News angles are approaches to journalism content often used to provide a way to present a new report from an event. One particular type of news angle is the LOCAL news angle where a local news outlet focuses on an event by emphasising a local connection. Knowledge graphs are most often used to represent knowledge about a particular entity in the form of relationships to other entities. In this paper we see how we can extract a knowledge sub graph containing entities and relevant relationships that are connected to the locality of a news outlet. The purpose of this graph is to use it for automated journalism or as an aid for the journalist to find local connections to an event, as well as how the local connection relate to the event. We call such a graph a relevance knowledge graph. An algorithm for extracting such a graph from a linked data source like DBpedia is presented and examples of the use of a relevance graph in a LOCAL news angle context are provided.
Knowledge graphs have found many applications lately [
Here we explore the knowledge graphs in relation to the concept of news angles. A news angle is a concept from journalism, emphasising a view on an event that address and focus on particular aspects of that event [11, pp.781–795]. News angles has been formalised for computational purposes in [12]. One such news angle is the Local or Proximity news angle [12, 13]. This is a news angle often applied by local news papers when relating to events outside of their normal geographical scope. The idea is that the particular event becomes interesting from a local perspective due to its reference to an entity that has local relevance.
Motta et al. [12] focuses on the place of an event in their description of the local/proximity angle. Thus only a place indicates the event’s relevance for some local news outlet. In their example formulations in common logic, events that occur in locally relevant local places, but also locally relevant remote places are included. However, they also suggest that events that include local actors could be locally relevant. We suggest that all sorts of entities that are somehow locally relevant would justify a local/proximity angle on an event. This includes places, persons, organisations, natural phenomena, cultural artifacts etc.
Motta et al. suggest using spatial reasoning on places involved in events to enable the local angle. They, however, did not specify how this could be done. Performing a relevance search for each and other event at the time needed would be costly without pre-existing information structures. The semantic web is such an information structure, but also this is too large to enable eficient searches. We suggest an approach where the search for a local news angle should use a knowledge graph that includes places and other entities that have relevance for the local news paper’s area of interest as well as known relationships between them. We call such a graph a relevance graph. The local news paper’s relevance graph may be used when checking external news items for entities involved in external events.
An example would be a local singer mentioned as participating in the choir in a Eurovision Song Contest performance. That would permit a local news angle on the song contest. If the local singer was included in the news paper’s relevance graph, external news items including the singer’s name should be flagged for the journalists. In addition the journalist could be informed about the relationships path from the news paper’s central entity to the singer.
One may of course construct relevance graph manually, but a less costly way is perhaps to use existing semantic web resources to construct this knowledge graph, at least for an initial version. Later on, journalists may themselves be able to extend the graph by supporting tools, combined with repeated construction to include the latest information from the semantic web.
In the remaining of this paper we address the automatic construction of such a relevance graph, using entities and relationships from DBpedia. We suggest various approaches to measuring relevance of an entity for another related entity, formalising what it means to be of relevance for the local news angle. Then we present an algorithm for constructing the relevance graph. We present some examples and discuss the experiences with those, particularly addressing the problems with using a source like DBpedia. We then describe related work relating to sub graph extraction and graph embeddings, before we conclude with ideas for some future work.
We start out with some suggested relevance measures. To define the various measures we need some notation. Assume we have a knowledge graph with entities = {1, 2, . . . , }. Further assume we have a collection of binary relations between the entities in , ℛ = {1, 2, . . . , } where ⊆ × , = 1, . . . , .
The first approach to finding the most important neighbours in the knowledge graph is based on the assumption that neighbours that are connected through many relations should have a connectivity value relative to the total number of connections. Now assume that the set () = ∪=1{|(, ) ∈ } is the left neighbour set of , and () = ∪=1{|(, ) ∈ } is the right neighbour set of . Then the neighbour set () = () ∪ (). For each of the elements ∈ () we can assign a connectivity score
(|) = |{|(, ) ∈ , = 1, . . . , }| + |{|(, ) ∈ , = 1, . . . , }| A normalized relevance score is further given by (1) (2) (3) (4) (5) (6) (|) = ̂︀
(|) max=1 (|) The remaining will have a relevance score of 0.0 to .
One way of considering relevance is that relations through which has fewer neighbours should contribute more to the total score for those neighbours. Relationships with fewer examples may convey more information, and may therefore be seen as more important. This can be done by weighting the score from a neighbour from a particular relation by 1 divided by the number of neighbours originating from that relation, i.e., 1/|{|(, ) ∈ }|. An alternative approach is to weigh by 1/ log2(|{|(, ) ∈ }| + 1). This gives the following weighted relevance scores 1 and 2 for entities in ()
1(|) = ∑︁ 1 ((, )) =1 |{|(, ) ∈ }|
+ ∑︁ 1 ((, ))
=1 |{|(, ) ∈ }| 2(|) = ∑︁ 1 ((, )) =1 log2(|{|(, ) ∈ }| + 1)
+ ∑︁ 1 ((, )) =1 log2(|{|(, ) ∈ }| + 1) and their normalized versions ̂︂1(|) = ̂︂2(|) =
1(|) max=1 1(|)
2(|) max=1 2(|)
As we know, relevance among entities in the real world is not a symmetric property. For example, in a local news paper, the local buildings, places, and persons born in the community are important, but the country of the locality may not be so interesting to cover. This even though the country and the locality may have a lot of connections in the overarching knowledge graph. One way of handling this is by considering what we call (neighbour) reciprocal relevance between the two entities. We consider two such dependent connectivity measures 1 and 2. They are based on the normalized expressions found in equations 2, 5 and 6: 1(|) = (|) · ̂︂1(|)
̂︀ 2(|) = (|) · ̂︂2(|)
̂︀ For practical use the normalised versions are ̂︁1(|) =
1(|) max=1 1(|) ̂︁2(|) = 2(|) (7) (8) (9) (10) max=1 2(|) These expressions have the efect that the relevance is higher if there are relations in which is one of relatively few entities related to . That is, as compared to the situation where is one of many entities related to .
As mentioned, to build a relevance graph for a local news paper could be a manual task, but also quite demanding for a single journalist. The web, and in particular Wikipedia, consists of exactly the kind of knowledge about relationships between entities that is of importance in the news context. However, these web pages sufer from • incompleteness of relevance links • the emphasis on particular content domains • links to irrelevant or marginally relevant entities • erroneous links These problems is of course due to the fact that these web pages are produced and edited on a voluntary basis. Anyhow, the content and links found in Wikipedia and further structured in DBpedia, is a good starting point for exploring algorithms for constructing relevance graphs.
The algorithm, found in Algorithm 1, for building the graphs starts out with the central entity, the node we want to build a relevance graph for. We expand the graph by first adding entities one link away from the central entity, then those that are two links away, etc., etc. At each new level we first remove entities that are irrelevant by certain heuristics (unspecified in algorithm, see page 6). Next, we compute the relevance value for the entity in mind by computing its relevance to the originating entity at the previous level (Eq 9 or Eq 10) and multiply with that entity’s relevance to the central entity. The central entity has relevance 1.0. At each level we expand only those entities that have relevance value above some boundary . If there are several links to an entity from the level below the entity’s relevance is increased slightly combining scores from all links (done by combining two scores with the commutative and transitive formula = 1.0 − (1.0 − 1)(1.0 − 2)). To keep the graph at a reasonable size we take away those entities that have a score lower than . After having chosen the entities to be included in the graph as nodes, we add to the graph all relationships found in DBpedia among entities in the graph.
Algorithm 1 Get Relevance Graph Nodes Input c_ent: a DBpedia entity to be the centre of the relevance graph.
← ← { ← 1 while ≤ ← ∅ − 1 ← { new Node(c_ent,level = 0, score = 1.0) }
max_level do for ∈ − 1 do ← { : (, r, ) ∈ DBpedia ∨ (, r, ) ∈ DBpedia} : ∈ such that level() = − 1 ∧ score() ≥ } ◁ Nodes at previous level that are relevant enough to be expanded
◁ Remove immediately irrelevant entities. See page 6. for ∈ do ◁ for all candidate entities = reciprocal relevance score for e in relation to (Eq 9 or Eq 10) = · score() if Node(e,level(e),score(e)) ∈
score(e) ← adjust(score(e),) else
← end if ∪ new Node(e,level = i, score = ) ◁ has e already been created ◁ adjust relevance score for e
◁ new node ◁ Initial node ◁ Initialise graph ◁ Augment graph to level ◁ The augmenting set ◁ Find nodes at level i end for end for ← ∪ end while for Node(e,level(e),score(e)) ∈ do if score(e) <
← ∖ {Node(e,level(e),rel(e))} end if end for return
◁ remove nodes with too small score We have explored this algorithm on two locations. One is Valdres in Norway which is a rural, and mountainous district with six municipalities (Norwegian: kommune) and a population of about 12,000. We experimented with various relevance boundaries for including entities in the graph, and with = 0.15 and = 0.05 (see Algorithm 1) we got a graph with 929 entities. The graph was visualised with a rdf-visualiser1. The size and complexity of the graph is indicated in Figure 1. The central entity is found amplified in the middle of the figure. We 1https://issemantic.net/rdf-visualizer also note that there is a hub structure in the graph. For example, there is a hub around the node with a ring, i.e., the node for the entity dbpedia:/Eastern_Norway.
The inclusion of the entity for Eastern Norway illustrates an issue with use of Wikipedia as a source. The relevance measures would find some relevance to Eastern Norway. But if you know more about Norway, you will understand that this entity is not relevant for the local newspaper of Valdres. Eastern Norway includes Valdres, but like Scandinavia or Norway it relates to a too big area to be considered an important key word for Valdres. For a large area like Norway this is solved as it removed because it passes the limit of 1.000 in going or out going links. This heuristic is used in Algorithm 1 together with some other heuristics to immediately exclude entities from the construction of the relevance graph. The heuristics include • Entities which have more than 1.000 in going or out going relationships • Entities that indicates a collection, ending in for example “_of_Norway” (Municipalities_of_Norway) • Files (start with “File:”) • Timelines (start with “Timeline_of”) • Years Notice that we have to, to some extent, specialize these heuristics to a Norwegian context.
As we see, our heuristics does not exclude Eastern_Norway, as its number of links are fewer than 1.000. To find a useful boundary will be a matter of trial and error, or we need to create more complex methods for assessing immediate exclusion of an entity.
We also explored the small town of Clifden, Ireland as a the central entity. Figure 2 shows the central node and the highest scoring nodes at level one.
We have suggested the use of such graphs to support automatic journalism algorithms that detect events that may have a local connection. Further, we envision that the detected event can be framed and presented in the context of a local news paper (using a LOCAL news angle). In our first version we detect external events by performing entity linking on texts from news papers that are collected continuously. To find a relevant event is rare for such rural areas as Clifden or Valdres, but it does happen. In a newspaper in a neighbouring county (“The Connaught Telegraph”) to Clifden we discovered this story:
Achill RNLI assist with medical evacuation on Inishturk. Achill Island RNLI responded to a request for assistance with the medical evacuation of a patient on Inishturk on bank holiday Monday. ... etc.
The connection form Clifden is through the “Wild Atlantic Way” as both Clifden and Inishturk are included in the DBpedia description of this tourist attraction (Figure 3). It is moreover interesting to observe that there is also a connection through an entitity named “South Inishturk”. This is a small island close to Clifden that also has the name Inishturk. The description of this island mentions that it is has the same name as the larger Inishturk, including a link to its Wikipedia page. This illustrate yet an important problem with using DBpedia for such a purpose. The links are general, as most of them are wikiPageWikiLink relationships which DBpedia uses to represent links from one Wikipedia page to another. Further, the authors of Wikipedia are not good at categorising the entities, so entities that have no useful meaning in this context do also show up. We already saw this with all the entities in DBpedia that are mainly lists of other entities of a particular type, although no type from the Wikipedia ontology is used neither for the list or the listed entities.
Any texts, not only news reports, could be used as source for identifying local connections. For example, this imagined tweet text does result in the connections from the relevance graph as seen in Figure 4:
The Mountjoy Prison is going to be renovated after decades of protests, the minister for justice said during his visit at the Connemara National Park.
The idea of extracting subgraphs from open KGs like Wikidata for news-related purposes has already been explored in the literature, although none of the proposals address news angles and locality. An early example is [14], which represents news messages as small KGs with edges extracted from Freebase, in order to support content-based news recommendation.
[15] also represents news articles as small anchor graphs, with edges and K-hop neighbourhops extracted from Wikidata by a sub-graph extractor trained jointly with a news recommender. The purpose is to provide explainable recommendations in real time.
Context-Aware Graph Embeddings (CAGE) [16] also extracts Wikidata sub-graphs to represent news texts. News-text graph embeddings are concatenated with embeddings that represent user behaviours to ofer new recommendations that take short-term user preferences into account.
[17] instead extracts entities and edges from Wikidata to enrich graphs that represent users short- and long-term interests. Entities extracted from candidate news texts are then compared with entities in interest graphs to predict which articles a user may find interesting.
The aim of KLG-GAT [18] is to enhance fact checking and verification by connecting nodes that represent the claim and evidence sentences with triples from Wikidata5M, a subset of Wikidata. A multi-head graph attention network (GAT) is trained to provide input to a claim classifier.
Finally, NewsLink [19] aims to support robust and explainable query answering by representing both news texts and user queries as connected sub-graphs extracted from an open KG.
A broader and more detailed review of these and other uses of KGs for the news is presented in [20]. In relation to the work presented here, some of them (e.g., [15, 16, 17]) demonstrate learned sub-graph extraction with Graph Neural Networks (GNNs) [21, 22] pointing forward to alternative ways to suggest relevance graphs in future work.
Many of the above sub-graph embedding techniques use graph-embedding techniques to make the structural and symbolic content in knowledge graphs content available for numerical and sub-symbolic deep learning. Graph embeddings represent graph nodes or sub-graphs as vectors in low-dimensional semantic spaces, so that semantic similarity is reflected in spatial distance between nodes and/or that semantic relations between nodes are reflected in spatial directions and distances.
Early graph-embedding approaches exploited the translation characteristic of knowledge graphs to generate node embeddings, so that an edge type used to connect nodes should be represented as a translation in semantic space from its head (or source) node to its tail (or target) [23]. Subsequent approaches were proposed, e.g., to account for one-to-many and many-to-many relations between nodes and for semantically diferent relations between nodes [24, 25, 26].
Inspired by word embeddings [27], other approaches used parameterised random walks to represent graphs as sets of walks [28, 29], similar to NL sentences, used as inputs to word embedding models like CBOW, Skip-Gram [30] or GloVe [27].
More recently, dedicated Graph Neural Networks (GNNs) have been proposed [21, 22], inspired by deep neural networks used for image and text analysis, including Graph Attention Networks (GATs), Convolutional Graph Networks (CGMs) and Recurrent Graph Networks (RCNs). For example, in an RCN, each node is assigned an initial vector which is updated in subsequent iterations based on the node’s neighbouring vectors until the network converges. The final node vectors can be fed, for example into a node classifier whose loss is used to train the RCN.
In relation to the work presented here, graph embeddings ofers an alternative way of discovering local connections to events. Computationally, graph embeddings can be combined with scaling similarity search algorithms such as Meta’s FAISS [31] and HSNW [32] to find connections eficiently. However, they are unlikely to be faster than simple entity look-up in precomputed relevance graphs, as we propose in this paper. And, unlike the graph-traversal based technique we have described, embedding approaches are unable to explain local connections by showing how an event is related to the location of interest. They nevertheless ofer a way of suggestion connections that are not represented in existing KGs and that should therefore be explored in future work.
In this paper we have shown how to construct a relevance graph, i.e., a knowledge graph intended to support identification of textual information sources that may be of use for automated journalism. The relevance graph extracts from a larger linked data set, here DBpedia, a smaller graph of entities that is considered to be of relevance for a news outlet, for example, a smaller local news paper.
The algorithm has been shown to work on a couple of examples, but its application still has to face the problems caused by the lack of structure in Wikipedia, and with that DBpedia. To solve the challenges of DBpedia, one may go to a diferent source, like Wikidata. This could be supplied with various approaches from network analysis. For example, measures of centrality could perhaps be used to identify irrelevant entities that are highly connected.
It is also pertinent to explore diferent graph embedding techniques. They can be used to assess similarity between entities. Techniques like e-walks and p-walks [33] can be used as alternatives or complement to the relevance measures described in Section 2. A similar approach can be use to identify new connections. Instead of computing the distance between two known nodes, we can search for the closest nodes in the low-dimensional space to build the location graph. However, some close nodes may lack an explicit connection in the knowledge base, which could be resolved by link prediction models like TransE [23].
Acknowledgments This research is funded by the Norwegian Research Council’s IKTPLUSS programme as part of the News Angler project (grant number 275872) and by MediaFutures partners and the Research Council of Norway as part of MediaFutures: Research Centre for Responsible Media Technology & Innovation (grant number 309339).