Topic extraction and categorization using a glove based technique for representing topics

doi:10.21203/rs.3.rs-3232176/v1

Download PDF

Research Article

Topic extraction and categorization using a glove based technique for representing topics

https://doi.org/10.21203/rs.3.rs-3232176/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Topic extraction and categorization is an important task because by doing that it is easy to find out which are the topics most discussed by the users in their tweets or opinions and need to be analyzed. In this work, topics are extracted from positive and negative opinions and then categorized into different groups. For performing this, first a collection of opinions are divided into two sets- positive opinions and negative opinions by using a sentiment analyzer. Then a method is proposed to find out the most discussed topics in the set of positive opinions and negative opinions. For extracting the topics from a set of opinions the noun words are extracted from the set of the opinions. After extracting the topics, the similar topics have been combined by using synonymy relation. Then the frequent topic words are represented with the help of GloVe embedding technique. Finally, the topics are categorized by using a clustering algorithm by applying it on the frequent topic words. For the evaluation of the proposed method, tweets from a Twitter User dataset are used. The results obtained from the experiments by applying the proposed method on the dataset give promising result and provide interesting and meaningful clusters of topics. Moreover an analysis of the result obtained for both positive and negative opinions is also presented.

Topic detection

topic categorization

Glove embedding

Synonymy relation

K-means algorithm

A huge amount of data is daily produced by large number of social media and websites. There are various types of social networks like micro-blogging platforms, instant messaging Apps, blogging etc. Public behavior and response to different issues and events can be understood by using data derived from these social media sites. So these can be important sources of data for understanding attitudes and behaviors of users [1]. Public debate and political communication can also be influenced by the information obtained from these data. In sentiment analysis one important task to detect and categorized topics from such kind of data. Topic detection or extraction is an effective way to analyze data and relate similar text together [2].

Understanding the content of the posts or tweets directly can be a time-consuming procedure if it has been done manually. However, in some cases, posts about specific area or domains have to be obtained and it is not possible to read all the posts or tweets manually for identifying relevant topics or issues. Knowing the topic in the tweets or posts can ease the process of analyzing the tweets or posts [3]. If topic detection is applied on positive tweets or opinions, then it will help to identify the topics or issues about which people are mostly talking in a positive way. Similarly, in case of negative tweets or opinions also, the topic detection or extraction techniques can be used to find out most discussed topics or issues in the negative tweets.

After detecting the topics, the topics can be categorized, where the categories belong to a pre-defined category set or grouping the topics into unknown categories. Categorizing topics using unsupervised technique can be useful for knowing information like which topics or matters that people are mostly talking about etc. In the supervised approach, predefined categories are used to label or group the topics. But other information than the pre-specified labels, can also be existed in a dataset. In today’s world, where there is vast amount of data available in web, pre labeled dataset always cannot be easy to find. So, unsupervised techniques can benefit for these problems. Moreover, for categorization of the topics, unsupervised methods based on clustering do not need to be trained on a training dataset. Thus, detecting the topic will be robust and more suitable to be applicable to real-world settings [4][G].

Twitter is one of the most used and popular social media site [5][D]. From the previous studies it is observed that tweets can be considered as social sensors of real-time events [6][E]. Government and most of the organizations use Twitter to communicate with public. However, to find relevant topics from tweets is a difficult task because millions of tweets are daily uploaded. Twitter data i.e., tweets is used in this work and is used for finding and categorizing topics from the tweets.

The aim of this study is to detect and categorize topics from positive and negative tweets from a collection of Twitter data i.e. from a collection of tweets. Tweets are unstructured data having noises like emoticons, slang etc. which may negatively affect the performance of the detection algorithms. That is why these data have to be preprocessed before actual processing [4] [G]. After preprocessing the tweets are separated into positive and negative tweets by using Vader [7], a lexicon-based sentiment analyzer and then topics are extracted from these positive and negative tweets separately. It is seen that, in most of the works the topics are extracted by using hashtags [8] [H]. However, all tweets related to a given topic are not always hashtagged. Thus in this work, a method based on extracting frequent nouns from the textual dataset has been used to extract topics from these positive and negative tweets sets separately. A synonymy relation is used to combine all the topics with same meaning present in a list into a single entry and then only the frequent topics are kept. Thus, there will be two lists of frequent topics, one for positive tweets and one for negative tweets.

After extracting the frequent topics, the similar topics are clustered using an unsupervised method. For that purpose, first each topic present in a frequent topic list is represented as a vector using Global Vectors (GloVe) word embedding technique [9]. To achieve this, words are mapped into a meaningful space, where the distance between words is related to semantic similarity. The benefit of using GloVe is that GloVe focuses on words co-occurrences over the whole corpus. Its embeddings involve to the probabilities for two words that they appear together. In this work, K-means clustering algorithm has been applied on the set of vectors to cluster the topics. Each cluster is labeled manually based on its topics.

The main contributions of the work are-

A topic detection method is proposed where extraction of topics is performed by extracting noun words from a textual dataset and then a synonymy relation is adopted to combine the synonym topics into a single entry. The list of frequent noun words will be the resultant list of topics.
A clustering-based method is proposed to categorize the topics into different clusters. In this method, a representation technique is presented which adapts the GloVe embedding technique to represent the topic words present in the list of topics.
Manual labeling is done on the clusters of the topics obtained. Then an analysis
of the results is presented.

The rest of the paper is structured as follows. In section 2, the works related to this study are discussed. In section 3, the proposed method and the related concept are described. In section 4, the description of the dataset used, the result obtained from the experiments using the proposed method and analysis of the result are presented. In section 5, the conclusion of this paper and future lines of research is given.

Automatic topic detection or extraction is an important and interesting task in sentiment analysis field. Various methods for topic detection in Twitter data have been proposed [10], [5], [11–13]. There are approaches without word embeddings versus with word embeddings, specified versus unspecified, supervised versus unsupervised (or combination of them), and online versus offline topic detection approaches.

Some of the works employing clustering for topic extraction are summarized below. Most of the papers explored English tweets only.

In the study done in paper [14], for reducing the dimension of the data before the learning process, the authors used singular value decomposition (SVD) using K-means Clustering. Comparative results were shown while calculating the accuracy of the combination of SVD and K-means Clustering methods but the computation time required with combination was faster than the method where without any reduction in advance. K-means Clustering was applied.

In another study done in paper [15] the problem of feature extraction was tried to be solved. For performing this, in this work, initially different features such as user activity features, location based features, content and text features and user profile features were extracted. Then using natural language models, Jenson-Shannon Divergence (JSD) measure was applied on the extracted text features to characterize each labeled tweet. The authors extracted different features from the trending topics collected from twitter. Then similar trending topics of tweet user profile are grouped into clusters. Similar user profiles with same trending topics of tweet are clustered using Fuzzy K-means (FKM) algorithm. After that, using fuzzy membership function, centers are determined to similar user profiles with same trending topics of tweet. Also, from the clustering result, they applied Extreme learning machine (ELM) algorithm for the analysis of the growing characteristics of spam with similar topics in twitter. Improved detection results were achieved with the experimental results.

Again in one of the latest works done in [16], the authors used three datasets from Twitter and Reddit, and on these datasets, various methods for document clustering and topic modeling techniques were evaluated. Four different feature representations were benchmarked which were derived from term-frequency inverse-document-frequency (TF-IDF) matrices and word embedding models combined with four clustering methods and for comparison the authors include a Latent Dirichlet Allocation topic model. For interpreting the clusters, a method was also demonstrated by them which were based on top-words using TF-IDF weights combined with embedding distance measures.

In a study done in paper [17], a high utility pattern clustering (HUPC) method over micro blog streams was proposed by the users. From the micro blog stream, a group of representative patterns were first extracted in this method. Then these patterns are grouped into topic clusters. On a large scale of micro blog streams, this method works well. Using this method, coherent topics and new emerging topics can simultaneously be detected. While experimented with Twitter streams and Sina Weibo streams, this method gives better performance over existing methods.

In another study done in the paper [18], a novel method was proposed by the author for topic detection and multi-view clustering integrating multi-relations among the tweets which includes social tag relations, semantic relations and temporal relations. A method was also proposed for measuring relations among tweets. Also, a new document similarity measure which was based on a suffix tree was proposed to measure the semantic similarity of tweets in a better way. Moreover, a new method for extracting the keywords was proposed which was also based on a suffix tree.

In the work done in paper [19], an approach for topic detection was presented where a document-pivot algorithm was applied for topic detection where the algorithm clustered the documents and the clusters were treated as a topic. They considered specific features of tweets which could strongly indicate that particular sets of tweets belong to the same cluster.

In the study done in the paper [8], a novel approach was proposed where the lack of Twitter annotation for topical clustering using hash tags was addressed. In their study, inspiration was taken from the systems like Google News and past work on document clustering and categorization. From their results, it was found that the produced clusters by traditional unsupervised methods can often be inherent, but supervised methods utilizing hash-tags as indicators of topics gives good results. Temporal effects of their methodology and training set size considerations were also discussed and they also present a simple method to find the most representative tweet in a cluster.

In this section a method is described to find out the topics most discussed by the users from positive and negative opinions or tweets and then group the topics into different categories. Thus, the method described here takes as input the dataset of tweets, categorize the tweets into positive and negative tweets and outputs the lists of topics mostly discussed by the users in the positive tweets and negative tweets and then categorizes the topics. The framework of the proposed method for topic extraction and topic categorization for positive and negative tweets is shown in Fig. 1.

The details of the proposed method have been described below:

3.1 Collection of data

The aim of this work is to find out the topics most discussed by the users from a collection of positive and negative tweets. Thus, at first a set of tweets has been collected as described in section 4.1. Then, the collected data are given as input into the next phase i.e. preprocessing of the dataset.

3.2 Preprocessing of the Dataset

Tweets are unstructured data having noises like emoticons, slang etc. which may negatively affect the performance of the detection algorithms. That is why these data have to be pre-processed before actual processing [4]. In this phase, the set of collected tweets has been pre-processed. In the case of tweets the “Hashtag” may carry important information, so first “Hashtag” information is extracted from the tweets and

stored separately for future use. Then basic cleaning related to the tweets including removing Emojis, Hashtags, Mentions, Smileys, URLs, Reserved words (RT, FAV) etc. is done. Then the “Hashtag” information previously stored is added to the cleaned tweets. After this, digits are removed from the tweets and the texts are converted to lowercase. Then punctuation marks are also removed from the tweets.

3.3 Categorizing the tweets into positive and negative categories

Since, the aim of the work is to extract and categorize topics for positive and negative tweets, thus, Vader, a lexicon-based sentiment analyzer has been used to perform sentiment analysis on the pre-processed dataset of tweets. The Vader will identify the pre-processed tweets as positive, negative and neutral. The positive and negative tweets are stored in two different text files and neutral tweets are not stored as it is not needed in the proposed method.

3.4 Extraction of the distinct topic from positive and negative tweets separately

The set of positive tweets and negative tweets obtained from the previous phase are processed in this phase. In the paper [20], the authors found that using nouns rather than other words improved the topics’ coherence. Thus, in this work only noun words from both the collections of positive tweets and negative tweets have been used for the extraction of the topics. All the noun words present in both the collections have been extracted separately using the nltk pos-tagger. In each collection, for each topic word, frequencies of occurrence of the topic word in that collection of tweets are counted and only distinct topic words with their frequencies are kept. Thus, the lists for both the collections (i.e. positive and negative) represent the lists of all distinct topic terms obtained from the collections of positive tweets and negative tweets.

3.5 Combining topics using synonymy relation separately for each list

In this step, these types of words are combined and treated as a single entry and frequencies of these words are also added. In this step, the topic words which have the synonymy relation are combined into a single entry in the list of topic words. This is performed on both the lists of topics for positive and negative tweets. A Synonymy Relation is a word sense relation which refers that whether near identity is represented by the two word senses (synonyms). For example “photo” and “picture" are synonym words. So in this step, these types of words are combined and treated as a single entry and frequencies of these words are also added.

For each topic word, w₁ in the list of topics W, the synonym list is extracted from the WordNet. Then for each word w₂ (other than w₁) present in the list of topics W, it is checked if w₂ is in the synonym list of w₁. If w₂ is in the synonym list of w₁, then only w₁ is kept in the list and frequency of w₁ is increased by frequency of w₂ and w₂ is eliminated from the topics list of W.

3.6 Keeping only frequent topics in both the lists

Since, the aim is to extract the most discussed topics from the tweets, thus only frequent topic words are kept in both the lists of topics for positive tweets and negative tweets. To do this, each topic word w₁ present in the list of topics, W is visited and check whether its frequency satisfies a given threshold value. If it does not satisfy the given threshold value then the topic word w₁ is eliminated from the topic list W. So the resultant topic list contains only the frequent topic words.

3.7 Representing topic terms using Glove embedding technique for both the lists

In this step, GloVe embedding technique is used for obtaining vector representations for the topic words in both the lists of topics for positive and negative tweets i.e. the GloVe embedding technique creates a vector representation for each topic word present in a topic list. It is an unsupervised learning algorithm which is used to obtain the vector representations of the words. The training of GloVe model is performed on the non-zero entries of a global word-word co-occurrence matrix, which showcase how frequently words co-occur with one another in a given corpus. GloVe can be used to represent words with semantic meaning. That is why, it is used in this work for representing the words so that semantic meaning of the words can be obtained and the words with same semantic meaning can be grouped together by using the advantage of GloVe embedding. The pre-trained word vectors, which contain 6B tokens and 400K vocab is used in this work for implementation of GloVe embedding technique for vector representation of a word. The dimension of each vector representing a topic word has been chosen as 50.

3.8 Categorization of the topics using clustering

After representing the topic words present in a topic list as a set of vectors, a clustering technique is applied on the set of vectors to categorize the topics. The popular K-means clustering algorithm is chosen for this purpose and applied on the sets of vectors obtained from both the positive tweets and negative tweets separately, such that it will add the similar topics present in a topic list into same group and dissimilar topics present in a topic list into separate group. Here, the K-means algorithm uses Cosine distance as the similarity measure.

Cosine distance:

Using this measure, the similarity of the two vectors is computed by measuring the cosine of the angle between them. The cosine distance between two vectors of attributes, A and B, is represented by:

$$\text{ cos}\theta =\frac{A.B}{ ⃦A ⃦ ⃦B ⃦}= \frac{{\sum }_{i=1}^{n}{A}_{i}{B}_{i}}{\sqrt{{\sum }_{i=1}^{n}{A}_{i}^{2}}\sqrt{{\sum }_{i=1}^{n}{B}_{i}^{2}}}$$

4.1 Experimental data

For evaluating the proposed method, a Twitter User dataset has been used which was constructed in the paper [21]. This dataset contains 150 tweets on average from 1557 users has been created by using Twitter API. Since in this work, only the tweets are needed, so the tweets of all the users are combined and stored in a single text file. The dataset contains a total of 197614 tweets. This dataset contains tweets from January, 2021 to June, 2021.

4.2 Evaluation criteria

The proposed method is evaluated using Silhouette measure. This is a metric using which the goodness of a clustering technique is computed [22]. Its value ranges from − 1 to 1.

4.3 Result and Analysis

At first, the preprocessing steps are performed on the tweets of the created Twitter User dataset, where each object is a tweet. Since the aim is to analyze the topics in positive and negative tweets, so Vader sentiment analysis tool is applied on all the tweets to categorize the tweets into positive tweets and negative tweets. After performing this, it is found that number of positive tweets is 72098, number of negative tweets is 39749 and number of neutral tweets is 85766. In this work, focus is only on positive and negative tweets. So, neutral tweets are not analyzed here.

4.3.1 Analysis of positive tweets

To analyze the positive tweets, first, 1925 numbers of noun words are extracted from 72098 numbers of positive tweets and stored in a list. These 1925 numbers of noun words will act as the list of all the topic words (i.e. list of 1925 topic words) which are discussed in the positive tweets. Then, similar topic words are combined using synonymy relation. After performing this, number of topic words obtained is 1398. Then a minimum threshold value of 5 is given as input and the list of frequent topic words is obtained. After this step, the final list of topics is obtained which contains 1171 topic words. Then each topic word is represented by using a vector which is obtained by using GloVe embedding. Then K-means clustering is applied on the set of vectors representing the all 1171 numbers of topic words and then performance is measured using Silhouette measure. In Fig. 2, Silhouette measures for the different values of k are shown.

From Fig. 2, it is seen that for positive tweets optimal number of clusters is obtained as 17. The clusters are manually inspected in terms of topic words present in it and a manual labeling is done on the clusters (as shown in Table 1) to categorize the topic words. In Fig. 3, number of topic words present in each category for positive tweets has been shown.

From Fig. 3, it is clear that in positive tweets, the topics related to ‘National issues’ are the most discussed topics. Moreover, the topic words related to the groups, ‘Journal and literature’, ‘Household items’, ‘Concept’, ‘Emotion’, ‘Technology’ are mostly discussed topics in the positive tweets in the dataset.

Table 1

Topic words present in the clusters and manual labeling of the clusters to categorize the topic words for positive tweets
Clusters	Topic words present in the clusters	Manual labeling of the clusters
1	road, bridge, houses, cities, parts, square, middle, gates, hall, streets etc.	Construction
2	person, man, actress, mother, friend, woman, hero, actor, sir, lady, brother, uncle etc.	Human being
3	Sugar, meat, eggs, orange, nuts, drink, meal, honey, rice, cheese etc.	Food
4	idea, question, situation, examples, impact, reason, concept, cases, feedback, answers, effect, stress etc.	Concept words
5	politics, job, nation, issues, movement, effort, hope, act, country, policies, work, initiative etc.	National issues
6	news, message, radio, conversation, story, writers, book, fiction, topic, articles, copy, language etc.	Journal and literature
7	players, tournament, sports, selection, victory, career, competition, baseball, tennis, games etc.	Sports
8	worry, desire, joy, sense, anger, thoughts, calm, cry, fear, laughter, feeling, passion etc.	Emotion
9	disease, tests, vaccine, discovery, research, virus, medicines, cases, effects, temperature etc.	Medical domain
10	shelves, rings, ring, pipe, eyes, thread, ear, watches, dress, stick, cap, shirt etc.	Household item names
11	media, data, device, store, computer, windows, users, companies, platforms, phone, services, model, system etc.	Technology
12	level, period, economy, market, increase, oil, number, sector, drop, accounts etc.	Market/Economy
13	research, economics, school, student, university, professor, education, training, learning, expert	Education
14	music, album, sound, show, episode, records, scene, film, video, image, drama, songs etc.	Entertainment
15	church, community, group, member, charity, mercy, family, traditions, priest, servant etc.	Society
16	grass, trees, leaf, orange, wood, water, rock, steam, stone, garden, dust etc.	Nature
17	summer, room, hour, night, morning, day, gift, dinner, winter, midnight, holiday, arrival etc.	Time

4.3.2 Analysis of negative tweets

For analyzing the negative tweets, 1863 numbers of noun words are extracted from the 39749 numbers of negative tweets and stored in a list. These 1863 numbers of noun words will act as the list of all the topic words which are discussed in the negative tweets. Then, similar topic words present in the list are combined using synonymy relation. After performing this, number of topic words present in the list becomes 1184. Then a minimum threshold value of 5 is given as input and all the frequent topic words present in the list are obtained. Thus, the final list of topics contains 1023 numbers of frequent topic words. Then each topic word is represented by using a vector which is obtained by using GloVe embedding. Then K-means clustering is applied on the set of vectors representing the 1023 numbers of topic words and then performance is measured using Silhouette measure. In Fig. 4, Silhouette measures for the different values of k are shown.

In Fig. 4, it is seen that, for negative tweets optimal number of clusters is obtained as 16. The clusters are manually inspected in terms of topic words present in it and a manual labeling is done on the clusters (as shown in Table 2) to categorize the topic words. In Fig. 5, number of topic words present in each category for negative tweets has been shown.

From the Fig. 5, it is clear that in negative tweets, the topics related to ‘Administrative’ are the most discussed topics. Besides this, the topic words related to the groups, ‘Medical’, ‘Concept’, ‘Technology’ are mostly discussed topics in the negative tweets in the dataset.

Table 2

Topic words present in the clusters and manual labeling of the clusters to categorize the topic words for negative tweets
Clusters	Topic words present in the clusters	Manual labeling of the clusters
1	Road, bridge, houses, cities, parts, square, middle, gates, hall, streets etc.	Construction
2	person, man, actress, mother, friend, woman, hero, actor, sir, lady, brother, uncle etc.	Human being
3	Sugar, meat, eggs, orange, nuts, drink, meal, honey, rice, cheese etc.	Food
4	journey, vehicles, driving, train, driver, stops, sea, air, captains, car, crew, passenger etc.	Journey and vehicle
5	idea, question, situation, examples, impact, reason, concept, cases, feedback, answers, effect, stress etc.	Concept
6	politics, job, nation, issues, movement, effort, hope, act, country, policies, work, initiative etc.	National issues
7	news, message, radio, conversation, story, writers, book, fiction, topic, articles, copy, language etc.	Journal and literature
8	point, wave, turn, bottom, right, lock, space, brake, wing, rain, structure, light etc.	Science and Mathematics
9	Players, tournament, sports, selection, victory, career, competition, baseball, tennis, games etc.	Sports
10	worry, desire, joy, sense, anger, thoughts, calm, cry, fear, laughter, feeling, passion etc.	Emotion
11	Cause, tests, vaccine, discovery, research, experts, medicines, cases, effects, temperature etc.	Medical domain
12	shelves, rings, ring, pipe, eyes, thread, ear, watches, dress, stick, cap, shirt etc.	Household item names
13	media, data, device, store, computer, windows, users, companies, platforms, phone, services, model, system etc.	Technology
14	level, period, economy, market, increase, oil, number, sector, drop, accounts etc.	Market/Economy
15	leader, case, statement, intelligence, election, state, post, cabinet, meeting, party, house, court etc.	Administrative
16	music, album, sound, show, episode, records, scene, film, video, image, drama, songs etc.	Entertainment

From the results obtained by applying proposed method on the positive and negative tweets, it can be said that, most of the categories of topics formed are same in positive tweets and negative tweets in the Twitter User dataset. However, ‘Society’, ‘Nature’, ‘Time related’ and ‘Education’ are the groups which occur in positive tweets but not in negative tweets. So, these groups may be interesting and useful for further analysis of the positive tweets in the Twitter User dataset, because, it can be said that people are mostly talking about the topics related to these groups positively in the Twitter User dataset. Similarly, ‘Journey and vehicle’, ‘Science and mathematics’ and ‘Administrative’ are the groups which are present in negative tweets but not in positive tweets. These groups may be interesting and useful for further analysis of the negative tweets in the Twitter User dataset, because, it can be said that people are mostly talking about the topics related to these groups negatively in the Twitter User dataset. Again, the groups ‘Construction’, ‘Human being’, ‘Food’, ‘Concept’, ‘National issues’, ‘Journal and literature’, ‘Sports’, ‘Emotion’, ‘Medical’, ‘Household items’, ‘Technology’, ‘Market/Economy’ and ‘Entertainment’ exist in both the positive and negative tweets. So people are talking about the topics related to these groups in both the positive and negative way.

If the topic words are considered individually then it is found that there are 342 topic words which exist only in the positive tweets, and do not exist in negative tweets. Some of these topic words are- 'model', 'neck', 'assignment', 'wedding', 'editor', 'emphasis', 'explanation', 'umbrella', 'fan', 'term', 'bees', 'balloons', 'sense', 'purpose', 'difference', 'outcomes', 'top', 'rule', 'elegance', 'chance' etc. Again, there are 194 topic words which exist only in negative tweets but not in positive tweets. Some of these topic words are- 'loneliness', 'clearance', 'party', 'cherry', 'intention', 'start', 'echo', 'respect', 'poison', 'count', 'obligation', 'employee', 'goal', 'steel', 'wall', 'voice', 'house', 'size', 'presentation', 'curiosity' etc. These 342 topic words existing only in positive tweets and 194 topic words existing only in negative tweets may be interesting because from these topic words, it could be easy to get information about which topics, people are talking mostly in positive way and about which topics people are talking mostly in negative way in a particular set of tweets.

From this study, it could be observed that the categories of topics obtained are interesting and meaningful and can be used for further analysis of the tweets. So, the topics mostly discussed in the positive tweets and negative tweets can be extracted and categorized into different groups using the proposed method.

In this work, from collections of tweets, positive tweets and negative tweets are separated using Vader, a lexicon-based sentiment analyzer. Then individual list of topic words and categories of similar topic words are extracted for both the positive tweets and negative tweets. For performing this, a method based on extracting frequent noun words is used for extracting the list of frequent topic words from the positive and negative tweets separately. For each list, the synonym topic words are combined using synonymy relation into a single entry. Then a clustering technique is applied on the topic words to obtain the topic categories, where topic words are represented by using GloVe embedding technique. From the results obtained by the proposed method, it is seen that, different topics which are mostly discussed in the positive tweets and negative tweets can be extracted using the proposed unsupervised technique and also, meaningful and interesting clusters can be obtained for the topics without using prior knowledge.

Acknowledgments

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Asgari-Chenaghlu M, Feizi Derakhshi MR, Farzinvash L, Balafar MA, Motamed C (2021) Topic detection and tracking techniques on twitter: A systematic review. Complexity 2021:1–15
Ibrahim R, Elbagoury A, Kamel MS, Karray F (2018) Tools and approaches for topic detection from twitter streams: survey. Knowl Inf Syst 54
Baghdadi H, Ranaivo-Malancon B (2011) An automatic topic identification algorithm. J Comput Sci 7:1363–1367
Mottaghinia Z, Feizi Derakhshi MR, Farzinvash L, Salehpour P (2020) A review of approaches for topic detection in twitter. J Experimental Theoretical Artif Intell 33:1–27
Choi HJ, Park C (2018) Emerging topic detection in twitter stream based on high utility pattern mining. Expert Syst Appl 115
Tembhurnikar S, Patil N (2015) Topic detection using bngram method and sentiment analysis on twitter dataset. 4th International Conference on Reliability, Infocom Technologies and Optimization 09:1–6
Hutto C, Gilbert E (2015) Vader: A parsimonious rule-based model for sentiment analysis of social media text. Proceedings of the 8th International Conference on Weblogs and Social Media, ICWSM 8
Rosa KD, Shah R, Lin B, Gershman A, Frederking RE (2011) Topical clustering of tweets
Pennington J, Socher R, Manning CD (2014) Glove: Global vectors for word representation. Empir Methods Nat Lang Process (EMNLP) 1532–1543
Cigarran Recuero JM, Castellanos A, Garcia-Serrano A (2016) A step forward for topic detection in twitter: An fca-based approach. Expert Syst Appl 57
Bogdanowicz A, Guan C (2022) Dynamic topic modeling of twitter data during the covid-19 pandemic. PLoS ONE 17
Wang M, Jayaraman PP, Solaiman E, Chen L, Li ZE, Jun S, Geor-gakopoulos D, Ranjan R (2018) A multi-layered performance analysis for cloud-based topic detection and tracking in big data applications. Future Generation Computer Systems 87
Negara ES, Triadi D (2019) Topic modelling twitter data with latent dirichllocation method 10:386–390
Aini K, Najahaty I, Hidayati L, Murfi H, Nurrohmah S (2015) Combination of singular value decomposition and k-means clustering methods for topic detection on twitter 10
Soman SJ, Murugappan S (2014) Detecting malicious tweets in trending topics using clustering and classification. 2014 International Conference on Recent Trends in Information Technology
Curiskis S, Drake B, Osborn T, Kennedy P (2019) An evaluation of document clustering and topic modelling in two online social networks: Twitter and reddit. Inform Process Manage 57
Huang J, Peng M, Wang H (2015) Topic detection from large scale of microblog stream with high utility pattern clustering. Proceedings of the 8th Workshop on Ph.D. Workshop in Information and Knowledge Management 3–10
Fang Y, Zhang H, Ye Y, Li X (2014) Detecting hot topics from twitter: A multiview approach. J Inform Sci 40:578–593
Petkos G, Papadopoulos S, Kompatsiaris Y (2014) Two-level message clustering for topic detection in twitter. SNOW-DC@WWW
Martin F, Johnson M (2015) More efficient topic modelling through a noun only approach. Proceedings of the Australasian Language Technology Association Workshop 111–115
Naznin F, Mahanrta AK (2023) Grouping of Twitter Users according to contents of their tweets. Indonesian J Electr Eng Comput Sci 31:876–884
Rousseeuw P (1987) Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. J Comput Appl Math 20:53–65

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Topic extraction and categorization using a glove based technique for representing topics

Status:

Version 1

Abstract

Figures

1 Introduction

2 Related Work

3 Proposed Method

3.1 Collection of data

3.2 Preprocessing of the Dataset

3.3 Categorizing the tweets into positive and negative categories

3.4 Extraction of the distinct topic from positive and negative tweets separately

3.5 Combining topics using synonymy relation separately for each list

3.6 Keeping only frequent topics in both the lists

3.7 Representing topic terms using Glove embedding technique for both the lists

3.8 Categorization of the topics using clustering

4 Experimental Results

4.1 Experimental data

4.2 Evaluation criteria

4.3 Result and Analysis

4.3.1 Analysis of positive tweets

4.3.2 Analysis of negative tweets

5 Conclusions and Lines for Future Work

Declarations

Acknowledgments

References

Additional Declarations

Status:

Version 1