016
[7] Long Chen and Joemon M Jose and Haitao Yu and Fajie Yuan, (2017) “A Semantic Graph-Based
Approach for Mining Common Topics from Multiple Asynchronous Text Streams”, 2017
International World Wide Web Conference Committee (IW3C2)
[8] Michael Thelwall, (2021) “Word Association Thematic Analysis: A Social Media Text volume 13,
pages i-111
[9] Andrew Cattle and Xiaojuan Ma, (2017) “Predicting Word Association Strengths”, 2017
Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages
1283–1288
[10] Matan Zuckerman & Mark Last, (2019) “Using Graphs for Word Embedding with Enhanced
Semantic Relations”, Proceedings of the Thirteenth Workshop on Graph-Based Methods for Natural
Language Processing (TextGraphs-13).
[11] Long Chen & Joemon M Jose & Haitao Yu & Fajie Yuan & Dell Zhang, (2016). "A Semantic Graph
based Topic Model for Question Retrieval in Community Question Answering", WSDM’16
[12] Jintao Tang & Ting Wang & Qin Lu Ji & Wang & Wenjie Li, (2011)."A Wikipedia Based Semantic
Graph Model for Topic Tracking in Blogosphere", IJCAI’11
[13] Stavros Souravlas & Angelo Sifaleras & M Tsintogianni & Stefanos Katsavounis, (2021). "A
classification of community detection methods in social networks: A survey", International Journal
ofGeneral Systems 50(1):63-91
[14] Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, Pierre-
Alain Muller: Deep learning for time series classification: a review. Data Min Knowl
Disc 33, 917–963 (2019)
[15] Nima Hatami, Yann Gavet, Johan Debayle: Classification of time-series images using deep
convolutional neural networks Conference: Tenth International Conference on Machine Vision
(ICMV 2017).
[16] Zhiguang Wang, Tim Oates: Encoding Time Series as Images for Visual Inspection and

Vol.11, No.2/3/4, July 2021
14
Classification Using Tiled Convolutional Neural Networks. Association for the Advancement of
Artificial Intelligence (www.aaai.org)(2015)
[17] Zhiguang Wang, Weizhong Yan, Tim Oates: Time series classification from scratch with deep neural
networks: A strong baseline. International Joint Conference on Neural Networks (IJCNN)(2017)
[18] "Sparkling Data Ocean - Data Art and Science in Spark”, http://sparklingdataocean.com/
[19] Yoav Goldberg&Graeme Hirst (2017) Neural Network Methods in Natural Language Processing,
Morgan & Claypool Publishers.
[20] "Word2Vec Model Training", http://sparklingdataocean.com/2017/09/06/w2vTrain/
[21] "Introduction to
Word2Vec2GraphModel”,http://sparklingdataocean.com/2017/09/17word2vec2graph
[22] Alex Romanova, (2020) “Building Knowledge Graph in Spark Without SPARQL”, Database and
Expert Systems Applications, DEXA 2020 International Workshops BIOKDD, IWCFS and
MLKgraphs, Bratislava, Slovakia, September 14–17, 2020, Proceedings.
[23] "Find New Associations in
Text",http://sparklingdataocean.com/2018/04/04/word2vec2graphInsights/
[24] "Word2Vec2Graph Model and Free
Associations",http://sparklingdataocean.com/2017/12/24/word2vec2graphPsychoanalysis/
[25] Practical Deep Learning for Coders, https://course.fast.ai/ (2020).
[26] Jeremy Howard, Sylvain Gugger: Deep Learning for Coders with fast.ai and Py-Torch. O’Reilly
Media, Inc. (2020).
[27] Time series/ sequential data study group, https://forums.fast.ai/t/time-series-sequential-data-study-
group/29686 (2019)
[28] "GoodTherapy: PsychPedia: Free Association",
https://www.goodtherapy.org/blog/psychpedia/free-association-in-therapy (2019).
[29] "Word2Vec2Graph to Images to Deep
Learning",http://sparklingdataocean.com/2019/03/16/word2vec2graph2CNN/
[30] “Practical Deep Learning applied to Time Series", https://github.com/oguiza
[31] “Motifs Findings in GraphFrames”, https://www.waitingforcode.com/apache-
spark-graphframes/motifs-finding-graphframes/read
[32] “Drawing graphs with
dot”,
https://www.ocf.berkeley.edu/~eek/index.html/tiny_examples/thinktank/src/gv1.7c/doc/dotguide.p
df
[33] “Visual network analysis with Gephi”, https://medium.com/@EthnographicMachines/visual-
network-analysis-with-gephi-d6241127a336
[34] “EEG Patterns by Deep Learning and Graph
Mining”,http://sparklingdataocean.com/2020/08/19/brainGraphEeg/
[35] Something2vec,https://gist.github.com/nzw0301/333afc00bd508501268fa7bf40cafe4e(2016)
AUTHOR
Alex Romanova Holds MS in mathematics from Faculty of Mechanics and Mathematics,
Moscow State University and Ph.D. in applied mathematics from Facultyof Geography,
Moscow State University, Moscow, Russia. She is currently a data scientist in Melenar, an
expert in Knowledge Graph, NLP, Deep Learning, Graph Mining and Data Mining. Sharing
her experience in technical blog: http://sparklingdataocean.com/

SEMANTICS GRAPH MINING FOR TOPIC DISCOVERY AND WORD ASSOCIATIONS

More Related Content

What's hot (18)

Similar to SEMANTICS GRAPH MINING FOR TOPIC DISCOVERY AND WORD ASSOCIATIONS (20)

Recently uploaded (20)

SEMANTICS GRAPH MINING FOR TOPIC DISCOVERY AND WORD ASSOCIATIONS