1 Introduction

With the widespread popularization of the Internet, it becomes easier and more convenient for people to get news from the Internet than other traditional media. Unfortunately, open Internet fuels the spread of a great many fake news without effective supervision. Fake news are news articles that are intentionally and verifiably false, and could mislead readers [AG17a]. With characteristics of low cost, easy access, and rapid dissemination, fake news can easily mislead public opinion, also disturb the social order, damage the credibility of social media, infringe the interests of the parties and cause the crisis of confidence [VRA18, SCV+17]. We all know how it has occurred and exerted an influence in the past 2016 US presidential elections [AG17a]. Hence, it is important and valuable to develop methods for detecting fake news.

Most existing works on fake news detection are based on styles, focusing on capturing the writing style of news content as features to classify news articles [GM17, Gil17, Wan17, JLY17]. Although they can be effective, these approaches cannot explain what is fake in the target news article. On the other hand, knowledge based (or content based) fake news detection, which is also known as fact checking [SSW+17], is more promising, as the detection is based on content rather than style. Existing content based approaches focus on path reachability trying to find a path in an existing knowledge graph [PVGPW17, PCE+17] for a given triple [LCSR+15, SFMC17, SW16]. However, there are a few limitations of the existing content-based approaches, which lead to the following research questions:

RQ1: What happens if we do not have a knowledge graph in the first place, but only have articles? For a fake news topic, it is likely that at the beginning we do not have the knowledge graph to rely on for fact-checking. Our idea is either to construct knowledge graphs based on (true and fake) news articles bases, or to utilize related sub-graphs from open knowledge graphs. In the former case, we can also construct two knowledge graphs on the same topic: one is based on fake news articles and the other one based on true news articles. It should be noted that fake news articles are also available in online fake news web sites, such as ‘the Onion’, which often provide different categories of fake news articles. In the latter case, we could extract the sub-graph centered on the background topic of news articles from the open knowledge graph. Hence, we construct an external knowledge graph for these news articles based on facts related to the background topic in DBpedia datasetFootnote 1.

RQ2: Can we use incomplete and imprecise knowledge graphs for fake news detection? All computational knowledge-based approaches mainly focus on simple common relations between entities, such as “country”, “child”, “employerOf”. And the knowledge graphs they use are too incomplete and imprecise to cover the complex relations that appeared in fake news articles. For example, the triple (Anthony Weiner, cooperate with, FBI) extracted from a news article has the entities of “Anthony Weiner” and “FBI”, and the relation of “cooperate with”. The entities are easily found in open knowledge but the relation is not. In this paper, our idea is to make use of knowledge graph embedding for computing semantic similarities, so as to accommodate incomplete and imprecise knowledge graphs. As far as we know, this is the first work on this direction. We use a basic knowledge graph embedding model, namely TransE  [BUGD+13] , to test the potential of knowledge graph embedding methods in content based fake news detection.

RQ3: How can we use Knowledge Graph Embedding for content based fake news detection? We firstly propose an approach to utilizing TransE [BUGD+13] to train a single model on a given knowledge graph, such as a subset of an open knowledge graph, or one that is constructed based on some news articles. Secondly, we propose a approach to generating a binary TransE model (B-TransE) which combines a negative model with a positive single model. Furthermore, in order to improve the performance, we also propose a hybrid approach to using a fusion strategy to combine the feature vectors produced by the models above.

Our major contributions of this paper are summarized as follows:

  • To the best of our knowledge, we are the first to propose the approach of content based fake news detection by making use of incomplete and imprecise knowledge graphs.

  • We proposed a few approaches to exploit knowledge graph embedding to facilitate content based fake news detection.

  • Our experiments show that our binary model approach outperforms our single model approach, and that our hybrid approach improve the performance of fake news detection.

  • Our experiments show that our approaches outperform the Knowledge Stream approach in the test datasets.

2 Related Works

2.1 Fake News Detection

An effective approach is of prime importance for the success of fake news detection that has been a big challenge in recent years. Generally, those approaches can be categorized as knowledge-based and style-based.

Knowledge-Based. The most straightforward way to detect fake news is to check the truthfulness of the statements claimed in news content. Knowledge-based approaches are also known as fact checking. The expert-oriented approaches, such as SnopesFootnote 2, mainly rely on human experts working in specific fields to help decision making. The crowdsourcing-oriented approaches, such as FiskkitFootnote 3 where normal people can annotate the accuracy of news content, utilize the wisdom of crowd to help check the accuracy of the news articles. The computational-oriented approaches can automatically check whether the given claims have reachable paths or could be inferred in existing knowledge graphs. Ciampaglia et al. [LCSR+15] take fact-checking as a problem of finding shortest paths between concepts in a knowledge graph; they propose a metric to assess the truth of a statement by analyzing path lengths between the concepts in question. Shiralkar et al. [SFMC17] propose a novel method called"Knowledge Stream(KS)" and a fact-checking algorithm called Relational Knowledge Linker that verifies a claim based on the single shortest, semantically related path in KG. Shi et al. [SW16] view fake news detection as a link prediction task, and present a discriminative path-based method that incorporates connectivity, type information and predicate interactions.

Style-Based. Style-based approaches attempt to capture the writing style of news content. Mykhailo Granik et al. [GM17] find that there are some similarity between fake news and spam email, such as they often have a lot of grammatical mistakes, try to affect reader’s opinion on some topics in manipulative way and use similar limited set of words. So they apply a simple approach for fake news detection using naive Bayes classifier due to those similarity. Gilda [Gil17] applies term frequency-inverse document frequency (TF-IDF) of bi-grams and probabilistic context free grammar (PCFG) detection and test the dataset on multiple classification algorithms. Wange [Wan17] investigates automatic fake news detection based on surface-level linguistic patterns and design a novel, hybrid convolutional neural network to integrate speaker related metadata with text. Jiang et al. [JLY17] find that some key words tend to appear frequently in the micro-blog rumor. They analyze the text syntactical structure features and presents a simple way of rumor detection based on LanguageTool.

2.2 Knowledge Graph Embedding

Bordes et al. [BUGD+13] propose a method, named TransE, which models relationships by interpreting them as translations operating on the low-dimensional embeddings of the entities. TransE is very efficient while achieving state-of-the-art predictive performance, but it does not perform well in interpret such properties as reflexive, one-to-many, many-to-one, and many-to-many. So, Wang et al. [WZFC14] propose TransH which models a relation as a hyperplane together with a translation operation on it. Lin et al. [LLS+15] propose TransR to build entity and relation embeddings in separate entity space and relation spaces. TransR learns embeddings by first projecting entities from entity space to corresponding relation space and then building translations between projected entities. Ji et al. [JHX+15] propose a model named TransD, which uses two vectors to represent a named symbol object (entity and relation), and the first one represents the meaning of a(n) entity (relation), the other one is used to construct mapping matrix dynamically.

3 Basic Notions

In this section we introduce some basic notions related to content-based classification of news articles with external knowledge.

A knowledge graph KG describes entities and the relations between them. It can be formalised as \( KG = \{E, R, S\}\), where E denotes the set of entities, R the set of relations and S the triple set. An article base AB is a set of news articles for each of which we have a title, a full content text and an annotation of true or fake. A knowledge graph may be a readily available for fact checking, such as DBpedia, or one needs to construct one from an article base.

In this paper, we use the knowledge graph embedding (KGE) method TransE to facilitate fake news detection. Typical knowledge graph completion algorithms are based on knowledge graph embedding (KGE). The idea of embedding is to represent an entity as a k-dimensional vector h (or t ) and defines a scoring function \(f_r\) (h, t ) to measure the plausibility of the triplet (h, r, t) in the embedding space. The representations of entities and relations are obtained by minimising a global loss function involving all entities and relations. Different KGE algorithms often differ in their scoring function, transformation and loss function. When a knowledge graph is converted into vector space, more semantic computations can be applied than just reasoning and querying.

The task of fact checking is to check if a target triple (h, r, t) is true based on a given knowledge graph. The task of content based fake news detection (or simply fake news detection), is to check if a target news article is true based on its title and content, as well as some related knowledge graph.

4 Our Approach

4.1 Framework Overview

To detect whether a news article is true or not, and to answer our research questions as outlined in Sect. 1, we propose a solution which uses, a tool to produce knowledge graphs (KG), a single B-TransE model, a binary TransE model and finally hybrid approaches. Firstly, we generate background knowledge by producing three different KG. This part addresses RQ1 and RQ2. Then we use a B-TransE model to build entity and relation embedding in low-dimensional vector space and detect whether the news article is true or not. We test a single TransE model and a binary TransE model and thus answer RQ3. Finally, we use some hybrid approaches to improve detection performance.

For the task of background knowledge generation, we consider three types of KGs: one is based on fake news article base; one is based on open KG, such as DBpedia, a crowd-sourced community effort to extract structured information from Wikipedia; one is based on true news article base from reliable news agencies.

The external KG extracted from open knowledge graph includes two parts: \(KG_1 = \{E_1,R_1,S_1\}\) based on entities from fake article base and \(KG_2 = \{E_2,R_2,S_2\}\) centered on the topic of news articles. These are further described in Sect. 5.2.

External KGs such as DBpedia are excellent for general knowledge facts, such as (Barack Obama, birthPlace, Hawaii). However, they are incomplete and imprecise as such KGs do not contain enough relations to represent current events, as the latter are generated daily. An example of such a relation is (Anthony Weiner, cooperate with, FBI), which is not contained in DBpedia. Despite this, in Sect. 5, we show that an incomplete and imprecise external open KG can perform well on the task of fake news detection.

The entities and the relation from the example above, however, can easily be extracted from an article on the topic. In order to be able to assess news items as true or fake, we propose an approach which uses external knowledge generated from real world news articles. We propose using a set of true and a set of fake articles to generate two models: \(\mathbf M \) and \(\mathbf M '\) as described in Sect. 3. We summarize these articles, as using the full article text causes redundancies and increases runtime. We further explore the performance of our approach, using only external knowledge from article bases, in order to answer the question what happens when we do not have a KG in the first place, but only news articles. In Fig. 1 we outline the methods used to generate a KG from an article base.

Fig. 1.
figure 1

Triple extraction from an article base

Full size image

To construct KG from news articles, we start with a set of news articles and use OpenIEFootnote 4 to extract triples first. However, OpenIE does not perform well in triple extraction of news, so we propose some methods to improve the quality of the triples, including Stanford NERFootnote 5 and others which are further discussed in Sect. 5.2. We then perform entity alignment and obtain the triples which constitute our article based KG.

Once we have generated our three external KG, we use TransE to train a single model on each of them and compare their performance. To the best of our knowledge, this is the first work to apply knowledge graph embedding for fake news detection. Thus we use the basic TransE model. Since all the translation-based models aim to represent entities and relations in a vector space and there is no great difference between these models on our dataset, we choose the most basic model TransE. The single model is further described in Sect. 4.2 and an outline of its usage can be seen in Fig. 2. Our results, presented in Sect. 5, show that the external open KG has the best performance.

Fig. 2.
figure 2

Single TransE model

Full size image

Then, we explore what happens when we combine a negative single model and a positive single model. The binary TransE (B-TransE) model is further described in Sect. 4.3 and an outline of its usage can be seen in Fig. 3. In Sect. 5 we then show that binary models perform better than single ones.

Fig. 3.
figure 3

Binary TransE model

Full size image

Finally, we use a hybrid approach using an early fusion strategy that combines the feature vectors produces by the models above in order to improve detection performance. Further details of this approach are in Sect. 4.4.

4.2 Single TransE Model

To judge whether a given news article is true or fake through a knowledge graph, we extract triples from the news article and represent the triples in vector space, so that we can judge whether the news article is true or fake by the vectors. We use a Knowledge Graph to train a TransE model, which represents triples as vectors, and we name our method Single TransE Model.

In the Single TransE Model, we define TransE model as \(\mathbf M \), and a triple based on \(\mathbf M \) as \((\mathbf h ,\mathbf t , \mathbf r )\). We denote the triples extracted from one news item as \(\mathbf TS \), so each triple is defined as \(triple_{i}=(\mathbf h _{i},\mathbf t _{i}, \mathbf r _{i})\), where i means the index of the triple in \(\mathbf TS \). We represent one news item as \(\mathbf{N }=\{\mathbf{TS }, \mathbf{M }\}\).

To classify one news item, we calculate the bias of each triple in \(\mathbf{TS }\). The bias of \(triple_i\) is defined as

$$\begin{aligned} f_{b}(triple_{i}) = ||\mathbf h _{i} + \mathbf r _{i} - \mathbf t _{i}||_{2}^{2} \end{aligned}$$
(1)

Then we use these biases to classify the news item through a classifier. There are two ways we use these biases to do classification.

Avg Bias Classification. For the first one, we use average bias of a triple set to classify the news item through a classifier and name it Avg Bias Classification. The average bias of a triple set is defined as

$$\begin{aligned} f_{avgB}(TS)=\frac{\sum _{i=1}^{n}f_{b}(triple_i)}{|\mathbf TS |} \end{aligned}$$
(2)

where the \(|\mathbf TS |\) refers to the size of the triple set.

Max Bias Classification. The second one, the Max Bias Classification uses the max bias of a triple set to judge whether a news item is true or fake. The max bias of a triple set is defined as

$$\begin{aligned} f_{maxB}(TS)=f_{b}(triple_{max}) \end{aligned}$$
(3)

Where max refers to the index of the triple whose bias is the maximum.

4.3 B-TransE Model

The single TransE model sometimes is not good enough, since there are some true triples whose biases are large on both the true single model and the fake single model, so that these news items would be incorrectly classified as fake news if we use just a single true TransE model.

To solve this problem, we propose to train two models, one model is trained based on the triples extracted from fake news and another is trained based on the triples extracted from true news, so that we can do classification by comparing the biases of the true model and the biases on the fake model. We name it B-TransE model:

  • the model based on true news is defined as \(\mathbf M \), and a triple based on \(\mathbf M \) is defined as \((\mathbf h ,\mathbf t , \mathbf r )\)

  • the model based on fake news is defined as \(\mathbf M '\), and a triple based on \(\mathbf M '\) is defined as \((\mathbf h ',\mathbf t ', \mathbf r ')\)

In the B-TransE Model, we represent one news item as \(\mathbf{N }=\{\mathbf{TS }, \mathbf{TS }', \mathbf{M }, \mathbf{M }'\}\), \(\mathbf TS \) refers to triple set extracted from the news based on \(\mathbf M \) and each triple is defined as \(triple_{i}=(\mathbf h _{i},\mathbf t _{i}, \mathbf r _{i})\), and \(\mathbf TS '\) refers to triple set based on \(\mathbf M '\) and each triple is \(triple_{i}'=(\mathbf h _{i}',\mathbf t _{i}', \mathbf r _{i}')\), where i refers to the index of the triple in each triple set.

We define the bias of \(triple_i\) and \(triple_i'\) as

$$\begin{aligned} f_{b}(triple_{i}) = ||\mathbf h _{i} + \mathbf r _{i} - \mathbf t _{i}||_{2}^{2} \end{aligned}$$
(4)
$$\begin{aligned} f_{b}(triple_{i}') = ||\mathbf h _{i}' + \mathbf r _{i}' - \mathbf t _{i}'||_{2}^{2}\end{aligned}$$
(5)

To judge whether a news item is true or fake, we propose two classify functions and do some experiments to verify the efficiency of each method. Max Bias Classify The first way, we use max bias on true single model and max bias on fake single model to do classification. And the Max Bias Classify function is defined as

$$\begin{aligned} f_{mc}(N)=0, if f_{b}(triple_{max})<f_{b}(triple_{max}') \end{aligned}$$
(6)
$$\begin{aligned} f_{mc}(N) = 1, otherwise \end{aligned}$$
(7)

where \(f_{mc}(N)=0\) means the news item is true, and \(f_{mc}(N) = 1\) means it is fake. Avg Bias Classify The another way, we use average bias on true single model and average bias on fake single model to do classification. And the Avg Bias Classify function is defined as

$$\begin{aligned} f_{ac}(N)=0, if f_{avgB}(TS)<f_{avgB}(TS') \end{aligned}$$
(8)
$$\begin{aligned} f_{ac}(N) = 1, otherwise \end{aligned}$$
(9)

where \(f_{ac}(N)=0\) means the news item is true, and \(f_{ac}(N) = 1\) means it is fake.

4.4 Hybrid Approaches

To improve the detection performance, we need a fusion strategy to combine the feature vectors from different models. The fusion strategy we use is known as early (feature-level) fusion, which means integrating different features first and using those integrated-features do classification.

In this part, we use the bias vector of the triple, whose bias is the maximum, rather than bias to do classifiction. The bias vector is defined as

$$\begin{aligned} \mathbf v _i = \mathbf h _{i} + \mathbf r _{i} - \mathbf t _{i} \end{aligned}$$
(10)

The max bias vector is defined as \(Vec_{max}\). We use two different feature vectors:

  1. 1.

    max bias vectors from the model based on true news is defined as \(Vec_{max}\);

  2. 2.

    max bias vectors from the model based on fake news is defined as \(Vec_{max}'\).

The integrated vector is defined as \(V\), so that:

$$\begin{aligned} V=(Vec_{max}, Vec_{max}') \end{aligned}$$
(11)

which means we concatenate two different max bias vectors to get an integrated vector, and we use this vector to do classification.

5 Experiments and Analysis

5.1 Data

Fake and True News Article Bases.

We use two article bases for our experiments: one with fake news and one with news that we regard as true. We use Kaggle’s ‘Getting Real about Fake News’ dataset, which contains news articles on the 2016 US Election, and we select 1,400 of this dataset as our Fake News Article Base (FAB). These articles have been manually labeled as Bias, Conspiracy, Fake, Bull Shit, which we regard as fake. Our True News Article Base (TAB) was produced by using the BBC News, Sky News and The Independent websites to scrape 1,400 news articles, which were on the topic of US Election and were published between 1st January and 31st December 2016. These articles have not been manually labeled, however, for the purposes of our experiments, we regard them as true. The statistics of two article bases are shown in Table 1. We divide each article base into two parts, 1,000 are for training a model and 400 are for testing.

Table 1. Statistics of fake and true news article bases.
Full size table

Knowledge Graphs. We produce three knowledge graphs for our experiments: one named FKG based on FAB, one named D4 (DBpedia 4-hop) from DBpedia, and one named NKG based on TAB.

FKG. FKG\( = \{E_0,R_0,S_0\}\) is constructed using the training set of FAB. FKG has the following characteristics: \(|E_0|\) = 4K entities, \(|R_0|\) = 1.2K relations, \(|S_0|\) = 8K triples.

D4. To build our KG from DBpedia with 4 hops, we use SPARQL query endpoint interfaceFootnote 6 to interview DBpedia dataset online. We selected 4 hops as it provides a good trade-off between coverage and noise level. There is a public SPARQL endpoint over the DBpedia datasetFootnote 7. DB4 includes two parts, they are \(KG_1\) and \(KG_2\). \(KG_1 = \{E_1,R_1,S_1\}\) based on entities from FAB. It has the following characteristics: \(|E_1|\) = 215K entities, \(|S_1|\) = 760K triples. \(KG_2 = \{E_2,R_2,S_2\}\) centered on 2016 US election. We take the entity “United States presidential election 2016"as \(h_0\), extract all triples within four hops. It has the following characteristics: \(|E_2|\) =  132K entities, \(|R_2|\) = 5,211 relations, \(|S_2|\) = 312K triples. The reason we extract 4-hop subgraph is that one more hop produces lots of repetitive triples, and most appear in the 4-hop one. We just need to make sure that we get triples related to the topic even some are not related tightly, which also makes the KG construction easier and general.

NKG. We produce NKG\( = \{E_3,R_3,S_3\}\) using the training set of TAB. NKG has the following characteristics: \(|E_3|\) = 15K entities, \(|R_3|\) = 3,751 relations, \(|S_3|\) = 19k triples.

5.2 Experiment Setup

Article Summarization. We use the titles and the first two sentences of each article to produce the summaries. We did some small-scale experiment, and found that the above summarisation works better than other choices. The intuition behind is that the main message of a news article is often contained in the title and the first two sentences.

Knowledge Extraction. We use an extraction model to extract train triples from 1k fake news, which is used to train FML, and extract train triples from 1k true news, which is used to train FML. Simultaneously, we use an extraction model to extract test triple sets from 400 fake news and 400 true news, which means translating each news item into a triple set with a fake or true label. We use OpenIE to perform triple extraction. However, OpenIE does not perform very well on triple extraction from news articles. Thus, we use the following four methods to improve the quality of the entities and relations in the triples extracted:

  • We disambiguate pronouns so that a text such as “The man woke up. He took a shower.” would be transformed to “The man woke up. The man took a shower”. We use Neuralcoref to do this.

  • We use NLTKs WordNetLemmatizer to transform any verbs in the triples to their present tense.

  • We shorten the length of the entities, which is extracted though OpenIE and is named OpenIEEntity. We find out the word which is real entity in the entity extracted though OpenIE and remove other words. Such as “western mainstream media like John Kerry" is shortened to “western mainstream media".

  • We use Stanford NER to extract entities from news, which is named NEREntity. Then align the OpenIEEntities to NEREntities.

To produce the two parts of the external KG from an open knowledge graph, as outlined in Sect. 4.1, we use the following steps:

  1. 1.

    \(KG_1 = \{E_1,R_1,S_1\}\) based on entities from a fake article base. Firstly, to obtain the set of entities \(E_1\) from triples in fake article base. And then, to extract all triples \(S_1\) from the open knowledge graph with these entities as subjects and objects respectively.

  2. 2.

    \(KG_2 = \{E_2,R_2,S_2\}\) centered on the topic of news articles. This sub-KG reflects true statements about the news topic in the real world. We take the entity \(h_0\) that is the most related to the topic as the center, and extract all triples \(S_2\) within a certain number of hops. As shown in Fig. 4, it is a simplified three-hop sub-graph example. Supposing the node “0” to be \(h_0\), firstly, to extract all triples denoted as \(T_1\) that has the formula as \((h_0, r, t)\). Secondly, to extract all triples denoted as \(T_2\) that has the formula as \((h_1, r, t)\), where \(h_1\) refers to an entity in \(T_1\), also one of the nodes “1” in the figure. And the rest can be done by analogy.

Fig. 4.
figure 4

A simplified three-hop sub-graph example

Full size image

Model Generation. We generate three single trained models based on TransE for our experiments: the first model FML using the negative knowledge graph FKG; the second model TML-D4 using the positive knowledge graph D4; the third model TML-NKG using the positive knowledge graph NKG.

FML. The TransE model gets the input of \(S_0\) and automatically produces the trained model FML.

TML-D4. The TransE model gets the input of \(S_1 + S_2\) and automatically produces the trained model TML-D4.

TML-NKG. The TransE algorithm gets the input of \(S_3\) and automatically produces the trained model TML-D4.

5.3 Fake News Detection

Using Single Models.

The results of the single TransE model with different bias function are shown in Table 2, which shows that: (1) TML-D4 model performs the best (in terms of F score) for the fake news detection task. It suggests that using incomplete knowledge graph can still be effective for fake news detection task. (2) FML and TML-NKG model also perform pretty well, which suggests that using imprecise knowledge graphs can also be effective for fake news detection. This also suggests that, if we do not have knowledge graph in the first place, but only have articles, contracting a knowledge graph from articles is a effective method. (3) Max Bias significantly outperforms Avg Bias in terms of F Score. Maybe there are a few true triples in the triple set of one true news, so that the average bias of the triple set becomes smaller. Since not all the triples extracted from one fake news is false, max bias is more useful in fake news detection task. (4) TML-D4 performs a little better than TML-NKG and FML. The results may correlate with the training data of the TransE model: There are 1 K training news of TML-NKG and FML, but there are 132 K entities and 312 K triples of the training data set of TML-D4.

Table 2. Performance of single TransE model.
Full size table

Using B-TransE Model.

The results of B-TransE Model with different bias function are shown in Table 3, from which we observe that the B-TransE Model is better than the Single TransE Model. This suggests thates the approach based on one related knowledge graph is not enough, and that one should combine related knowledge graph with external knowledge graphs.

Table 3. Performance of different models.
Full size table

Hybrid Approaches. In this section, we do experiments on the test sets using the hybrid approach described in Sect. 4.4. Experimental results of combining different models are shown in Table 4. We use vectors from a single TransE model and integrated vectors from a B-TransE Model. The classification we use is SVM [Joa98, SS02], and we choose ‘poly’, ‘linear’ and ‘rbf’ as kernel functions. From Table 4, we can draw a conclusion that: the hybrid approach can further improve the single and binary model approaches.

Table 4. Performance of different models.
Full size table

Knowledge Stream. Finally, we test Knowledge Stream approach [SFMC17] on the 400 true articles and 400 fake articles required that a file exists for each article which contains all of the triples extracted from the given article and with IDs for each entity and relation accordingly. Once these files existed, they were run in Knowledge Stream to produce scores for each triple in each file. Table 5 shows the results of the comparison of the performance of the TransE FML (which is not even the best single model from our approach, as discussed above) and that of Knowledge Stream. From the table we observe that while Knowledge Stream has a very high recall value, TransE outperforms it significantly. Therefore, we conclude that: Our single TransE model is better than Knowledge Stream on the task of fake news detection when the background knowledge graph is constructed from real news articles.

Table 5. Performance of different models.
Full size table

6 Conclusion and Future Work

In this paper, we tackle the problem of content based fake news detection. We have proposed some novel approaches of fake news detection based on incomplete and imprecise knowledge graphs, based on the existing TransE model and our B-TransE model. Our findings suggest that even incomplete and imprecise knowledge graph can help detect fake news.

As for future work, we will explore the following directions: (1) To combine our content based approaches with style-based approaches. (2) To provide explanations for the results fake news detection, even with incomplete and imprecise knowledge graphs. (3) To explore the use of the schema of knowledge graphs as well as approximate reasoning  [PRZ16] and uncertain reasoning [PTRT12, SFP+13, JGC15] in fake news detection.