Fake news refers to deceptive online content and is a problem which causes social harm [1]. Early detection of fake news is therefore a critical but challenging problem. In this paper we attempt to determine if state-of-the-art models, trained on the LIAR dataset [2] can be leveraged to reliably classify short claims according to 6 levels of veracity that range from “True” to “Pants on Fire” (absolute lies). We investigate the application of transformer models BERT [3], RoBERTa [4] and ALBERT [5] that have previously performed significantly well on several natural language processing tasks including text classification. A simple neural network (FcNN) was also used to enhance each model's result by utilising the sources' reputation scores1. We achieved higher accuracy than previous studies that used more data or more complex models. Yet, after evaluating the models' behaviour, numerous flaws appeared. These include bias and the fact that they do not really model veracity which makes them prone to adversarial attacks. We also consider the possibility that language-based, fake news classification, on such short statements is an ill-posed problem.
Social media has made the creation and spreading of information easier, quicker and cheaper
than ever before. This scenario has resulted in an epidemic of what is termed as ‘fake news’
content that deliberately gives false information to deceive and manipulate, often with negative
results [
Since fake news spread fast, early detection is necessary in order to limit the spread and,
consequently, the harm. The use of Machine Learning (ML) techniques for Natural Language
Processing (NLP) is one way to build classifiers that could serve as potential early detectors.
Two distinct approaches that use NLP-based models [
CEUR 1. Model-Based approaches: A ML model is used to find some reliable features in a dataset that correlate to the classification label. 2. Feature-Based approaches: The ML model relies on pre-defined, linguistic or textual feature(s) assumed to indicate deception.
Our approach is model-based, since transformer models are used to learn features that correlate short statements with their respective class labels.
Transformers are deep learning architectures that have changed the face of NLP research in
recent years. A transformer is initially trained on a large corpus of text to build a ‘language
model’. This model represents words as vectors, yet unlike traditional word embeddings (like
Word2Vec), the representation of each word is sensitive to the context within which the word
occurs. The trained transformer can then be used for various NLP tasks, such as Question
Answering, Named Entity Recognition, Text Classification and others [
Google’s BERT [
The LIAR dataset [
Each statement in LIAR also comes with textual metadata including the job title of the speaker,
the speaker’s afiliation, state of origin and the context in which the claim was uttered, (namely
an interview, a debate or another event) [
Among the studies leading to this one, two are very relevant to our approach. Kirilin & Strube
(2018) [
Kirilin & Strube represented the statements and the textual metadata as FastText [
Liu et al used BERT-base to classify the statements and textual metadata (except for the
speaker’s names). BERT’s output-vector and the reputation were then utilized in an attention
network followed by a simple neural network to carry out the classification. This entire layout
was repeated twice. The first one produced a coarse grained (true or false) classification. This
output was passed to the second segment, together with the initial input, to derive a final 6-way
classification [
The mathematical definition of a well-posed problem is attributed to mathematician Jacques
Hadamard [
A problem is well-posed if it satisfies the following 3 criteria [
2. It has only one, uniquely defined solution; and
3. The solution’s behaviour changes continuously with the initial conditions
Thus, an ill-posed problem is a problem that fails one or more of the above criteria [
When considering a few of the short statements from the LIAR dataset, it becomes apparent that many statements can have multiple possible truth-levels based on context, time or who the speaker is. For instance, the veracity of the statement: “I am pro-life, he is not” will depend on the speaker, the subject and their opinions at a given time.
These multiple, possible solutions violate the second rule for a well posed problem, giving a
strong reason to believe that purely language-based classification of fake news is an ill-posed
problem. Another class of problems is called ill-conditioned. Such problems may not satisfy
the definition of ill-posed problems but are considered similarly unstable for practical purposes
since a small change in the input results in a large change in the output [
The principal aim of this research was to build a Fake-News Classifier, that matches, or exceeds, previous classifiers’ accuracy scores on the LIAR dataset. Objectives to reach this aim included: 1. To analyse the efectiveness of transformers, given that they are the current state-of-theart NLP models, for the classification task. 2. To investigate whether better trained or larger transformers can achieve a higher accuracy. 3. To evaluate whether the overall classification result can be enhanced by adding neural network layers that use both the transformer’s output and the source’s reputation score.
Two other objectives were added with the aim of determining the reliability of the resulting models.
4. To investigate the bias, behaviour and learning of the models achieved. 5. To investigate the argument that fake news classification, performed using only NLP (without the classifier having knowledge of the real world), is an ill-posed or ill-conditioned problem.
Three BERT variants were used in order to determine if the diferences in transformer size, pre-training or optimisations matter (as initially hypothesised). These were:
6-way classification for diferent levels of truthfulness was performed on the statements in LIAR without the use of any metadata.
Fine tuning the transformers was performed manually for reasons of limited disk space. 80% of LIAR’s data was used for fine tuning (training), 10% was used for validation and 10% for testing. LIAR comes already split into these segments allowing for a fair comparison with results reported in diferent studies.
BERT and RoBERTa converged after two epochs at learning rates of 1.8−5 and 2.2−5
respectively. ALBERT took 4 epochs at a learning rate of 2.2−5. The batch size for training all models
was 64.
3.1.1. Using the Reputation Score
To use both the statement and the reputation scores for classification we created FcNN (Fully
Connected Neural Network). This was necessary because NLP-transformers are not applicable
for classification with numerical data such as the reputation score [
Each of the transformers produces a classification vector consisting of 6 values in its final layer, which can be extracted programmatically. These are input to the FcNN together with the 6 values of the reputation score (Figure 1) after the latter are normalised (divide by 200, since this is a value close to the largest reputation score). FcNN is then trained at a learning rate of 9-4. To avoid overfitting their result was checked every 500 epochs of training and the best iftting model was used. In every run, the FcNN managed to fit in less than 9000 epochs.
In a separate attempt, FcNN was applied on its own, for classification using only the reputation vector. This provided a baseline against which other models could be compared to (Table 3).
None of the textual metadata was used because a speaker’s name, job or similar details were considered to be unrelated to a statement’s truthfulness. Furthermore, these values either repeat frequently, are often null or not normalized (non-atomic and diferent spelling can be found for the same value). Because of this, we were concerned that it would bias the models unnecessarily.
Relying too heavily on the individual’s reputation may result in labelling liars instead of lies [
This test-set’s 226 statements are truths from liars and lies from mostly-honest speakers. These were chosen by computing each speaker’s honesty ratio P, a measure of how honest a speaker is, based on each speaker’s classification of his or her claims, such that: = 1.5( − ) + ( − ) (1)
The numerical diference of a speaker’s pants-on-fire statements from true ones was multiplied to give it a higher weighting. Speakers with values close to zero (balanced liars) were ignored. Those with scores less than -15 are considered liars so we take their truthful statements. Speakers scoring more than 4 are honest ones, for which we take their lies. The reason for these cut-of points was because speakers with honesty ratio between -15 and 4 were ones with relatively fewer claims. The inequality resulted from the fact that the labels of liars are skewed to begin with (3 false, 1 neutral and 2 true labels). Thus, the set of 226 statements was collected.
While only the FcNN-only model (trained normally on LIAR’s training set) was used to classify these 226 statements, it was expected that even the FcNN models using the transformer’s output will be prone to this same bias, if confirmed.
We also trained the same models on datasets with diferent data quality than that of LIAR and compared the results. This was meant to reveal the efect that the quality of the data has on the models’ learning and also if the models are truly able to learn the intended classification task or not.
For this task, two variations of LIAR were created. The first is called Shufled-LIAR and was obtained by randomly shufling the spoken claims attribute among all entries in the training set, while leaving every other attribute (column) untouched. By having a dataset with randomised text and all other attributes untouched, we can better determine how much the text really afects the result. If the same results on the actual, unshufled set are also achievable on a completely random set, then one can conclude that the results are accidental and hence insignificant.
Additionally, the Cleaned-LIAR dataset was created in order to allow training and testing
on data of better quality (less errors) [
The spelling and grammar of the statements were also corrected manually, under the assumption that since the transformers were pre-trained on good quality text and have a limited vocabulary, classification may receive a boost from these corrections. If accuracy is not improved when training on this set, this may suggest that the flawed entries were responsible for the higher accuracy on the original (unchanged) LIAR.
On Cleaned-LIAR, BERT was trained for 2 epochs at a 2−5 learning rate. RoBERTa and ALBERT were trained at a learning rate of 1.2−5 for 3 and 2 epochs respectively. The models failed to fit for Shufled-LIAR.
If instances of the same basic model produce highly varied classifications when they are trained on data that difers gradually (when tested on the same test data); it may indicate that the problem is ill-conditioned (at least in the way the problem is being treated here).
Five copies of the same transformer were trained with training data that varies proportionally each time. The dataset’s original training and validation portions were joined and their order 2Spelling and other mistakes in LIAR mostly result from how the data was scraped from polifact.com to produce the dataset. randomised. The resulting set was then stratified, splitting it in 5 folds (parts), such that all folds contain virtually an identical number of statements and variety of labels (truth levels). This keeps the label balance identical for each fold and thus for the 5 folds. For each of the five training runs, a diferent combination of 4 folds would be used for training, and the fith would be used for validation. The test set was the same in each of the 5 runs.
This classification was performed with both LIAR and Cleaned-LIAR separately, using the
BERT-base model. Then this was all repeated with RoBERTa-Large. For comparison, the same
procedure was repeated using the two transformers to carry out a 5-way sentiment analysis on
the Stanford Sentiment Treebank (SST-5) dataset [
The same hyperparameters were used to train the transformers (2−5 learning rate, 64 batch size for 2 epochs).
In our evaluation, we managed to achieve a higher accuracy than other results reported in literature. However, all other test results suggest that our models are flawed despite their higher accuracy. This is described in more detail below.
The transformer-only classifiers had a performance similar to Wang’s previous attempts that
utilise statements alone [
All of our Transformer+FcNN models exceeded accuracy results by Kirilin & Strube (2018) and Liu et al (2019) in spite of these studies using more data. This vindicates our decision to avoid using textual metadata. Furthermore, our BERT model performs better than Liu et al’s system, despite it having a far simpler architecture (Figure 1). All transformers produced similar accuracy scores. A bigger or a better trained transformer only marginally improves fake news classification.
When classifying truthful statements from liars and lies from honest speakers, FcNN displays a clear bias caused by utilising reputation. This is clearly visible in the Confusion Matrix found in Table 2.
The fact that the transformers did not manage to properly fine-tune for Shufled-LIAR indicates that the models are correlating some features to the labels, whilst no such correlating feature occurs randomly. However, when trained on the less noisy, Cleaned-LIAR the performance of the transformers, without FcNN was generally poorer (Table 3). RoBERTa is the exception in this case, since the cleaned set resulted in marginally better performance.
This unexpected result raised the question of whether the models are really modelling veracity. A test to this efect was done as described below.
Consider the following true statement that was classified correctly: “One out of every four homeless people on our streets is a veteran.” 27.7% 27.3% 28.2% 48.0% 47.9% 48.6%
Classifier FcNN only BERT-base Only BERT+FcNN RoBERTa Only RoBERTa+FcNN ALBERT Only ALBERT+FcNN
LIAR
A change in the fine-grained classification of the statement is expected if any of the following changes is done: • Negation of the same statement: “One out of every four homeless people on our streets is not a veteran.” • Reducing probability of the statement: “One out of every four homeless people on our streets is a friendly veteran.”; and • Contradiction: “One out of every four homeless people on our streets is not homeless.” Nevertheless, all such modifications are still classified as fully-true, showing that the models are not modelling deception or veracity, thus making them vulnerable to adversarial attacks.
Fake news classification results varied with gradual changes in input data over 4 times that of Sentiment Analysis on SST-5 (Table 4). Taking the Mean Square Error (MSE) for each run one would measure the diference in classifications from their target label. Taking an Overall MSE for the 5 runs, fake news classification shows considerable changes in output (Table 4).
By the definition of ill-conditioned problems [
Factors supporting the case that NLP-based, fake news classification is an ill-posed problem
include:
1. There appears to be no indicator of truthfulness or deception within LIAR’s statements
unless one has knowledge of the real world. Sentiment Analysis by contrast, can be based
on the presence of certain words or expressions.
2. Feature Based detection does not generalise over domains [
All these point to the likelihood that there cannot be a model that maps a string of text to truth levels without knowledge of the world. This likelihood is strong for the models trained on LIAR and demonstrated for our models despite their relatively higher accuracy.
Our best models achieve a higher accuracy on the LIAR dataset utilising the spoken statements and the speakers’ reputation alone, outperforming methods that either used more data, more complex models or both. BERT and BERT variants can be leveraged to classify short statements more accurately.
Bigger and better trained transformers yielded only a marginal improvement over the smaller BERT-base transformer. In our case, although ALBERT-Large did perform better than BERT-base, fake news classification accuracy did not scale in proportion to the transformer size.
The most important insights resulted from testing beyond accuracy scores. Flaws were found and these led to questioning the whole idea of language-based classification of content according to deception or veracity. Issues were also identified with the LIAR dataset and these flaws were used to test the efect that data quality has on the models’ ability to learn the task. Specifically, we show that although the models’ accuracy on LIAR is better than random, the language transformers’ contribution to the classification was generally poorer when trained and tested on cleaner data.
3Good models give a high accuracy. These models give a high accuracy; therefore, they are good. This is a logical fallacy known as Afirming the Consequent.
Furthermore, when compared with sentiment classification, fake news classification appears to be an unstable problem. We put forward arguments that suggest that purely NLP-based, fake-news classification on short statements, such as those found in LIAR, is not robust since it presents traits of ill-posed and ill-conditioned problems.
A simple test indicates that these models are not really modelling deception or veracity and are thus vulnerable to adversarial attacks.
The models herein, while improving on previous accuracies were thus proven unreliable for classifying arbitrary claims. The biggest contributor to the higher score was the reputation score which was shown to bias the models.
We used only the text and the speaker’s reputation in our tests, achieving a better score. An ablation study can also be done to analyse the impact of each attribute on the result.
Future studies can attempt similar investigations with the use of constructed features like part-of-speech tagging or dependency parsing together with those utilised internally by the transformer.
It would also be interesting to establish and standardise a variety of tests and metrics to assess the quality of a fake news classifier, by testing behaviour rather than mere accuracy scores; such as the ability to truly model veracity, stability (whether it is well conditioned or not), its ability to generalise over domains, and tests for bias. The ability to truly model veracity or deception deserves particular attention in future work, since it determines a classifier’s robustness against adversarial attacks.
Future studies are recommended to treat purely NLP-based, fake news detection as ill-posed, especially those utilising arbitrary or non-explainable features. Using knowledge-graphs to store knowledge about the real world, is likely one potential way to regularise the ill-posed problem.
Lastly, our models stand as examples of why analysis of a model’s behaviour should be a better judge of how good the model is, rather than mere accuracy. Going forward, we believe this to be essential for mitigating the fake news problem efectively. [20] Z. Zhou, H. Guan, M. Bhat, J. Hsu, Fake news detection via NLP is vulnerable to adversarial attacks, Proceedings of the 11th International Conference on Agents and Artificial Intelligence (ICAART 2019) (2019). [21] D. B. Buller, J. K. Burgoon, Interpersonal deception theory, Communication Theory 6 (1996) 203–242. [22] J. K. Burgoon, D. B. Buller, C. H. White, W. Afi, A. L. S. Buslig, The role of conversational involvement in deceptive interpersonal interactions, Personality and Social Psychology Bulletin 25 (1999) 669–686.