Microblog topic identification using Linked Open Data

Much valuable information is embedded in social media posts (microposts) which are contributed by a great variety of persons about subjects that of interest to others. The automated utilization of this information is challenging due to the overwhelming quantity of posts and the distributed nature of the information related to subjects across several posts. Numerous approaches have been proposed to detect topics from collections of microposts, where the topics are represented by lists of terms such as words, phrases, or word embeddings. Such topics are used in tasks like classification and recommendations. The interpretation of topics is considered a separate task in such methods, albeit they are becoming increasingly human-interpretable. This work proposes an approach for identifying machine-interpretable topics of collective interest. We define topics as a set of related elements that are associated by having posted in the same contexts. To represent topics, we introduce an ontology specified according to the W3C recommended standards. The elements of the topics are identified via linking entities to resources published on Linked Open Data (LOD). Such representation enables processing topics to provide insights that go beyond what is explicitly expressed in the microposts. The feasibility of the proposed approach is examined by generating topics from more than one million tweets collected from Twitter during various events. The utility of these topics is demonstrated with a variety of topic-related tasks along with a comparison of the effort required to perform the same tasks with words-list-based representations. Manual evaluation of randomly selected 36 sets of topics yielded 81.0% and 93.3% for the precision and F1 scores respectively.


Introduction
Microblogging systems are widely used for sharing short messages (microposts) with online audiences. They are designed to support the creation of posts with minimal effort, which has resulted in a vast stream of posts relating to issues of current relevance such as politics, product releases, entertainment, sports, conferences, and natural disasters. Twitter [1], the most popular microblogging platform, reports that over 500 million tweets are posted per day [2]. Such systems have become invaluable resources for learning what people are interested in and how they respond to events. However, making sense of such large volumes of posts is far from topics. First, the potential elements of topics are determined by processing linked entities in a collection of posts. Then, the elements are assigned to topics by processing a co-occurrence graph of the entities. Finally, the topics are created by processing the elements and representing them with the Topico ontology. To the authors' knowledge, this is the first approach that utilizes semantic Web and LOD to identify topics within collections of microposts.
To assess the viability of the proposed approach, we developed a prototype to generate topics for 11 collections of tweets gathered during various events. The utility of these topics (totaling 5248) is demonstrated with a variety of topic-related tasks and a comparison of the effort required to perform the same tasks with words-list-based (WLB) representations. An evaluation of randomly selected 36 sets of topics yielded 81.0% and 93.3% for the precision and F 1 scores respectively.
The main contributions of this work are: • an approach for identifying semantic topics from collections of microposts using LOD, • the Topico ontology to represent semantic topics, • an analysis of semantic topics generated from 11 datasets with over one million tweets, and • a detailed evaluation of the utility of semantic topics through tasks of various complexities.
To enable the reproducibility of our work and to support the research community, we contribute the following: • a prototype to generate semantic topics [29].
• the semantic topics generated from the datasets and the identifiers of the tweets of 11 datasets [30,31], • a demonstration endpoint for performing semantic queries over the generated topics (http:// soslab.cmpe.boun.edu.tr/sbounti), and • the manual relevancy-annotations of S-BOUN-TI topics from 36 sets corresponding to approximately 5760 tweets [31].
The remainder of this paper is organized as follows: The Related work section provides an overview of topic identification approaches. The key concepts and resources utilized in our work are presented in the Background section. The proposed approach is described in the Approach to identifying semantic topics section. An analysis of the topics generated from various datasets and their utility is detailed in the Experiments and results section. Our observations related to the proposed approach and the resulting topics are presented along with future directions in the Discussion and future work section. Finally, in the Conclusions section, we remark on our overall takeaways from this work.

Related work
The approaches to making sense of text can be characterized in terms of their input (i.e., sets of short, long, structured, semi-structured text), their processing methods, the utilized resources (i.e., Wikipedia, DBpedia), and how the results are represented (i.e., summaries, words, hashtags, word-embeddings, topics).
Various statistical topic models, such as latent semantic analysis (LSA) [7,8], non-negative matrix factorization (NMF) [9,10], and latent Dirichlet allocation (LDA) [32], aim to discover topics in collections of documents. They represent documents with topics and topics with words. The topics are derived from a term-document matrix from which a document-topic and a topic-term matrix are produced. LSA and NMF methods achieve this with matrix factorization techniques. LDA learns these matrices with a generative probabilistic approach that assumes that documents are represented with a mixture of a fixed number of topics, where each topic is characterized by a distribution over words. The determination of the predefined number of topics can be difficult and is typically predetermined based on domain knowledge or experimentation. The sparseness of the term-document matrix stemming from the shortness of the posts presents challenges to these approaches [33][34][35]. The topics produced by these approaches are represented as lists of words and will be referred as words-lists-based approaches (WLB). The interpretation of the topics is considered a separate task. LDA has been widely utilized for detecting topics in microposts. Some of these approaches associate a single topic to an individual post [36,37], whereas others consider a document to be a collection of short posts that are aggregated by some criteria like the same author [38], temporal or geographical proximity [39], or content similarity that indicates some relevance (i.e., hashtag or keyword) [40]. Fig 2 shows the top ten words of some LDA topics resulting from tweets that we collected during the 2016 U.S. presidential debates (produced by Twit-terLDA [41]). Some approaches use the co-occurrence of words in the posts to capture meaningful phrases and use these bi-terms in the generative process [4,42,43]. The determination of the predefined number of topics for large collections of tweets that are contributed by numerous people can be quite difficult.
More recently, word-embeddings learned from the Twitter public stream and other corpora (i.e., Wikipedia articles) have been used to improve LDA based [5,6] and NMF based [44] topics. In some cases, word-embeddings are used to enhance the posts with semantically relevant words [17,45]. Another utility of word-embeddings is in assessing the coherence of topics by determining the semantic similarity of their terms [46]. Alternatively, in micropost topic identification, some approaches consider topics as a set of similar posts, such as those based on the fluctuation in the frequency of the terms of interest (i.e., words and hashtags) [11,12]. The evolution of such trending topics can be traced using the co-occurrences of terms. Marcus et al. [14] created an end-user system that presents frequently used words and representative tweets that occur during peak posting activity which they consider as indicators of relevant events. Petrović et al. [47] use term frequency and inverse document frequency (tf-idf) to determine similar posts and select the first post (temporally) to represent topics. An alternative similarity measure is utilized by Genc et al. [48] who compute the semantic distances between posts based on the distances of their linked entities in Wikipedia's link graph. In these approaches, the interpretation of what a topic represents is also considered a separate task.
Sharifi et al. [20] produce human-readable topics in the form of a summary phrase that is constructed from common consecutive words within a set of posts. BOUN-TI [21] represents topics as a list of Wikipedia page titles (that are designed to be human-readable) which are most similar (cosine similarity of tf-idf values) to a set of posts. While these topics are humancomprehensible, they are less suitable for automated processing.
The approaches mentioned thus far have been domain-independent, however, in some cases domain-specific topics may be of interest. In the health domain, Prieto et al. explore posts related to specific sicknesses to track outbreaks and epidemics of diseases [49] by matching tweets with illness-related terms that are manually curated. Similarly, Parker et al. utilize sickness-related terms which are automatically extracted from Wikipedia [50]. Eissa et al. [51] extract topic related words from resources such as DBpedia and WordNet to identify topics PLOS ONE related to user profiles. As opposed to machine learning-based approaches that generate topics, these approaches map a collection of posts to pre-defined topics.
Entity linking approaches [52] have been proposed to identify meaningful fragments in short posts and link them to external resources such as Wikipedia pages or DBpedia resources [22][23][24][25]. These approaches identify topics related to single posts. Such approaches may not adequately capture topics of general interest since they miss contextual information present in crowd-sourced content. Semantic Web technologies [53] are frequently utilized to interpret documents with unique concepts that are machine-interpretable and interlinked with other data [54][55][56]. For example, events and news documents have been semantically annotated with ontologically represented elements (time, location, and persons) to obtain machine-interpretable data [57][58][59]. Parliamentary texts are semantically annotated with concepts from DBpedia and Wikipedia using look-up rules and entity linking approaches [60,61]. Biomedical references have been normalized with an unsupervised method that utilizes the OntoBiotope ontology [62] that models microorganism habitats and word-embeddings [63]. In this manner, they can map text like children less than 2 years of age to the concept pediatric patient (OBT:002307) that bears no syntactic similarity.
Our work utilizes semantic Web technologies to identify topics from domain-independent collections of microposts and to express them. Like many of the other approaches, we aggregate numerous posts. The ontology specification language OWL [64] is used to specify Topico to represent topics. The elements of topics are identified via entity linking using LOD resources. The collective information gathered from sets of posts is utilized in conjunction with the information within LOD resources to improve the topic elements. The elements of topics are related based on having co-occurred in several posts. In other words, numerous posters have related these elements by posting them together. This can result in topics that may seem peculiar, such as the FBI and the U.S. presidential candidate Hillary Clinton, which became a hot subject on Twitter as a result of public reaction. A co-occurrence graph is processed to determine the individual topics. The topic elements are the URIs of web resources that correspond to fragments of posts. Various fragments may be associated with the same resource since our approach aims to capture the meaning (i.e., "FBI", "feds" and "Federal Bureau of Investigation" to http://dbpedia.org/resource/Federal_Bureau_of_Investigation). Semantically represented topics offer vast opportunities for processing since short unstructured posts are mapped to ontologically represented topics consisting of elements within a rich network of similarly represented information.

Background
This section describes the basic concepts and tools related to semantic Web and ontologies, entity linking, and Linked Open Data that are used in this work.
Ontology is an explicit specification that formally represents a domain of knowledge [28,65]. It defines inter-related concepts in a domain. The concepts are defined as a hierarchy of classes that are related through properties. Ontologies often refer to definitions of concepts and properties in other ontologies, which is important for reusability and interoperability. Resource Description Framework Schema (RDFS) definitions are used to define the structure of the data. Web ontology language (OWL) [64] is used to define semantic relationships between concepts and the data. Ontology definitions and the data expressed with ontologies are published on the Web and referred to with their unified resource identifiers (URI). To easily refer to the ontologies and data resources, the beginning of URIs are represented with namespace prefixes. For example dbr: is often used to refer to http://dbpedia.org/resource/. A specific entity is referred to with its namespace prefix and the rest of the URI following it such as dbr: Federal_Bureau_of_Investigation for http://dbpedia.org/resource/Federal_Bureau_of_ Investigation (which is the definition of FBI in DBpedia).
We use OWL language to define the Topico ontology to express microblog topics. Other ontologies that Topico refers to are DBpedia to express encyclopedic concepts, (FOAF) [66] to express agents (with emphasis on people), W3C basic Geo vocabulary [67,68] and Geonames [69] to express geolocations, Schema.org [70,71] to express persons and locations, and W3C time [72] to express intervals of topics. The namespace prefixes that are referred to in this paper are given in S1 Table. Entity linking is used to identify fragments within text documents (surface forms or spots) and link them to external resources that represent real-world entities (i.e., dictionaries and/or encyclopedias such as Wikipedia) [52]. Entity linking for microposts is challenging due to the use of unstructured and untidy language as well as the limited context of short texts. We use TagMe [24] for this purpose as it offers a fast and well-documented application programming interface (API) [73]. TagMe links text to Wikipedia articles to represent entities. This is suitable for our purposes since the articles are cross-domain and up-to-date. Given a short text, TagMe returns a set of linked entities corresponding to its spots. For each result, TagMe provides a goodness value (ρ) and a probability (p) which are used to select those that are desirable. The chosen entities are treated as candidate topic elements. Fig 3 illustrates the response of TagMe for a short text. Here, the spot FBI is linked to https://en.wikipedia.org/wiki/Federal_Bureau_ of_Investigation with ρ = 0.399 and p = 0.547. We use ρ and p to determine the viability of a topic element.
Linked Data [26,74] specifies the best practices for creating linked knowledge resources. Linked Open Data (LOD) refers to the data published using Linked Data principles under an open license. It is an up-to-date collection of interrelated web resources that spans all domains of human interests, such as music, sports, news, and life sciences [75,76]. LOD contains 1,255 datasets with 16,174 links among them (as of May 2020) [77]. With its rich set of resources, LOD is suitable for representing the elements of topics, such as http://dbpedia.org/resource/ Federal_Bureau_of_Investigation to represent "FBI". Among the most widely used data resources in LOD are DBpedia [78] with more than 5.5 million articles derived from Wikipedia (as of September 2018 [79]) and Wikidata [80,81] with more than 87 million items (as of June 2020 [82]). DBpedia is a good resource for identifying entities such as known persons, places, and events that often occur in microposts. For example, in the short text: "POLL: The Majority DISAGREE with FBI Decision to NOT Charge #Hillary-#debatenight #debates #Debates2016" the entity linking task identifies the spots Hillary and FBI that are linked to dbr:Hillary_Clinton and dbr:Federal_Bureau_of_Investigation respectively. Both Wikidata and DBpedia support

PLOS ONE
semantic queries by providing SPARQL endpoints [83,84]. SPARQL [85] (the recursive acronym for SPARQL Protocol and RDF Query Language) is a query language recommended by W3C for extracting and manipulating information stored in the Resource Description Framework (RDF) format. It utilizes graph-matching techniques to match a query pattern against data. SPARQL supports networked queries over web resources which are identified with URIs [86]. This work uses LOD resources in SPARQL queries to demonstrate the utility of the proposed approach and to identify if topic elements are persons or locations.

Approach to identifying semantic topics
This work focuses on two main aspects related to extracting topics from collections of microposts: their identification and their representation. More specifically, the determination of whether LOD is suitable for capturing information from microposts and if semantically represented topics offer the expected benefits. The key tasks associated with our approach are (1) identifying the elements of topics from collections of microposts, (2) determining which elements belong to which topics, and (3) semantically representing the topics. This section presents a topic identification approach and describes its prototype implementation which is used for evaluation and validation purposes. First, we describe the ontology developed to represent topics since it models the domain of interests and, thus, clarifies the context of our approach. Then, we present a method for identifying topics from micropost collections, which will be represented using this ontology. While describing this method, aspects relevant to the prototype implementation are introduced in context. Finally, various implementation details are provided at the end of this section.

Topico ontology for representing topics
In the context of this work, a topic is considered to be a set of elements that are related when numerous people post about them in the same context (post). Here, we focus on the elementary aspects of the topics that are most common in social media. For this purpose, we define an ontology called Topico that is specified with the Web Ontology Language (OWL) [64] using Protègè [87] according to the Ontology 101 development process [88]. The main classes and object relations of the ontology (Topico) are shown in Fig 4. Further object properties are shown in S1 Appendix. This section describes some of the design decisions and characteristics relevant to Topico. In doing so, the references to the guidelines recommended for specifying ontologies are shown in italic font. The prefix topico is used to refer to the Topico namespace.
The first consideration is to determine the domain and scope of the ontology. Representing topics that emerge from collections of microblog posts is at the core of our domain. Thus, the ontology must reflect the concepts (classes) and properties (relations) common to microblogs. It aims to serve as a basic ontology to represent general topics that could be extended for domain-specific cases if desired. The simplicity is deliberate to create a baseline for an initial study and to avoid premature detailed design.
To enumerate the important terms in the ontology, we inspected a large volume of tweets. We observed the presence of well-known people, locations, temporal expressions across all domains since people seem to be interested in the "who, where, and when" aspects of topics. What the topic is about varies greatly, as one would expect. As a result we decided to focus on the agents (persons or organizations), locations, temporal references, related issues, and metainformation of topics. Based on this examination, a definition of classes and a class hierarchy was developed. The main class of Topico is topico:Topic which is the domain of the object properties topico:hasAgent, topico:hasLocation, and topico:hasTemporalExpression which relate a topic to people/organizations, locations, and temporal expressions. To include all other kinds of topic elements we introduce the topico:isAbout property (i.e., Topic1 topico:isAbout dbr: Abortion). We defined several temporal terms as instances of the topico:TemporalExpression class. Also, since the subjects of conversation change rapidly in microblogs, the required property topico:observationInterval is defined which corresponds to the time interval corresponding a collection (timestamps of the earliest and latest posts). This information enables tracking how topics emerge and change over time, which is specifically interesting for event-based topics like political debates and news. A topic may be related to zero or more of elements of each type. Topics with no elements would indicate that no topics of collective interest were identified. Such information may be of interest to those tracking the topics in microblogs. In

PLOS ONE
our prototype, however, an approach that yields topics with at least two elements was implemented since we were interested in the elements of the topics.
With respect to the consider reusing existing ontologies principle, we utilize the classes and properties of existing ontologies whenever possible such as W3C OWL-Time ontology [72], FOAF, Schema.org, and Geonames. FOAF is used for agents and persons. The classes schema:Place, dbo:Place, geonames:Feature, and geo:Point are defined as subclasses of topico:Location. Temporal expression of W3C OWL-Time ontology [72] are grouped under topico:TemporalExpression. The temporal expressions of interest which were not found are specified in Topico.
A topic related to the first 2016 U.S. presidential debate (27 September 2016) is shown in Fig 5. This topic is related to the 2016 U.S. presidential candidate Donald Trump, the journalist Lester Holt (topico:hasPerson), racial profiling, and terry stopping (topico:isAbout) in the United States (topico:hasLocation) in 2016 (topico:hasTemporalTerm). The subject of racial profiling and terry stopping (stop and frisking mostly of African American men) frequently emerged during the election.

Identifying topics
The task of topic identification consists of identifying significant elements within posts and determining which of them belong to the same topic. An overview for identifying topics is shown in Fig 6, which takes a set of microposts and results in a set of topics represented with Topico (S-BOUN-TI topics). Semantic topics are stored in RDF repositories to facilitate processing. Algorithm 1 summarizes the process of generating semantically represented topics given a collection of microposts. It has three phases: the determination of candidate topic elements, topic identification, and topic representation.
Algorithm 1 Topic extraction from microposts   T.add(sTopic) ⊳ Add to topics for collection P 29: end for 30: return T The first phase determines the candidate topic elements that are extracted from each post (Lines 10-13). P is a set of posts. The function entities(p) returns the entities within a post p. We denote a post and its corresponding entities as hp, li where l are linked entities. Determining the candidate elements entails the use of an entity linker that links elements of microposts to external resources and a rule-based temporal term linker. We defined temporal term linking rules [31] to detect frequently occurring terms like the days of the week, months, years, seasons, and relative temporal expressions (i.e., tomorrow, now, and tonight) to handle the various is the co-occurrence graph of candidate topic elements; gt is the set of sub-graphs of G 0 whose elements belong to the same topic; and T is the set of semantic topics represented in OWL. TagMe, Wikidata, DBpedia, and Wikipedia are external resources used during entity linking. Topico is the ontology we specified to express semantic topics. All topics are hosted on a Fuseki SPARQL endpoint.
https://doi.org/10.1371/journal.pone.0236863.g006 ways in which they are expressed in social media. We denote linked entities as [spots] ↣ [URI], where all the spots that are linked to an entity are shown as a list of lowercase terms and the entities are shown as URIs. For example, [north dakota, n. dakota] ↣ [dbr:North_Dakota] indicates the two spots north dakota and n. dakota that are linked to dbr:North_Dakota where some posts refer to the state North Dakota with its full name and others have abbreviated the word north as "n.". The entity linking process may yield alternatively linked spots or unlinked spots. For example, the spot "Clinton" may be linked to any of dbr:Hillary_Clinton, dbr:Bill_ Clinton, dbr:Clinton_Foundation or not at all (unlinked spot). Such candidate topic elements may be improved by examining the use of patterns within the collective information for agreement among various posters. Thus, the linked entities retrieved from all the posts are used to improve the candidate elements attempting to link previously unlinked spots or altering the linking of a previously linked spot (Lines 14-18). In our prototype, entities are retrieved using TagMe which links spots to Wikipedia pages. We map these entities to DBpedia resources which are suitable for the semantic utilization goals of our approach (see the Background section). At the end of this phase, any remaining unlinked spots are eliminated, yielding the final set of candidate topic elements.
The second phase decides which elements belong to which topics. We consider the limited size of microposts to be significant when relating elements since the user chose to refer to them in the same post. The more often a co-occurrence is encountered the more significant that relation is considered since the aim is to capture what is of collective interest. In this work, the term co-occurring elements/entities is defined to be the co-occurrence of the spots within a post to which these elements are linked.
To identify topics, we construct a co-occurrence graph of the candidate topic elements.
Let w: E ! R ½0;1� be a function that returns the weight of an edge and is defined as: Fig 7 shows an example co-occurrence graph constructed from four micropost texts. The linked entities obtained from these posts are dbr:Donald_Trump, dbr:Lester_Holt, dbr: Social_Profiling, dbr:Terry_Stop, dbr:Constitutionality. Within four posts, the co-occurrence between some of these entities ranged between 0.25 to 0.75. Thus, we have extracted a significantly rich set of information from the posts in terms of relating them to web resources which themselves are related to other resrouces via data and object properties.
To represent collective topics (those of interest to many people) the weak elements are eliminated prior to identifying the topics (Line 20). The weak edges (w(e)<τ e ) are removed. All vertices that become disconnected due to edge removal are also removed. The following equations describe how G = (V, E) is pruned to obtain the final co-occurrence graph The co-occurrence graph G 0 represents all related topic elements (Line 20). S2 Fig shows a co-occurrence graph obtained at this step. G 0 is processed to yield sets of related topic elements, each of which will represent a topic (Line 21). The criteria for determining the topics (sub-graphs of G 0 ) are (1) an element may belong to several topics since it may be related to many topics, (2) topics with more elements are preferable as they are likely to convey richer information, (3) topics with few elements are relevant if their relationships are strong (i.e., topics of intense public interest such as the death of a public figure).
The maximal cliques algorithm [90] is used to determine the sub-graphs, where for a graph Maximal cliques are sub-graphs that are not subsets of any larger clique. They ensure that all elements in a sub-graph are related to each other. An examination of the maximal cliques obtained from co-occurrence graphs revealed that many of them had a few (two or three) vertices. This is not surprising since it is unlikely that many elements become related through many posts. Another observation is that some elements (vertices) occur with a far lower frequency than the others. Since topics with few very weak elements are not likely to be of great interest, they are eliminated with the use of the τ sc threshold: freqðvÞ jPj < t sc where v 2 V and freq(v) = |{hp, li|v 2 l^hp, li2LEP}|.  After the elements of the topics are determined, additional information necessary to represent the final topics is obtained. The observation interval is computed using the posts with the earliest and the latest timestamps in P (Line 22). Since our S-BOUN-TI topics represent persons, locations, and temporal expressions the entity types of elements are resolved (Lines 23-25). The type of temporal expressions is determined while they are being extracted. The types of other elements are identified with semantic queries. For example, if the value of the rdf:type property includes foaf:Person or dbo:Person its type is considered to be a person. To determine if an entity is a location, first, the value of rdf:type is checked for a location indicator (schema:Place, dbo:PopulatedPlace, dbo:Place, dbo:Location, dbo:Settlement, geo:SpatialThing and geonames: Feature). Locations are quite challenging as they may be ambiguous and be used in many different manners. Then, the contexts of the spots corresponding to entities are inspected for location indicators (succeeding the prepositions in, on, or at) within the post-collection. Again, we employ a threshold (τ loc ) to eliminate weak elements of location type. Finally, an entity v is considered a location if jlocationÀ prepositionsðv;LEPÞj jPj > t loc . For example, the entity FBI in "FBI reports to both the Attorney General and the Director of National Intelligence." is considered as an agent, whereas in "I'm at FBI" it is considered as a location.
At this stage, we have the sub-graphs of G 0 (topics) along with the types of topic elements. The final phase is to represent these topics with the Topico ontology. Each t 2 T is mapped to an instance of topico:Topic (Lines [26][27][28][29]. The topics are related with their elements in accordance in accordance with their types. For example, elements of type person are associated with

PLOS ONE
a topic with the topico:hasPerson property. The property topico:isAbout is used for all elements of type other than the types of person, location, and temporal expression. The observation intervals are associated with the topico:observationInterval property. The instantiated topics are referred to as S-BOUN-TI topics and are ready for semantic processing.
Note that other graph algorithms could be used to obtain topics. Also, alternative pre-and post-processing steps could be utilized. For example, it may be desirable to eliminate or merge some topics to yield better results. For illustration purposes, let's consider the consequence of using the maximal-cliques algorithm on pruned graphs. The maximal-cliques algorithm requires all of its elements to be related. The pruning of weak bonds introduces the potential of severing the relations necessary to be identified as a topic. In such cases, very similar topics may emerge, such as those that differ in only a single element. This conflicts with our desire to favor a variety of topics with higher numbers of elements. Not pruning the graph to prevent such cases would unreasonably increase the cost of computation since the original graphs are very large and consist of many weak relations. A post-processing step could be introduced to merge similar topics with the use of two thresholds: τ c for topic similarity and t e min for an absolute minimum edge relevancy weight. Let T 0 be the set of cliques (T 0 � P(V 0 )). The set of merged cliques, T, is obtained by: Higher values for τ c or t e min lead to more topics that are similar to one another.

Prototype infrastructure
The services used to acquire external information are: the TagMe API for entity linking suggestions, the DBpedia and Wikidata for fetching semantic resources to be used as topic elements, and the Phirehose Library [91] for continuously fetching posts from the Twitter streaming API filter endpoint [92]. TagMe and Twitter have granted us access tokens to make API requests. DBpedia makes resources available under the Creative Commons Attribution-ShareAlike 3.0 License and the GNU Free Documentation License. Wikidata resources are under CC0 license which is equivalent to public domain access (both of which provide a public SPARQL endpoint). Our implementation has complied by all the terms and conditions related to the use of all services. The prototype was deployed on a virtual machine based on VMWare infrastructure running on Intel Xeon hardware with 2 GBs of RAM and Linux operating system (Ubuntu). The implementation of the maximal-cliques algorithm [90] is run within the R [93] statistical computation environment. In all of the processes mentioned above, we use a local temporary cache to reduce unnecessary API calls to reduce network traffic. Finally, all topics are represented as instances of topico:Topic, serialized into OWL, and stored in a Fuseki [94] server (a SPARQL server with a built-in OWL reasoner) for further processing.

Experiments and results
The main focus of this work is to examine the feasibility of using LOD resources to identify useful processable topics. Accordingly, our evaluation focuses on the examination of the characteristics and the utility of the generated S-BOUN-TI topics. We considered it important to generate topics from real data, which we gathered from Twitter. The quality and utility of the resulting topics are examined by: • inspecting the characteristics of their elements (Semantic topic characteristics section),

PLOS ONE
• comparing the effort required to perform various tasks in comparison to topics generated by words-list-based (WLB) approaches (The utility of semantic topics in comparison to WLB topics section), and • manually assessing their relevancy (Topic relevancy assessment section).
Furthermore, to gain insight into the similarity of topics with topics generated by other methods we compared them with human-readable (Comparison with human-readable topics section) and WLB topics (Comparison with WLB topics section).

Datasets
For evaluation purposes S-BOUN-TI topics were generated from 11 datasets consisting of 1, 076, 657 tweets collected during significant events [31]. The Twitter Streaming API was used to fetch the tweets via queries which are summarized in Table 1. The first four sets were fetched during the 2016 U.S. election debates, which are significantly greater than the others. We expected that there would be a sufficient quantity of interesting tweets during the debates, which was indeed yielded plenty of divers tweets (�48 tweets/second) resulting in very large datasets. The remainder of the datasets (except [PUB]) were collected during other notable events. These are focused due to a particular person such as Carrie Fisher or a concept such as concert. The debates related sets were collected for the duration of the televised debates and the remainder were collected until they reached at least 5000 posts. The [PUB] dataset was collected to inspect the viability of topics emerging from tweets arriving at the same time but without any query (public stream). Note that the Twitter API imposes rate limits on the number of tweets it returns during heavy use. Although they do not disclose their selection criteria, the tweets are considered to be a representative set.  Table 2 shows the number of posts and the ratios of distinct posters. The number of posts during the debates (pd 1 ,pd 2 ,pd 3 , and [VP]) are fairly similar. For all datasets, the percentage of unique contributors is generally greater than 70%, which is desirable since our approach aims to capture topics from a collective perspective.

PLOS ONE
S-BOUN-TI topics are generated from collections of tweets. The debate datasets were segmented into sets of tweets posted within a time interval to capture the temporal nature of topics. Throughout the remainder of this paper, a collection of posts will be denoted as where ds id is the name of the dataset, and t s and t e are the starting and ending times of a time interval. For example, pd 1 [10,12) refers to the tweets in pd 1 that were posted between the 10 th to the 12 th minutes of the 90-minute long debate. The earliest tweet is considered to be posted at the 0 th minute, thus t s = 0 for the first collection of a dataset.

Experiment setup
The first consideration is to determine the size of the collections. Streams of posts can be very temporally relevant, as is the case during events of high interest (i.e., natural disaster, the demise of a popular person, political debates). Furthermore, the subjects of conversation can vary quite rapidly. Short observation intervals are good at capturing temporally focused posts. Processing time is also significant in determining the size of the collections. When the rate of posts is high, the API returns approximately 5800 tweets per two minutes. Under the best of circumstances (when all required data is retrieved from a local cache), the processing time required for a collection of this size is approximately four minutes. Whenever API calls are required the processing time increases. We experimented with generating S-BOUN-TI topics with collections of different sizes and decided on limiting the size of collections to about 5000-8000 posts. This range resulted in meaningful topics with reasonable processing time. During heavy traffic, it corresponds to approximately 2-3 minutes of tweets, which is reasonable when topics tend to vary a lot.
As described earlier, our approach favors topics with a higher number of elements of significant strength. Table 3 shows the values of the thresholds we used to generate the topics where all values are normalized by the collection size.
The thresholds τ ρ and τ p are confidence values used to link entities as defined by the TagMe API, are set to the recommended default values. Higher values yield fewer candidate topic elements, thus fewer topics.

PLOS ONE
Crowd-sourcing platforms typically exhibit long tails (a few items having relatively high frequencies and numerous items having low frequencies), which is also observed for entities we identified. For example, highly interconnected and dominant six entities in a co-occurrence graph extracted from pd 1 are: Debate, Donald_Trump, Hillary_Clinton, year:2016, Tonight, and Now with weights of 0.12, 0.11, 0.11, 0.10, 0.07, and 0.03 respectively (maximum edge weight is 0.12). Similar distributions are observed in other collections.
All the thresholds are set heuristically based on experimentation. The threshold for eliminating weak edges (τ e ) is set to 0.001 based on the desire to capture entities with some agreement among posters. The threshold τ sc that is used to eliminate small cliques consisting of weak elements is set to 0.01. The threshold for clique similarity τ c is set to 0.8. Similar cliques were merged as a post-processing step (as explained in the Approach to identifying semantic topics section) where t e min is set to 0.0005 (τ e /2). Finally, τ loc = 0.01 to decide whether an entity is collectively used as a location. The cliques that remain after applying these thresholds are considered as collective topics.
To examine the impact of pruning applied before identifying topics, we traced the topic elements to the posts from which they were extracted. The percentage of posts that end up in the topics vary according to the dataset, with an average of 58% for vertices and 43% for edges (see S2 Table). Since the remaining vertices and edges are relatively strong, the resulting topics are considered to retain the essential information extracted from large sets of tweets. Table 4 summarizes the type of elements of the generated S-BOUN-TI topics. Most of them have persons, which is not surprising since tweeting about people is quite common. Topics with persons emerged regardless of whether the query used to gather the dataset included persons. Temporal expressions occurred more frequently in topics that were generated from datasets that correspond to events where time is more relevant (i.e., [CO]).

Semantic topic characteristics
The viability of our method requires the ability to link posts to linked data, thus the linked entities must be examined. From our datasets, some of the spots and its corresponding topic element that we observe are: [ Tweets can be very useful in tracking the impact of certain messaging since people tend to post what is on their mind very freely. Considering political campaigns, much effort is expended on deciding their talking points and how to deliver them. The ability to track the impact of such choices is important since it is not easily observable from the televised event or ) would yield topics, which we expected they would not since there would not be sufficient alignment among posts. Indeed no topics were generated, however, some entities were identified. We speculate that in public datasets collected during major events, such as earthquakes and terrorist attacks, the strength of ties could be strong enough to yield topics, although this must be verified.

The utility of semantic topics in comparison to WLB topics
The main purpose of this work is to produce topics that lend themselves to semantic processing. To demonstrate the utility of S-BOUN-TI topics, we provide a comparative analysis with topics represented as lists of terms (WLB) in terms of the effort required to perform various topic related tasks. The effort is described with the use of the helper functions shown in Table 5. To achieve the same results with WLB representations, it is necessary to determine the words that represent the person Hillary Clinton, other people, and count them. Thus, a type-resolution (TR) task to identify persons is needed which requires entity identification (EI) using an external resource (EX) with information about people.

Task 2-When do the topics related to women's issues occur?
This type of query is useful to know when certain topics emerge and to track whether they trend, persist, or diminish within streaming content. The time of observation of subjects is significant since they tend to rapidly change on social media.
The query in Listing 2 identified 166 topics in 66 time intervals related to topics about abortion, rape, and women's health. Retrieving the time intervals is straightforward, since the topico:observationInterval property captures this information. An inspection of the linked entities revealed that the posters used different terms related to these concepts, such as [rape, raped, rapist, rapists, raping, sexual violence, serial rapist] ↣ [dbr:Rape].

Listing 2 Query: When do the topics related to women's issues occur?
SELECT To achieve the same results with WLB representations, the terms related to the women's issues and the time intervals (TI) corresponding to the time they appeared must be determined.

Task 3-When do the top 50 issues related to topics including Hillary Clinton and/or Donald Trump appear?
This type of query is relevant for tracking how a particular messaging resonates with the public, such as for political and marketing campaigns that involve immense preparation.
The query in Listing 3 retrieves the top 50 issues (topic elements) associated with the topics including Donald Trump and/or Hillary Clinton and when they were observed. To do this, first, the top 50 issues (topico:isAbout) in the topics related to dbr:Donald_Trump or dbr: Hillary_Clinton are retrieved. Then, when these issues occurred is determined. This query returned 3061 results, among which: time: "2016-10-10T01:38:00Z"^^xsd:dateTime issueOfInterest: dbr:Patient_Protection_and_Affordable_Care_Act person: dbr:Donald_Trump that indicates that the issue of patient protection and affordable care act occurred in topics with Donald Trump on 09 October 2016 at 21:38 EST (during the 2 nd presidential debate). Fig  9 summarizes some of the issues that co-occurred with Hillary Clinton and/or Donald Trump for each two-minute interval during the 90-minute long debates (pd 1 ,pd 2 ,pd 3 ,[VP]). To gain some insight regarding how the resulting issues corresponded with the actual debates we inspected them along with their transcripts [95][96][97][98]. For example, racism was an issue that was mostly discussed by the candidates during the second half of the first presidential debate and the first half of the third debate. The topics we identified also revealed that racism was mostly posted during the same time (see the rows labeled White_people and Black_people in Fig 9). Furthermore, an inspection of the tweets posted during the same time also included tweets related to racism that were posted in pro-Republican and pro-Democratic contexts. For the topics related to Tax and Donald Trump only ([VP] [48,50) and pd 3 [80,82)) the corresponding tweets were indeed related to only Donald Trump. On the other hand, while the candidates were talking about ISIS, Iraq, and the position of the United States in the Middle East (pd 3 [68,70)) the identified topics were related to illegal immigration and income tax. In this case, we observe a lack of resonance between what was transpiring during the debate and the topics of interest to the posters (who preferred to post about other matters).

Listing 3 Query: When do the top 50 issues related to topics including Hillary Clinton and/or Donald Trump appear?
This query demonstrates the use of the inverse relationships topico:isAPersonOf and topico:inTopic. These relationships are inferred trough reasoning according to the definitions in Topico ontology (see descriptions of object relationships in Description Logic in S1 Appendix).
To achieve the same result with WLB topics, the terms indicating Hillary Clinton, Donald Trump, and the terms corresponding to the top 50 issues must be identified (EI) which requires reference to external resources (EX). Finally, when the issues emerged must be identified (TI).

Task 4-Which politicians occur in the topics?
This task requires determining the occupation of persons, which may be of interest to known people. While S-BOUN-TI topics include the persons, the DBpedia entities that represent them often do not include the dbo:occupation or worse are related to incorrect entities. However, Wikidata entities utilize wdt:P106 (occupation) property with persons quite systematically. Since DBpedia refers to equivalent Wikidata entities with the owl:sameAs property, the occupation of a person can be retrieved from Wikidata. The query in Listing 4 fetches the politicians in topics extracted from the debates by: (1) fetching all persons in the S-BOUN-TI topics from our endpoint, (2) retrieving the Wikidata identifiers of these persons from the DBpedia endpoint, and (3) identifying the persons whose occupation (wdt:P106) is a Politician (wd: Q82955) from the Wikidata endpoint. Among the results are: dbr:Abraham_Lincoln, dbr:Bill_ Clinton, dbr:Colin_Powell, dbr:Bernie_Sanders, and dbr:Saddam_Hussein. Query optimization (QO) is performed to reduce the search space by prioritizing the sub-queries according to their expected response sizes. This example shows the benefits of using LOD in our topics, where the links within the entities lead to a multitude of options. To achieve this with WLB topics, the terms referring to the bands and locations (EI and TR) are needed. External resources (EX) are needed to identify bands, locations, and the genre of the bands. Also, as explained in the Identifying topics section, the context of the location terms must be examined to determine if they were indeed used as locations.

Task 6-Which of the issues related to Barack Obama during the 2012 and 2016 U.S. election debates are the same?
In politics, it is useful to know which issues persist over time. The query shown in Listing 6 queries the topics identified during the 2012 and the 2016 U.S. election debates. For this purpose, topics were generated from the tweets gathered during the 2012 U.S. Presidential debates [99]. Central_Intelligence_Agency. There are many issues one would expect to see in a presidential debate such as taxes, violence, and the economy. Among other issues that appear in both years are racism, immigration, Muslim, Iraq, and Russia. One might be surprised to see golf (dbr: Golf) in this list; alas, the amount of golf played by candidates seems to be a matter of public interest. An inspection of the tweets confirms that the amount of golf that Barack Obama played became a topic of discussion.
Listing 6 Query: Which of the issues related to Barack Obama are the same in the 2016 and 2016 U.S. election debates? This is a federated query that queries two endpoints, one for the debates in 2012 and one for the debates in 2016. SELECT  To obtain a similar result with WLB representations, words common to topics of 2012 and 2016 must be retrieved. The results would be terms rather than concepts. For conceptual results, entity identification (EI and EX) could be used.

Task 7-Which religions and ethnicity were mentioned during the 2012 and 2016 debates?
Topico explicitly represents only persons, location, and temporal elements. To detect other types, external resources must be used. The query shown in Listing 7 utilizes knowledge about religions and ethnicities in Wikidata to retrieve related topics in the 2012 and the 2016 U.S. elections' debate topics with Query 1 that retrieves all religions from the Wikidata endpoint and Query 2 that retrieves the topics that include any of the items fetched in Query 1. A program that optimizes this query by feeding the output of Query 1 to Query 2 is used for this task (QO). The same process is repeated for ethnic groups. The tweets themselves refer to specific religions or ethnicities (i.e., Christian and Mexican). This query enables retrieving information about religions and ethnicities independent of any specific instance.
Listing 7 Query: Which religions were mentioned during the 2012 and 2016 debates? This query is issued using two queries. Query 1: Get the religions from Wikidata, where the property P279 � means all subclasses and Q9174 is the identifier for the religion class. Query 2: Get the topics that include religions. The query for 2012 returned only dbr:Catholicism, whereas for 2016 it returned dbr:Isla-m_in_the_United_States, dbr:Islam, and dbr:Sunni_Islam. A manual inspection of tweets confirms the difference in tweeting about religion. In 2012, Catholicism was a subject of concern related to abortion and in 2016 Islam became an issue in the context of the Iraq War and the 9/ 11 terrorist attacks.
Furthermore, we observed that the topic elements that co-occur with dbr:African_Americans also varied. With the support and opposition to the black lives matter movement, the elements dbr:Police and dbr:Racism were observed in 2016.
To accomplish this task with WLB representations, the identification of religions and ethnic groups are needed (TR) that requires entity identification (EI) using an external resource (EX).

Task 8-Which people are related to the same topics?
Semantic representation enables inference from present information, such as introducing the vcard:hasRelated property that specifies relationships among people and organizations (see vCard ontology [100]). This property can be used to relate people that occur in the same topic using the Semantic Web Rule Language (SWRL) [101]  In the WLB case, persons in topics must be identified (TR), which requires entity identification (EI) using external resources (EX). There must be some way of expressing this relation so it can be referenced.

Task 9-What are the categories of topics?
Topic enrichment [102][103][104] through external resources (EX) may greatly enhance the utility of topics. A useful enrichment for S-BOUN-TI topics would be to relate them to their DBpedia subject categories through the dct:subject property. For example, the category of the dbr:Job is dbc:Employment, which indirectly relates all S-BOUN-TI topics having dbr:Job as a topic element to dbc:Employment. The following SWRL rule (RD) enriches S-BOUN-TI topics with topico:isAbout relations to the categories of their elements: In this query, when the DBpedia category dbc:Employment is replaced with dbc:Law_enfor-cement_operations_in_the_United_States the results include topics with the element dbr:Stopand-frisk_in_New_York_City. Therefore, with this simple rule definition, it becomes possible to relate the topics with their categories and query the topics according to these categories.
A similar enrichment for WLB topics requires external resources (EX) and functionality such as semantic analysis (SA) of topics that could require considerable programming. Table 6 summarizes the effort to perform Task-1 through Task-9 for S-BOUN-TI and WLB approaches in terms of the subtasks that must be performed. The subtasks are described in terms of the helper functions, where those that must be performed numerous times are indicated with a subsequent parenthesized number (i.e., TR(2), for two type resolutions). For the sake of brevity, we assume the existence of primitive functions (i.e., string, set, and list operations) and query support, which are not indicated in the comparison.
Semantically represented topics offer many opportunities when they are utilized in conjunction with resources and ontologies within LOD. The utility of S-BOUN-TI topics is most apparent when it yields results that are not directly accessible in the source content. The use of semantic rules enables enriching topics with general or highly domain-specific information. The latter being quite lucrative for domain-specific applications.

Topic relevancy assessment
A comparative evaluation of the relevancy of S-BOUN-TI topics is difficult since the proposed approach has no precedence and produces topics that are significantly different from other approaches. The effort required for manual evaluation is complex, highly time-consuming, and error-prone since it involves the simultaneous examination of large sets of tweets (approximately 5800 per collection) and many semantic resources for every topic. The level of effort and diligence required to evaluate topics through surveys or services such as Amazon Mechanical Turk [105] that rely on human intelligence was deemed prohibitive. However, to gain insight regarding the relevancy of the topics, a meticulous evaluation was performed by the authors of this work with the assistance of a web application we developed for this purpose (see S1 Fig). This tool presents a set of topics to be annotated as very satisfied, satisfied, minimally satisfied, not satisfied, or error (when URIs are no longer accessible) along with optional comments to document noteworthy observations. During annotation, an evaluator may view the tweets from which the topics were generated as well as a word cloud that presents the words in proportion to their frequency. Also, the linked entities and temporal expressions extracted from the tweets can be inspected.
For evaluation purposes, 10 topics from randomly selected 36 intervals (9 from each debate) were annotated (S3 Table). Two annotators evaluated 24 intervals, 12 of which were identical to compute the inter-annotator agreement rate. The topics to be evaluated were selected based on a higher number of topic elements since they result from higher levels of alignment among posters. As such, they were deemed more significant to evaluate. Of the topics shown to annotators, there were 3 of size 8, 13 of size 7, 66 of size 6, 162 of size 5, 147 of size 4, 87 of size 3, and 2 topics of size 2. The topics are presented per interval since they are all identified from the same collection. The evaluator is expected to inspect each topic to determine if it is related to tweet collection from which it was generated (by also inspecting the tweets). Each element of each topic is inspected by visiting their DBpedia resources to determine their relevancy to the collection in the context of the other elements of the topic. An element that is related to the tweet set, but not in the context of the other elements is considered irrelevant. Each topic is labeled as: very satisfied only if all of the topic elements are valid; satisfied only if one of the topic elements is incorrect; minimally satisfied if more than one element is incorrect while retaining significantly valuable information; and not satisfied if several topic elements are incorrect (i.e., the relative temporal expression may be true but does not convey sufficiently useful information). Note that the evaluation was performed in a strict manner, where a penalty is given for any kind of dissatisfaction-regardless of the source of the error. For example, if a web resource on DBpedia has incorrect information (which happens), the annotation of that topic is penalized. This was done to avoid subjective and relative evaluation as well as to assess the viability of the resources being used. Furthermore, since S-BOUN-TI topics are produced for machine interpretation the accuracy of topic elements is quite significant. It is also easier to identify mistaken elements in contrast to assessing a whole document as an error.
The results are examined in two ways: for topics marked either Very satisfied or Satisfied (assuming general satisfaction) and for topics annotated exclusively as Very satisfied. The evaluation resulted in the precision and F 1 scores of 74.8%, 92.4% when considering only those marked as Very satisfied, and 81.0%, 93.3% when Very satisfied or Satisfied. The F 1 scores (computed as defined by Hripcsak and Rothschild [106]) indicate a high degree of agreement among annotators.

Comparison with human-readable topics
In an earlier work (BOUN-TI [21]), we identified human-readable topics from collections of microblogs (Wikipedia page titles). BOUN-TI models collections of posts as bags of words and compares their tf-idf vector with the content of Wikipedia pages to identify a ranked list of topics. The titles of the pages represent topics that are easily human-interpretable. BOUN-TI topics are satisfactory for human consumption, especially since they are descriptive titles produced by the prolific contributors of Wikipedia. Misleading topics can result when several subjects are posted about with similar intensities, such as the topic Barack Obama citizenship conspiracy theories derived from the words Barack and citizen whereas context of citizen was in Hillary is easily my least favorite citizen in this entire country-clearly not related to Barack Obama. Such cases occur as a consequence of using bag-of-words to model the documents. S-BOUN-TI overcomes this issue by considering both the wider context of the collection and the local context of posts while identifying topics. The context of individual tweets is used to determine potential topic elements, while the context of collections to capture the collective interest and patterns of use.
We inspected and compared BOUN-TI and S-BOUN-TI topics by deriving them from the same datasets (see S1 Fig). To give example, for pd 1 [26][27][28], some of BOUN-TI topics are: Donald Trump, Hillary Clinton, Bill Clinton, Barack Obama's Citizenship, and Laura Bush. Since S-BOUN-TI topics include many elements, we will suffice by mentioning some of the topic elements: persons dbr:Hillary_Clinton, dbr:Donald_Trump, dbr:Lester_Holt (the moderator of the debate) and other elements dbr:Debate, dbr:ISIS, dbr:Fact, dbr:Interrupt, dbr:Watching, and dbr:Website. These elements are identified because people were talking about ISIS, a high level of interruptions during the debate, and Hillary Clinton's fact-checking website.
The evaluation of BOUN-TI topics yielded 79.3% and 89.0% for precision and F 1 scores for those marked Very Satisfied only and 88.9% and 94.0% if annotated as Very Satisfied or Satisfied. The scores for BOUN-TI are higher, which is largely influenced by two factors. Firstly, BOUN-TI topics are single titles that tend to be high level, thus they tend to be relevant even when not very specific. For example, the Debate topic would be considered relevant to a collection of posts about a specific debate at a specific time, even though it is very general. In S-BOUN-TI, topics are more granular with more elements, thus the evaluator scrutinized each element to determine relevancy in a manner that penalizes mistakes. Since S-BOUN-TI topics are intended for machine processing, a harsher judgment is called for.
In summary, we have found some similarities between S-BOUN-TI and BOUN-TI topics. In some cases, the corresponding DBpedia resources of BOUN-TI topics (Wikipedia pages) were elements of S-BOUN-TI topics that indicate a similarity between the results of BOUN-TI and S-BOUN-TI. Both approaches produce relevant topics, while S-BOUN-TI produces a greater variety and more granular topics in comparison to BOUN-TI topics. In general, BOUN-TI captures higherlevel human-readable (encyclopedic) topics, while S-BOUN-TI picks up on lower-level elements that provide conceptual information that lend themselves to a greater variety of machine-interpretation such as Barack Obama is a person and was a president.

Comparison with WLB topics
Latent Dirichlet allocation (LDA) is one of the most popular topic models, which makes a comparison with S-BOUN-TI interesting. To perform a comparison, LDA topics are generated with TwitterLDA [41] with the two-minute intervals of the datasets pd 1 , pd 2 , pd 3 , and [VP] that are used for generating S-BOUN-TI topics (using the default values of LDA α = 0.5, β = 0.01, number of iteration = 100). Topics were generated for alternative values of LDA parameters for the expected number of topics: N = 2 − 10, 20, 30, 40, 50.
The topic representations of S-BOUN-TI and LDA are very different where LDA topics capture terms expressed by contributors as words-list-based (WLB) topics, S-BOUN-TI topics map original content to instances in LOD which are expressed with LOD resources and the OWL language [64]. With S-BOUN-TI a set of alternative words that are contributed may be mapped to the same semantic entity, capturing the intended meaning rather than how it was articulated.
To get a rough idea about the similarity of topics, we utilized the label (rdfs:label) of the S-BOUN-TI topic elements. The union of the lowercase form of words in the labels of all elements is compared using Jaccard similarity with topmost 10 terms of LDA topics (according to their distributions). We observed that there are cases that LDA and S-BOUN-TI topic elements are the same but not matching due to some syntactic difference. For example, if an LDA topic element is "emails", and S-BOUN-TI topic element is dbr:Email, the strings "emails" and "email" do not match which results in lower similarity scores. To address similar issues, we assumed that the cases that one term is a substring of the other are matching. Each S-BOUN-TI topic is elementwise compared with LDA topics that are generated for the same input set.
The maximum similarity of an S-BOUN-TI topic in an interval is considered its similarity. The average of such similarities in an interval is the similarity measure obtained from that interval. And, the average similarity for all intervals is the average for a dataset, which ranges between 60-70% with a maximum of 77%. Since the comparison of the topics is performed on elements of different levels, the results give a very rough idea. The semantic similarity is expected to be higher. We would have been concerned if the comparisons resulted in very low values since that would indicate a significantly different relation among topic elements. As a result, we observe considerable coverage between the topics identified by these approaches, which is interesting for future work towards alternative methods for identifying topic elements. For both methods, and in the case of using any other comparison methods to compare S-BOUN-TI topics with words-list-based topics, there is still the issue of word versus entity comparison. Automatically assessing the relevancy of topics without a gold standard is a challenging issue that requires domain knowledge and understanding of "topics" in the domain. We address these operations for future work.

Evaluation summary
To assess the proposed approach, S-BOUN-TI topics were generated from sets of tweets and examined by inspecting their characteristics, using them in processing tasks, and comparing them with topics generated from BOUN-TI. Our main inquiry was to assess the viability of generating topics from collections of microposts with the use of resources on LOD. We found that considerable links between tweets and LOD resources were identified and that identifying topics from the constructed entity co-occurrence graph yielded relevant topics. With semantic queries and reasoning, we saw that it was possible to reveal information that is not directly accessible in the source (tweets), which could be very useful for those (i.e., campaign managers, marketers, journalists) who are following information from social media.
The proposed approach is a straightforward one aimed at gaining a basic understanding of the feasibility of mapping sets of tweets to semantically-related entities. If possible, this would facilitate a vast number of applications that harvest the richly connected web of data. Our observations lead us to believe that this is possible. Furthermore, this approach would improve by enhancing the techniques used to identify and relate topic elements, refining the topic representation, and with the increasing quality of data on LOD, which have been improving in terms of quantity and quality during the span of this work, a most encouraging prospect. Potential improvements are elaborated in the following section.

Discussion and future work
In this section, we discuss some of our observations regarding the approach we proposed and present some future directions. The main objective of this work was to examine the feasibility of linking informal, noisy, and distributed micropost content to semantic resources in LOD to produce relevant machine-interpretable topics. We specifically focused on subjects of significant interest from a collective perspective. Topics of general interest lead to vast numbers of microposts. We generated semantic topics from a variety of tweets collections and represented them with an ontology that we developed for this purpose. The semantic topics were subjected to various tasks to examine their utility. The results show that relevant topics were identified for a diverse set of subjects. In the Experiments and results section, we presented the semantic topics generated from collections of tweets with emphasis on a complete set collected during the four major debates of U.S.elections (a total of 1036800 tweets). The utility of the resulting topics (respectively 1221, 1120, 1214, 1511 number of topics) was demonstrated through various tasks that facilitated the understanding of the issues relevant to the debate watchers, such as the persons, the locations, the temporal and other aspects of interest. Furthermore, issues at higher conceptual levels such as violence, ethnicities, and religions were revealed.
In our experiments, we observed that our approach produces relevant topics for diverse contexts. The topics of an entirely different context can be observed in a subject that is of great interest during the final preparations of this article, namely the coronavirus 2 (SARS-CoV-2) pandemic (a.k.a.  that is widely reflected on social media. A preliminary exploration of topics generated from collections of tweets related to COVID-19 also yielded relevant topics. In this case collections of tweets posted during the same time for 53 consecutive days were inspected to get a general sense of the issues of relevance. There were 140 people, 32 locations, 46 temporal expressions, 1097 issues that distinctly occurred in the topics. Among the occupations of people are politicians (several heads of states), journalists, singers, and athletes. The locations were dominated by China, Wuhan, Italy. While many locations were across the globe (i.e., Germany, West Bengal, and London) others were regions within the U.S. (i.e., Texas, Michigan, and Louisiana). This is reasonable since the time intervals of the tweets correspond to midday in the United States and COVID-19 cases were spiking in various parts of the country. There were also numerous temporal references, the most frequent ones being now, today, tonight, the months of January through May. The occurrence and the frequency of the specific temporal terms are significantly different from those encountered in the debate related sets, which did not have such a diverse set of temporal expressions (mostly the year 2016 now, tonight). Such differences capture the nature of contributions where the temporal aspect of the pandemic is indeed of much more significance due to the interest in how fast the rate of cases change and speculation about when things would improve. The resulting topics were processed to see when various issues emerged. Fig 10 shows when the about topic elements (topico:isAbout) were observed daily. Upon observing the references to drugs, we checked if other drugs were also referenced simply by querying DBpedia if the element type is dbo:Drug. This identified the other drugs referenced in tweets as: BCG vaccine, Cocaine, Doxycycline, Favipiravir, Generic drug, Pharmaceutical drug, Polio vaccine, Ibuprofen, Paracetamol, Chloroquine, Azithromycin, Antiviral drug, Hydroxychloroquine. The most frequently referenced one was Hydroxychloroquine-a drug mentioned by the president of the United States several times. Obviously, tweets from such small intervals are insufficient to inspect such a vast issue. A more comprehensive examination with collections that covers all time zones would be required. Nevertheless, even with this small set, it is evident that this approach produces relevant topics.
The results we obtained are encouraging leaving us with many future directions to pursue, which we elaborate in the remainder of this section.

Topic element detection
Since semantic topics consist of topic elements, correctly identifying them is important. Here, the main challenges are the inability to link to anything at all and incorrect linking. Obviously, entity linking fails when suitable entities are not represented on LOD, such as when new subjects emerge. In recent years the significance given to the creation and accessibility to open data resources has led to a rapid increase in the data represented on LOD [74] (see Background section for information about LOD). In this work, we eliminate unlinked spots, which could be particularly problematic for spots with high frequencies since that indicates common interest. To alleviate this matter, such spots could be linked to an instance of owl:Thing indicating that there is some thing of significance whose type is unknown.
For entities that exist but not successfully linked, better approaches are required. Named entity recognition and linking are active research areas that are improving across all domains and languages. Another approach to determining the correct entity type ranks all relevant types using taxonomies and ontologies such as YAGO, and FREEBASE [107,108]. Also, additional pre-processing steps can be taken prior to entity linking, such as tweet normalization [109] and hashtag segmentation [110].
As discussed in the Experiments and results section, identifying locations is challenging since many entities can be considered a location in some context. We imposed some rules to determine if such elements qualified as a location in the context it was used. Our evaluation revealed that although all elements we deemed to be locations were correct. Unfortunately, we missed identifying some of them since they did not match our rules-mostly due to how tweets were articulated. Location prediction on Twitter is known to be challenging and is of significant interest since there are many areas of application [111,112]. It is of interest for many purposes, such as disaster tracking and mitigation and with the emergence of the COVID-19 pandemic crisis, this work has been intensified. Many studies appear as preprints, which are not yet vetted, however with the immense motivation improved location detection is expected. Our rules for detecting locations must be revisited. Also, indirect location indicators such as those found in profiles and geotagged content [113], and in co-occurrence patterns [114] could offer hints that improve the detection of locations. For ethical reasons, we do not (and do not intend to) use any profile information, but could consider utilizing other indirect signals.
A more troublesome issue stems from ambiguous terms, which is most prevalent in person names. For example, the spot Clinton was inaccurately linked to the 42 nd U.S. president Bill Clinton instead of the 2016 U.S. presidential candidate Hillary Clinton in tweets regarding lying about Obamacare. In this case, both persons are politicians, one was a U.S. president and the other a U. S. presidential candidate, and they are spouses. Although this example is particularly challenging, the ambiguity of person names is generally challenging. A similar issue arises since the titles of songs, movies, albums, and books are terms of ordinary conversation, such as time (Time magazine), cure (music band The Cure), and WHO (television character Dr. Who). We encountered such cases in our experiments, albeit not frequently, since the entity linker typically assigns low confidence rates for such links. Furthermore, our approach eliminates the links to entities that occur infrequently (see the Identifying topics section) in a collaborative filtering manner.
Recently, word-embedding techniques that capture semantic similarity among terms are being applied to named entity recognition (NER) and disambiguation [115,116], and entity linking [117,118]. These techniques represent terms as vectors in a high dimensional vector space and obtain them via machine learning given a corpus. The semantics of terms are captured from the context of the terms. The vectors of semantically similar terms are close in the vector space. Since emerging entities are expected to be included in the knowledge bases, the topic and/or topic elements could be periodically revisited for opportunities for improvement. These advances in named entity detection and linking are very promising and are expected to positively improve the detection of topic elements.
In our experiments, we focused on English tweets and named entity recognition so that we could interpret the results. Several tools work for other languages (including TagMe that we used in our prototype). Furthermore, the natural language processing community strongly emphasizes work on low resource languages, which is resulting in additional knowledge resources and tools. The goal is to work with multiple languages, which link to the same conceptual entity. Thereby, being able to glean information regarding content that is globally produced. This is important for many tasks of global interest, such as pandemic diseases, disasters, news, entertainment, and learning material.

Semantic topic identification
In this work, we chose maximal cliques to identify the topics so as to assure that all the elements are related by virtue of having been posted together. The co-occurrence graphs from which topics are extracted have relatively few nodes with high degree centrality (i.e., Hillary Clinton and Donald Trump in the debate sets) with the remaining node being relatively weak (see Fig 7). Thus, several topics extracted from such graphs tend to share the dominant nodes, which reflect the narratives related to the dominant nodes. On the other hand, the nodes that are connected to the dominant nodes tend to fall into different topics since they are usually not connected to each other. This fairly accurately reflects micropost content (i.e., many different topics involve Donald Trump). However, this results in some topics seeming very similar or repetitive. It is worth investigating more relaxed graph algorithms to increase the elements of topics while preserving the context. For example, k-cliques (the maximal sub-graphs where the largest geodesic distance between any two vertices is k) constrained by type rules could yield richer topics. However, caution must be exercised, since the volume of microposts and their limited context is likely to yield many potential yet unrelated candidates for k > 1, which is computationally challenging and costly. Note that the goal is not to simply increase the number of topics or their elements since we are aiming to reduce large sets of tweets to higher-level topics. Rather, the aim to increase the quality of topics by associating related elements.
The size of the post collections we used was limited by the rate limits of the Twitter streaming API and our computational resources. For the debates, this corresponded to the tweets posted withing 2-minute intervals. During heavy posting conditions the subjects change frequently and short windows are suitable since the topics change and the number of tweets to detect collective interest is sufficient. During slower posting conditions the subjects don't change as fast, thus sets collected over longer durations are appropriate. Dynamically varying collection durations based on how frequently the subjects of topics change over time would be valuable. Topico is capable of representing such intervals, however, they must be determined through time series analysis.

Semantic topic representation
Topico specifies an elementary set of topic element types, namely person, location, temporal expressions, and other entities (those related by topico:isAbout. It encompasses basic classes, object properties, and data properties to represent commonly occurring elements. Inferred relations and classes support convenient processing. This ontology could be extended to cover additional types (such as events, art, currency, character, products, drug, and natural objects) as well as refine existing types (such as facility, address, astral body, organization, and market) [119]. While covering a wider type of cross-domain entities is of interest, we expect that customized for a specific domain will be quite interesting. For this purpose, ontologies relevant to the domain of interest and associated data resources are required. There are many useful ontologies and resources, especially in the life-sciences domain. Naturally, domain-specific tasks would also be defined. The tasks shown in the Experiments and results illustrate the kinds of topic-related tasks that could be of interest to campaign managers, journalists, and political enthusiasts.

Semantic topics utilization
The purpose of focusing on semantic topics is for their semantic processing potentials. We demonstrated how semantic topics can be utilized through various semantic tasks in the Experiments and results section. The vision is to deliver this power to an end-user who is following the rapidly flowing distributed microposts. Towards this end, higher-level tasks should be defined such as similarity, sentiment analysis, and recommendations. Tracking topics, such as when they emerge, if they persist, if and when they spike, and if they exhibit some pattern is useful information. Reports that provide statistical information will enable those who are interested in the topics to take action. Other interesting tasks are tracking the evolution of and predicting topics. One of the future directions is an explorer for S-BOUN-TI topics which requires the generation of human-interpretable topics. Using such an explorer users could search and browse topics, view ranked topics, graphs, and charts that provide relational and temporal information (trends), view social network analysis. They must be able to view the results of the processing of semantic topics, which may be predefined domain-specific. Such a system should recommend topics and topic related observations such as trends, newly emerging issues. Furthermore, multimedia presentations that depict the lifecycles of topics persist over a given time can be generated with dynamic summarization techniques [120,121].
Eventually, a tool that is customizable with domain-specific knowledge resources for detecting and processing topics. The specific nature of the subject and desired processing will vary depending on the context. A domain-specific topic detection system customized for diseases with knowledge bases like the International Classification of Diseases (ICD) [122] for diseases and SNOMED-CT (Systematized Nomenclature of Medicine-Clinical Terms) [123] would be useful in tracking the for pandemics related topics of public interest. Such topic explorers would enable users to glean domain-specific insights that are very difficult to obtain by direct experience with a vast number of microposts.
One of the most interesting potentials of semantic resources is revealed via federated queries that search across distributed resources. Unfortunately, the performance of federated queries can be quite inefficient. The order in which queries are executed must be carefully designed to achieve reasonable response times. Generally, electing to execute the more restrictive queries prior to others which restricts the search space is considered a good approach. Finally, generating streams of semantic topics could be facilitated with stream reasoning [124] and queried with a stream query language such as C-SPARQL [125] and C-SPRITE [126].

Conclusions
This work investigates the viability of extracting semantic topics from collections of microposts via processing their corresponding linked entities that are LOD resources. To this end, an ontology (Topico) to represent topics is designed, an approach to extracting topics from sets of microposts is proposed, a prototype of this approach is implemented, and topics are generated from large sets of posts from Twitter.
The main inquiry of this work was to examine whether an approach based on linking microposts to LOD resources could be utilized in generating semantically represented and machine-interpretable topics. The proposed approach extracts a significantly rich set of information from the posts in terms of relating them to web-resources which themselves are related to other resources via data and object properties. We demonstrate the benefits of using LOD and ontologies while identifying as well as utilizing the topics. During the identification phase, we were able to identify candidate elements and resolve their types. The ontologically represented topics consisting of entities enabled processing opportunities that revealed information about collections of microposts that are not readily observable even if each post were to be manually inspected. Also, we notice an increase in the quality of the generated topics over time thanks to the efforts related to the continued expansion and correction of LOD resources.
Our main goal of producing machine-interpretable topics was for their utilization in further processing. We demonstrated such utilization with several examples of various levels of complexity, where information that is not readily available in the original posts is revealed. A user evaluation (with 81.0% precision and F 1 of 93.3%) and regularly performed manual inspections show that the identified topics are relevant. In summary, we are encouraged by the results we obtained and list many research opportunities to improve the topic identification approach and process topics in general and in domain-specific manners.  Table. The namespace prefixes that are utilized in Topico and referred to in this paper. (PNG) S2 Table. Percentage of tweets in the post sets that produce the vertices (topic elements), edges (co-occurring elements), and topics. This table shows the percentage of tweets in the post sets that produce the vertices (topic elements), edges (co-occurring elements), and topics. The columns labeled Before and Pruned show the impact of pruning the graph. The columns labeled Topic show how many were retained in the topic. (PNG) S3 Table. The intervals within the datasets that were used for evaluating topics. (PNG)