Explainable Tsetlin Machine framework for fake news detection with credibility score assessment
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Explainable Tsetlin Machine framework for fake news detection with
credibility score assessment
Bimal Bhattarai Ole-Christoffer Granmo Lei Jiao
University of Agder University of Agder University of Agder
bimal.bhattarai@uia.no ole.granmo@uia.no lei.jiao@uia.no
Abstract particularly problematic as they seek to deceive
people for political and financial gain (Gottfried
The proliferation of fake news, i.e., news in-
and Shearer, 2016).
arXiv:2105.09114v1 [cs.CL] 19 May 2021
tentionally spread for misinformation, poses
a threat to individuals and society. Despite In recent years, we have witnessed extensive
various fact-checking websites such as Politi- growth of fake news in social media, spread across
Fact, robust detection techniques are required news blogs, Twitter, and other social platforms.
to deal with the increase in fake news. Sev- At present, most online misinformation is manu-
eral deep learning models show promising re- ally written (Vargo et al., 2018). However, natural
sults for fake news classification, however, language models like GPT-3 enable automatic gen-
their black-box nature makes it difficult to ex-
eration of realistic-looking fake news, which may
plain their classification decisions and quality-
assure the models. We here address this prob- accelerate future growth. Such growth is problem-
lem by proposing a novel interpretable fake atic as most people nowadays digest news stories
news detection framework based on the re- from social media and news blogs (Allcott and
cently introduced Tsetlin Machine (TM). In Gentzkow, 2017). Indeed, the spread of fake news
brief, we utilize the conjunctive clauses of poses a severe threat to journalism, individuals, and
the TM to capture lexical and semantic prop- society. It has the potential of breaking societal be-
erties of both true and fake news text. Fur-
lief systems and encourages biases and false hopes.
ther, we use the clause ensembles to calcu-
late the credibility of fake news. For eval-
For instance, fake news related to religion or gender
uation, we conduct experiments on two pub- inequality can produce harmful prejudices. Fake
licly available datasets, PolitiFact and Gossip- news can also trigger violence or conflict. When
Cop, and demonstrate that the TM framework people are frequently exposed to fake news, they
significantly outperforms previously published tend to distrust real news, affecting their ability
baselines by at least 5% in terms of accuracy, to distinguish between truth and untruth. To re-
with the added benefit of an interpretable logic- duce these negative impacts, it is critical to develop
based representation. Further, our approach
methods that can automatically expose fake news.
provides higher F1-score than BERT and XL-
Net, however, we obtain slightly lower accu- Fake news detection introduces various challeng-
racy. We finally present a case study on our ing research problems. By design, fake news in-
model’s explainability, demonstrating how it tentionally deceives the recipient. It is therefore
decomposes into meaningful words and their difficult to detect fake news based on linguistic con-
negations. tent, style, and diverseness. For example, fake news
may narrate actual events and context to support
1 Introduction
false claims (Feng et al., 2012). Thus, other than
Social media platforms on the Internet have be- hand-crafted and data-specific features, we need to
come an integral part of everyday life, and more employ a knowledge base of linguistic patterns for
and more people obtain news from social rather effective detection. Training fake news classifiers
than traditional media, such as newspapers. Key on crowdsourced data may further provide a poor
drivers include cost-effectiveness, freedom to com- fit for future news events. Fake news is emerging
ment, and shareability among friends. Despite continuously, quickly rendering previous textual
these advantages, social media exposes people to content obsolete. Accordingly, some studies, such
abundant misinformation. Fake news stories are as (Guo et al., 2018), have tried to incorporate so-cial context and hierarchical neural networks using Lukoianova, 2015). Many studies such as (Wang,
attention to uncover more lasting semantic patterns. 2017) (Mitra and Gilbert, 2015) (Shu et al., 2020a)
Despite significant advances in deep learning- incorporate publicly available datasets, providing a
based techniques for fake news detection, few ap- basis for detailed analysis of fake news and detec-
proaches can explain their classification decisions. tion methods.
Currently, knowledge on the dynamics underlying Recently, deep learning-based latent representa-
fake news is lacking. Thus, explaining why cer- tion of text has significantly improved fake news
tain news items are considered fake may uncover classification accuracy (Karimi and Tang, 2019).
new understanding. Making the reasons for deci- However, the latent representations are generally
sions transparent also facilitates discovering and difficult to interpret, providing limited insight into
rectifying model weaknesses. To the best of our the nature of fake news. In (Castillo et al., 2011),
knowledge, previous research has not yet addressed the authors introduced features based on social con-
explainability in fake news detection. text, obtained from the profiles of users and their
Paper contributions: In this paper, we propose activity patterns. Other approaches depend upon so-
an explainable framework for fake news detection cial platform-specific features such as likes, tweets,
built using the Tsetlin Machine (TM). Our TM- and retweets for supervised learning (Volkova et al.,
based framework captures the frequent patterns that 2017) (Ruchansky et al., 2017).
characterize real news, distilling both linguistic and While the progress on detecting fake news has
semantic patterns unique for fake news stories. The been significant, limited effort has been directed to-
resulting model constitutes a global interpretation wards interpretability. Existing deep learning meth-
of fake news. To provide a more refined view on ods generally extract features to train classifiers
individual fake news (local interpretability), we without giving any interpretable explanation. This
also propose a credibility score for measuring the lack of transparency makes them black boxes when
credibility of news. Finally, our framework allows it comes to understandability (Du et al., 2019). In
the practitioner to see what features are critical for this paper, we propose a novel fake news detection
making news fake (global interpretability). method that builds upon the TM (Granmo, 2018).
Paper organization: The remainder of the pa- The TM is a recent approach to pattern classifica-
per is organized as follows. In Section 2, we briefly tion, regression, and novelty detection (Abeyrathna
summarize related work. We then detail our ex- et al., 2021; Yadav et al., 2021; Abeyrathna et al.,
plainable framework in Section 3. In Section 4, the 2020; Bhattarai et al., 2021) that attempts to bridge
datasets and experimental configurations are pre- the present gap between interpretability and accu-
sented. We analyze our empirical results in Section racy in state-of-the-art machine learning. By using
5, before we conclude the work in the last section. the AND-rules of the TM to capture lexical and
semantic properties of fake news, our aim is to
2 Related Work improve the performance of fake news detection.
More importantly, our framework is intrinsically
The problem of detecting deception is not new to interpretable, both globally, at the model level, and
natural language processing (Vargo et al., 2018). locally for the individual false news predictions.
Significant application domains include detecting
false online advertising, fake consumer reviews, 3 Explainable Fake News Detection
and spam emails (Ott et al., 2011) (Zhang and Framework
Guan, 2008). The detection of fake news focuses
on uncovering spread of misleading news articles In this section, we present the details of our TM-
(Zhou and Zafarani, 2020) (Zhou et al., 2019). Typ- based fake news detection framework.
ical detection techniques use either text-based lin-
guistic features (Potthast et al., 2018) or visual 3.1 TM Architecture
features (Gupta et al., 2013). Overall, fake news The TM consists of a team of two-action Tsetlin
detection methods fall into two groups: knowledge- Automata (TAs) with 2N states. Each TA performs
based models based on fact-checking news arti- either action “Include” (in State 1 to N ) or action
cles using external sources (WuYou et al., 2014), “Exclude” (in State N to 2N ). Jointly, the actions
and style-based models, which leverage linguis- specify a pattern recognition task. Action “Include”
tic features capturing writing style (Rubin and incorporates a specific sub-pattern, while actionText Input
Input
Binary Encoded Input
Real News Fake News
Clauses
Summation
Threshold Argmax
Output y
Figure 1: Interpretable Tsetlin machine architecture.
“Exclude” rejects the sub-pattern. The TAs are up- clauses. Positive polarity is assigned to one half of
dated based on iterative feedback in the form of the clauses, denoted by Cj+ . These are to capture
rewards or penalties. The feedback signals how patterns for the target class (y = 0 for T M1 and
well the pattern recognition task is solved, with the y = 1 for T M2 ). Negative polarity is assigned to
intent of maximizing classification accuracy. Re- the other half, denoted by Cj− . Negative polarity
wards reinforce the actions performed by the TAs, clauses are to capture patterns for the non-target
while penalties suppress the actions. class (y = 1 for T M1 and y = 0 for T M2 ). In
By orchestrating the TA team with rewards and effect, the positive polarity clauses vote for the in-
penalties, a TM can capture frequent and discrimi- put belonging to the target class, while negative
native patterns from labeled training data. Each pat- polarity clauses vote against the target class.
tern is represented as a conjunctive clause in propo- Any clause Cj+ for a certain target class is
sitional logic, based on the human-interpretable formed by ANDing a subset L+ j ⊆ L of the lit-
disjunctive normal form (Valiant, 1984). That is, eral set, written as:
the TM produces AND-rules formed as conjunc-
Cj+ (X) =
^ Y
tions of propositional variables and their negations. lk = lk , (1)
lk ∈L+ lk ∈L+
A two-class TM structure is shown in Figure 1, j j
consisting of two separate TMs (T M1 and T M2 ).
As seen in the Input-step, each TM takes a Boolean where j = (1, . . . , m/2) denotes the clause in-
(propositional) vector X = (x1 , . . . , xo ), xk ∈ dex, and the superscript decides the polarity of the
{0, 1}, k ∈ {1, . . . , o} as input, which is ob- clause. For instance, the clause Cj+ (X) = x1 x2
tained by booleanizing the text input as suggested consists of the literals L+ j = {x1 , x2 }, and it out-
in (Berge et al., 2019; Yadav et al., 2021). That puts 1 if both of the literals are 1-valued. Similarly,
is, the text input is modelled as a set of words, we have clauses Cj− for the non-target class.
with each Boolean input signaling the presence The final classification decision is done after
or absence of a specific word. From the input summing up the clause outputs (Summation-step
vector, we obtain 2o literals L = (l1 , l2 , . . . , l2o ). in the figure). That is, the negative outputs are
The literals consist of the inputs xk and their substracted from the positive outputs. Employing a
negated counterparts x̄k = ¬xk = 1 − xk , i.e., single TM, the sum is then thresholded using the
L = (x1 , . . . , xn , ¬x1 , . . . , ¬xn ). unit step function u, u(v) = 1 if v ≥ 0 else 0, as
shown in Eq. (2):
A TM forms patterns using m conjunctive
clauses Cj (Figure 1 – Clauses). How the pat-
m/2 m/2
terns relate to the two output classes (y = 0 and Cj+ (X) − Cj− (X) .
X X
ŷ = u (2)
y = 1) is captured by assigning polarities to the j=1 j=1Type I (a) feedback Type I (b) feedback Type II feedback
Include Exclude Include Exclude Include Exclude
time stamp for time stamp for time stamp for
Type I (a) feedback Type I (b) feedback Type II feedback
Include Exclude Include Exclude Include Exclude
time stamp for time stamp for time stamp for
Figure 2: Visualization of Tsetlin machine learning.
The summation of clause outputs produces when y = 1 and to negative polarity clauses when
an adaptive ensemble effect designed to y = 0. Type II feedback, on the other hand, in-
help dealing with noisy and diverse data creases the discriminating power of the patterns,
(Granmo, 2018). For example, the classifier suppresses false positives, and makes clauses eval-
ŷ = u (x1 x̄2 + x̄1 x2 − x1 x2 − x̄1 x̄2 ) captures the uate to 0. Type II feedback is given to positive
XOR-relation. polarity clauses when y = 0 and to negative polar-
With multi-class problems, the classification is ity clauses when y = 1.
instead performed using the argmax operator, as The feedback is further regulated by the sum of
shown in the figure. Then the target class of the votes v for each output class. That is, the voting
TM with the largest vote sum is given as output. sum is compared against a voting margin T , which
is employed to guide distinct clauses to learn differ-
3.2 Interpretable Learning Process ent sub-patterns. The details of the learning process
As introduced briefly above, the conjunctive can be found in (Granmo, 2018).
clauses in a TM are formed by a collection of We use the following sentence as an example
TAs. Each TA decides whether to “Include” or to show how the inference and learning process
“Exclude” a certain literal in a specific clause based can be interpreted: X= [“Building a wall on the
on reinforcement, i.e., rewards, penalties, or inac- U.S-Mexico border will take literally years.”], with
tion feedback. The reinforcement depends on six output target “true news” i.e., y = 1. First, the in-
factors: (1) target output (y = 0 or y = 1), (2) put is tokenized and negated: X = [“build” =
clause polarity, (3) clause output (Cj = 0 or 1), (4) 1, “¬build” = 0, “wall” = 1, “¬wall” =
literals value (i.e., x = 1, and ¬x = 0), (5) vote 0, “U.S − M exico” = 1, “¬U.S − M exico” =
sum, and (6) the current state of the TA. 0, “take” = 1, “¬take” = 0, “years” = 1,
The TM learning process carefully guides the “¬years” = 0]. We consider two positive polar-
TAs to converge towards optimal decisions. To ity clauses and one negative polarity one (i.e., C1+ ,
this end, the TM organizes the feedback it gives C2+ , and C1− ) to show different feedback conditions
to the TAs into two feedback types. Type I feed- occurring while learning the propositional rules.
back is designed to produce frequent patterns, com- The learning dynamics are illustrated in Figure 2.
bat false negatives, and make clauses evaluate to 1. The upper part of the figure shows the current con-
Type I feedback is given to positive polarity clauses figuration of clauses and TA states for three dif-ferent scenarios. On the left side of the vertical 3.3 Problem Statement for Fake News
bar, the literals are included in the clause, and on Detection
the right side of the bar, the literals are excluded. We now proceed with defining the fake news detec-
The upper part also shows how the three different tion problem formally. Let A be a news article with
kinds of feedback modifies the states of the TAs, ei- si sentences, where i = 1 to S and S is the total
ther reinforcing “Include” or “Exclude”. The lower number of sentences. Each sentence can be written
part depicts the new clause configurations, after the as si = (w1i , w2i , . . . , wW
i ), consisting of W words.
reinforcement has been persistently applied. Given the booleanized input vector X ∈ {0, 1}o ,
the fake news classification is a Boolean classifica-
For the first time stamp in the figure, we as- tion problem, where we predict whether the news
sume the clauses are initialized randomly as fol- article A is fake or not, i.e., F : X → y ∈ {0, 1}.
lows: C1+ = (build ∧ wall ∧ years), C2+ = We also aim to learn the credibility of news by
(build∧wall ∧years∧¬take∧¬U.S −M exico), formulating a TM-based credibility score that mea-
and C1− = (build ∧ take ∧ years). Since clause sures how check-worthy the news is. With this
C1+ consists of non-negated literals of value 1 addition, the classifier function can be written as
in the input X, the output becomes 1 (because F : X → (y, Q), where y is a classification output
C1+ = 1 ∧ 1 ∧ 1). This invokes the condition and Q is the credibility score.
(C1+ = 1 and y = 1). So, as shown in upper left 3.4 Credibility Assessment
of Figure 2, when the actual output is 1 and the
clause output is 1, Type I (a) feedback is used to re- The credibility score can be calculated as fol-
inforce the Include action. Therefore, literals such lows. Firstly, the TM architecture in Section 3.1 is
as U.S −M exico and take are eventually included, slightly tweaked for the score generation. In the
because they appear in X. This process makes the architecture, instead of identifying the class with
clause C1+ eventually have one sub-pattern cap- the largest voting sum using Argmax, we obtain the
tured for y = 1, as shown on the lower left side of raw score from both the output classes. The raw
Figure 2. I.e., take and U.S − M exico have been score is generated using clauses, which represent
included in the clause. frequent lexical and semantic patterns that charac-
terizes the respective classes. Therefore, the score
contains information on how the input resembles
Similarly, in C2+ , the clause output is 0 because the patterns captured by the clauses. We thus use
of the negated inputs ¬take and ¬U.S − M exico this score to measure the credibility of news. For
(i.e., C2+ = 1 ∧ 1 ∧ 1 ∧ 0 ∧ 0). This invokes the instance, consider the vote ratios of 2:1 and 10:1
condition (C2+ = 0 and y = 1). Here, so-called for two classes (i.e., real vs. fake news) obtained
Type I (b) feedback is used to reinforce Exclude from two different inputs. Then the first class wins
actions. In this example, literals such as ¬take the majority vote in both cases, however, the 10:1
and ¬U.S − M exico are eventually excluded from vote ratio suggests higher credibility than for the
C2+ as shown in the middle part of Figure 2, thus input that gave a ratio of 2:1.
establishing another sub-pattern for C2+ . To normalize the credibility score so that it falls
between 0 and 1, we apply the logistic function to
the formula in Eq. (2) as shown in Eq. (3).
Type II feedback (right side of figure) only oc-
curs when y = 0 (for positive polarity clauses) 1
Qi = F −v T
. (3)
and y = 1 (for negative polarity clauses). We 1 + exp−k(vi i )
here use the negative polarity clause C1− to demon-
strate the effect of Type II feedback. This type of Above, Qi is the credibility score of the ith sample,
feedback is triggered when the clause output is 1. and k is the logistic growth rate. viF and viT are
The goal is now to make the output of the affected the total sum of votes collected by clauses for fake
clause change from 1 to 0 by adding 0-valued in- and true news respectively. Here, k is a user config-
puts. Type II feedback can be observed on the urable parameter deciding the slope of the function.
right side of Figure 2, where all negated literals For example, consider the scores (43, -47) and (124,
finally are included in C1− to ensure that the clause -177), both predicting fake class. The credibility
outputs 0. scores obtained from Eq. (3) with k = 0.012 are0.74 and 0.97, which shows that the second news Dataset #Real #Fake #Total
is more credible than the first one. PolitiFact 563 391 954
GossipCop 15,338 4,895 20,233
4 Experiment Setup
Table 1: Dataset statistics.
In this section, we present the experiment configu-
rations we use to evaluate our proposed TM frame- • RST (Rubin et al., 2015): Rhetorical Structure
work. Theory (RST) represents the relationship be-
tween words in a document by building a tree
4.1 Datasets structure. It extracts the news style features
from a bag-of-words by mapping them into a
For evaluation, we adopt the publicly available
latent feature representation.
fake news detection data repository FakeNewsNet
(Shu et al., 2017). The repository consists of news • LIWC (Pennebaker et al., 2015): Linguistic
content from different fact-checking websites, in- Inquiry and Word Count (LIWC) is used to ex-
cluding social context and dynamic information. tract and learn features from psycholinguistic
We here use news content annotated with labels and deception categories.
by professional journalists from the fact-checking
websites PolitiFact and GossipCop. PolitiFact is • HAN (Yang et al., 2016): HAN uses a hier-
a fact-checking website that focuses on U.S polit- archical attention neural network (HAN) for
ical news. We extract news articles published till embedding word-level attention on each sen-
2017. GossipCop focuses on fact-checking of en- tence and sentence-level attention on news
tertainment news collected from various medias. content, for fake news detection.
While GossipCop has more fake news articles than
PolitiFact, PolitiFact is more balanced, as shown in • CNN-text (Kim, 2014): CNN-text utilizes
Table 1. convolutional neural network (CNN) with pre-
trained word vectors for sentence-level clas-
4.2 Preprocessing sification. The model can capture different
granularities of text features from news arti-
The relevant news content and the associated la- cles via multiple convolution filters.
bels, i.e., True or False news, are extracted from
both of the datasets. The preprocessing steps in- • LSTM-ATT (Lin et al., 2019): LSTM-ATT
clude cleaning, tokenization, lemmatization, and utilizes long short term memory (LSTM) with
feature selection. The cleaning is done by remov- an attention mechanism. The model takes
ing less important information such as hyperlinks, a 300-dimensional vector representation of
stop words, punctuation, and emojis. The datasets news articles as input to a two-layer LSTM
are then booleanized into a sparse matrix for the for fake news detection.
TM to process. The matrix is obtained by build-
ing a vocabulary consisting of all unique words • RoBERTa-MWSS: The Multiple Sources of
in the dataset, followed by encoding the input as Weak Social Supervision (MWSS) approach,
a set of words using that vocabulary. To reduce built upon RoBERTa (Liu et al., 2019), was
the sparseness of the input, we adopt two methods: proposed in (Shu et al., 2020b).
1) Chi-squared test statistics as a feature selection
technique, and 2) selecting the most frequent liter- • BERT (Devlin et al., 2019): Bidirectional
als from the dataset. The experiment is performed Encoder Representations from Transformers
using both methods, and the best results are in- (BERT) is a Transformer-based model which
cluded. For the PolitiFact and GossipCop datasets, contains an encoder with 12 transformer
we selected the 20 000 and 25 000 most significant blocks, self-attention heads and hidden shape
features, respectively. size of 768.
• XLNet (Yang et al., 2019): XLNet is a gener-
4.3 Baseline
alized autoregressive pretraining model that
We first summarize the state-of-the-art of the fake integrates autoencoding and a segment recur-
news detection approaches. rence mechanism from transformers.In addition, we compare our results with other is arguably because HAN can capture the syntactic
machine learning baseline models such as logis- and semantic rules using hierarchical attention to
tic regression (LR), naive Bayes, support vector detect fake news. Similarly, the LIWC performs
machines (SVM), and random forest (RF). For a better than RST. One possible explanation for this
fair comparison, we select the methods that only is that LIWC can capture the linguistic features in
extract textual features from news articles. The news articles based on words that denote psycholin-
performance for these baseline models has been re- guistic characteristics. The LSTM-ATT, which has
ported in (Shu et al., 2019) and (Shu et al., 2020a). extensive preprocessing using count features and
sentiment features along with hyperparameter tun-
4.4 Training and Testing ing (Lin et al., 2019), has similar performance com-
We use a random training-test split of 75%/25%. pared with HAN in PolitiFact, however, outper-
The classification results is obtained on the test forms HAN on GossipCop. One reason for this can
set. The process is repeated five times, and the be that the attention mechanism is able to capture
average accuracy and F1 scores are reported. To the relevant representation of the input. In addition,
provide robust results, we calculated an ensemble we see that machine learning models such as LR,
average by first take the average of 50 stable epochs, SVM, and Naive Bayes are not very effective in
followed by taking the average of the resulting five neither of the datasets.
averages. We run our TM for 200 epochs with
The recent transformer-based models BERT and
hyperparameter configuration of 10 000 clauses, a
XLNet outperform our model in terms of accuracy.
threshold T of 200, and sensitivity s of 25.0. To
Our TM-based approach obtains slightly lower ac-
ensure a fair comparison of the results obtained
curacy, achieving 87.1% for PolitiFact and 84.2%
by other machine learning algorithms, we use the
for GossipCop. However, the F1 score for poli-
Python Scikit-learn framework that includes LR,
tiFact is 0.90, whereas for GossipCop it is 0.89,
SVM and Naı̈ve Bayes, using default parameter
which is marginally better than XLNet and BERT.
settings.
Thus, we achieve state-of-the-art performance with
5 Results and Discussion respect to F1-score overall, and with respect to
accuracy when comparing with interpretable ap-
We now compare our framework with above men- proaches. Since GossipCop is unbalanced with
tioned baseline models. sample ratio of 3:1 for real and fake news, we sub-
mit that the F1 score is a more appropriate perfor-
5.1 Comparison with State-of-the-Art
mance measure than accuracy. Overall, our model
The experimental results are shown in Table 2. is much simpler than the deep learning models be-
Clearly, our model outperforms several of the other cause we do not use any pre-trained embeddings
baseline models in terms of accuracy and F1 score. for preprocessing. This also helps in making our
model more transparent, interpretable, and explain-
PolitiFact GossipCop able.
Models
Acc. F1 Acc. F1
RST 0.607 0.569 0.531 0.512 The credibility assessment is performed as de-
LIWC 0.769 0.818 0.736 0.572 scribed in Subsection 3.4. The Figs. 3 and 4 show
HAN 0.837 0.860 0.742 0.672 the credibility scores for a fake news sample from
CNN-text 0.653 0.760 0.739 0.569
LSTM-ATT 0.833 0.836 0.793 0.798 PolitiFact and GossipCop, resepectively. As seen,
LR 0.642 0.633 0.648 0.646 the fake news can be ranked quite distinctly. This
SVM 0.580 0.659 0.497 0.595 facilitates manual checking according to credibility,
Naı̈ve Bayes 0.617 0.651 0.624 0.649
RoBERTa-MWSS 0.825 0.805 0.803 0.807 allowing users to focus on the fake news articles
BERT 0.88 0.87 0.85 0.79 with highest scores. However, when we observe the
XLNet 0.895 0.90 0.855 0.78
TM 0.871 0.901 0.842 0.896
soft scores e.g. obtained from XLNet, most of the
samples are given rather extreme scores. That is,
Table 2: Performance comparison of our model with 8 classifications are in general submitted with very
baseline models. high confidence. This is in contrast to the more
cautious and diverse credibility scores produced by
One can further observe that HAN outperforms TM clause voting. If we for instance set a credi-
LIWC, CNN-text, and RST for both datasets. This bility score threshold of 0.8 in TM, we narrow the1.0 negated features from the negative polarity clauses
of other classes, as well as the the plain features
0.8 from positive polarity clauses for the intended class.
Taking the positive- and negative polarity claues
0.6
Credibility
together, one gets stronger discrimination power.
Table 3 and Table 4 exemplify the above be-
0.4 havior for fake news detection. To showcase the
explainability of our approach, we list the ten most
0.2 captured plain and negated words per class, for
both datasets. We observe that negated literals ap-
0.0 pear quite frequently in the clauses. This allows the
0 100 200 300 400 trained TM to represent the class by the features
Samples
that characterize the class as well as the features
Figure 3: Credibility assessment for fake news in Poli- that contrast it from the other class. TM clauses that
tiFact. contain negated and unnegated inputs are termed
non-monotone clauses, and these are crucial for
1.0 human commonsense reasoning, as explored in
“Non-monotonic reasoning” (Reiter, 1987).
0.8
Also observe that the TM clauses include both
descriptive words and discriminative words. For ex-
0.6
Credibility
ample, in PolitiFact, the Fake class captures words
like “trump”, while the True class captures the
0.4
negated version, i.e., “¬trump”. However, this does
not mean that all the news related to Trump are
0.2
Fake. Actually, it means that if we break down the
clauses into literals, we see “trump” as a descrip-
0.0
tor/discriminator for the Fake News class. There-
0 1000 2000 3000 4000 5000
Samples fore, most of the clauses captured the word “trump”
in plain and in negated form (for the True news
Figure 4: Credibility assessment for fake news in Gos- class). However, one single word cannot typically
sipCop. produce an accurate classification decision. It is
selection of fake news down to around 300 out of the joint contribution of the literals in a clause that
400 for PolitiFact and to around 2 800 out of 4 895 contributes to the high accuracy. Hence, we have
for GossipCop. to look at all word pattern captured by the clauses.
When an input is passed into the trained TM, the
5.2 Case Study: Interpretability clauses from both classes that capture the input
We now investigate the interpretability of our TM word pattern are activated, to vote for their respec-
approach in fake news classification by analyzing tive class. Finally, the classification is made based
the words captured by the clauses for both true on the total votes gathered by both classes for the
and fake news. In brief, the clauses capture two particular input.
different types of literals:
6 Conclusions
• Plain literals, which are plain words from the
target class. In this paper, we propose an explainable and inter-
pretable Tsetlin Machine (TM) framework for fake
• Negated literals, which are negated words
news classification. Our TM framework employs
from the other classes.
clauses to capture the lexical and semantic features
The TM utilizes both plain and negated word pat- based on word patterns in a document. We also ex-
terns for classification. When we analyze the plain the transparent TM learning of clauses from
clauses, we see that most of the sub-patterns cap- labelled text. The extensive experimental evalua-
tured consist of negated literals. This helps TM tions demonstrate the effectiveness of our model
make decisions robustly because it can use both on real-world datasets over various baselines. OurPolitiFact
Fake True
Plain times Negated times Plain times Negated times
trump 297 candidate 529 congress 136 trump 1252
said 290 debate 413 tax 104 profession 1226
comment 112 civil 410 support 70 navigate 1223
donald 110 reform 369 senate 64 hackings 1218
story 78 congress 365 president 60 reported 1216
medium 63 iraq 361 economic 57 arrest 1222
president 48 lawsuit 351 americans 49 camps 1206
reported 45 secretary 348 candidate 48 investigation 1159
investigation 38 tax 332 debate 44 medium 1152
domain 34 economy 321 federal 41 domain 1153
Table 3: Top ten Literals captured by clauses of TM for PolitiFact.
GossipCop
Fake True
Plain times Negated times Plain times Negated times
source 357 stream 794 season 150 insider 918
insider 152 aggregate 767 show 103 source 802
rumors 86 bold 723 series 79 hollywood 802
hollywood 80 refreshing 722 like 78 radar 646
gossip 49 castmates 721 feature 70 cop 588
relationship 37 judgment 720 video 44 publication 579
claim 33 prank 719 said 33 exclusively 551
split 32 poised 718 sexual 32 rumor 537
radar 32 resilient 714 notification 25 recalls 535
magazine 30 predicted 714 character 25 kardashian 525
Table 4: Top ten Literals captured by clauses of TM for GossipCop.
results show that our approach is competitive with and fake news in the 2016 election. NBER Working
far more complex and non-transparent methods, in- Paper Series.
cluding BERT and XLNet. In addition, we demon- Geir Thore Berge, Ole-Christoffer Granmo, T. Tveit,
strate how fake news can be ranked according to a M. Goodwin, Lei Jiao, and Bernt Viggo Matheussen.
credibility score based on classification confidence. 2019. Using the tsetlin machine to learn human-
We finally demonstrate the explainability of our interpretable rules for high-accuracy text catego-
rization with medical applications. IEEE Access,
model using a case study. In our future work, we 7:115134–115146.
intend to go beyond using pure text features, also
incorporating spatio-temporal and other meta-data Bimal Bhattarai, Ole-Christoffer Granmo, and L. Jiao.
features available from social media content, for 2021. Measuring the Novelty of Natural Language
Text Using the Conjunctive Clauses of a Tsetlin Ma-
potentially improved accuracy. chine Text Classifier. In 13th International Confer-
ence on Agents and Artificial Intelligence (ICAART
2021). INSTICC.
References
C. Castillo, Marcelo Mendoza, and B. Poblete. 2011.
K. Darshana Abeyrathna, Bimal Bhattarai, Morten Information credibility on twitter. In WWW.
Goodwin, Saeed Gorji, Ole-Christoffer Granmo,
Lei Jiao, Rupsa Saha, and Rohan K. Yadav. 2021. J. Devlin, Ming-Wei Chang, Kenton Lee, and Kristina
Massively Parallel and Asynchronous Tsetlin Ma- Toutanova. 2019. Bert: Pre-training of deep bidirec-
chine Architecture Supporting Almost Constant- tional transformers for language understanding. In
Time Scaling. In The Thirty-eighth International NAACL-HLT.
Conference on Machine Learning (ICML 2021).
ICML. Mengnan Du, Ninghao Liu, and Xia Hu. 2019. Tech-
niques for interpretable machine learning. Commun.
K. Darshana Abeyrathna, Ole-Christoffer Granmo, ACM, 63:68–77.
Xuan Zhang, Lei Jiao, and Morten Goodwin.
2020. The Regression Tsetlin Machine - A Novel Song Feng, Ritwik Banerjee, and Yejin Choi. 2012.
Approach to Interpretable Non-Linear Regression. Syntactic stylometry for deception detection. In
Philosophical Transactions of the Royal Society A, ACL.
378.
Jeffrey A. Gottfried and E. Shearer. 2016. News use
H. Allcott and Matthew Gentzkow. 2017. Social media across social media platforms 2016.Ole-Christoffer Granmo. 2018. The Tsetlin machine - N. Ruchansky, Sungyong Seo, and Y. Liu. 2017. Csi: A
a game theoretic bandit driven approach to optimal hybrid deep model for fake news detection. Proceed-
pattern recognition with propositional logic. ArXiv, ings of the 2017 ACM on Conference on Information
abs/1804.01508. and Knowledge Management.
Han Guo, J. Cao, Y. Zhang, Junbo Guo, and J. Li. Kai Shu, Limeng Cui, Suhang Wang, D. Lee, and Huan
2018. Rumor detection with hierarchical social at- Liu. 2019. defend: Explainable fake news detec-
tention network. Proceedings of the 27th ACM Inter- tion. Proceedings of the 25th ACM SIGKDD In-
national Conference on Information and Knowledge ternational Conference on Knowledge Discovery &
Management. Data Mining.
A. Gupta, H. Lamba, P. Kumaraguru, and A. Joshi.
Kai Shu, Deepak Mahudeswaran, Suhang Wang, Dong-
2013. Faking sandy: characterizing and identifying
won Lee, and Huan Liu. 2020a. FakeNewsNet: A
fake images on twitter during hurricane sandy. In
data repository with news content, social context,
WWW ’13 Companion.
and spatiotemporal information for studying fake
Hamid Karimi and Jiliang Tang. 2019. Learning hierar- news on social media. Big Data, 8(3):171–188.
chical discourse-level structure for fake news detec-
tion. In Proceedings of the 2019 Conference of the Kai Shu, Amy Sliva, Suhang Wang, Jiliang Tang, and
North. Association for Computational Linguistics. Huan Liu. 2017. Fake news detection on social me-
dia: A data mining perspective. ACM SIGKDD Ex-
Yoon Kim. 2014. Convolutional neural networks for plorations Newsletter, 19(1):22–36.
sentence classification. In EMNLP.
Kai Shu, Guoqing Zheng, Yichuan Li, Subhabrata
J. Lin, Glenna Tremblay-Taylor, Guanyi Mou, Di You, Mukherjee, Ahmed Hassan Awadallah, Scott W.
and Kyumin Lee. 2019. Detecting fake news arti- Ruston, and Huan Liu. 2020b. Leveraging multi-
cles. 2019 IEEE International Conference on Big source weak social supervision for early detection
Data (Big Data), pages 3021–3025. of fake news. ArXiv, abs/2004.01732.
Y. Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Leslie G Valiant. 1984. A Theory of the Learnable.
Joshi, Danqi Chen, Omer Levy, M. Lewis, Luke Communications of the ACM, 27(11):1134–1142.
Zettlemoyer, and Veselin Stoyanov. 2019. Roberta:
A robustly optimized bert pretraining approach. Chris J. Vargo, Lei Guo, and Michelle A. Amazeen.
ArXiv, abs/1907.11692. 2018. The agenda-setting power of fake news: A
T. Mitra and E. Gilbert. 2015. Credbank: A large-scale big data analysis of the online media landscape from
social media corpus with associated credibility anno- 2014 to 2016. New Media & Society, 20:2028 –
tations. In ICWSM. 2049.
Myle Ott, Yejin Choi, Claire Cardie, and Jeffrey T. Han- Svitlana Volkova, Kyle Shaffer, Jin Yea Jang, and
cock. 2011. Finding deceptive opinion spam by any Nathan Oken Hodas. 2017. Separating facts from
stretch of the imagination. In The 49th Annual Meet- fiction: Linguistic models to classify suspicious and
ing of the Association for Computational Linguis- trusted news posts on twitter. In ACL.
tics: Human Language Technologies, Proceedings
of the Conference, 19-24 June, 2011, Portland, Ore- William Yang Wang. 2017. ”liar, liar pants on fire”: A
gon, USA, pages 309–319. The Association for Com- new benchmark dataset for fake news detection. In
puter Linguistics. ACL.
J. Pennebaker, Ryan L. Boyd, K. Jordan, and Kate G. WuYou, K. AgarwalPankaj, LiChengkai, Yang-jun,
Blackburn. 2015. The development and psychomet- and YuCong. 2014. Toward computational fact-
ric properties of liwc2015. checking. In VLDB 2014.
Martin Potthast, Johannes Kiesel, K. Reinartz, Janek Rohan Yadav, Lei Jiao, Ole-Christoffer Granmo, and
Bevendorff, and Benno Stein. 2018. A stylometric Morten Goodwin. 2021. Human-level interpretable
inquiry into hyperpartisan and fake news. In ACL. learning for aspect-based sentiment analysis. In
AAAI.
R Reiter. 1987. Nonmonotonic reasoning. Annual Re-
view of Computer Science, 2(1):147–186.
Z. Yang, Zihang Dai, Yiming Yang, J. Carbonell,
Victoria L. Rubin, Niall Conroy, and Yimin Chen. R. Salakhutdinov, and Quoc V. Le. 2019. Xlnet:
2015. Towards news verification: Deception detec- Generalized autoregressive pretraining for language
tion methods for news discourse. understanding. In NeurIPS.
Victoria L. Rubin and Tatiana Lukoianova. 2015. Truth Zichao Yang, Diyi Yang, Chris Dyer, X. He, Alex
and deception at the rhetorical structure level. Jour- Smola, and E. Hovy. 2016. Hierarchical atten-
nal of the Association for Information Science and tion networks for document classification. In HLT-
Technology, 66. NAACL.Linfeng Zhang and Yong Guan. 2008. Detecting click fraud in pay-per-click streams of online advertising networks. 2008 The 28th International Conference on Distributed Computing Systems, pages 77–84. X. Zhou, R. Zafarani, Kai Shu, and H. Liu. 2019. Fake news: Fundamental theories, detection strate- gies and challenges. Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining. Xinyi Zhou and Reza Zafarani. 2020. A survey of fake news. ACM Computing Surveys, 53(5):1–40.
You can also read
