=Paper=
{{Paper
|id=Vol-2696/paper_134
|storemode=property
|title=Team Buster.ai at CheckThat! 2020: Insights and Recommendations to Improve Fact-Checking
|pdfUrl=https://ceur-ws.org/Vol-2696/paper_134.pdf
|volume=Vol-2696
|authors=Mostafa Bouziane,Hugo Perrin,Aurélien Cluzeau,Julien Mardas,Amine Sadeq
|dblpUrl=https://dblp.org/rec/conf/clef/BouzianePCMS20
}}
==Team Buster.ai at CheckThat! 2020: Insights and Recommendations to Improve Fact-Checking==
https://ceur-ws.org/Vol-2696/paper_134.pdf
Team Buster.ai at CheckThat! 2020: Insights
And Recommendations To Improve
Fact-Checking
Mostafa Bouziane, Hugo Perrin,
Aurélien Cluzeau, Julien Mardas, and Amine Sadeq
Buster.AI, France
contact@buster.ai
Abstract. As part of the CheckThat! 2020 Task 2, we investigated sen-
tence similarity using transformer models. In Task 2, the goal was to
effectively rank claims based on their relevancy compared to an input
tweet. While setting our baseline on sentence similarity for fact-checking,
we gathered insights we felt compelled to share in this paper. We learned
how multi-modal data utilization could foster significant uplifts in model
performance. We also gained knowledge on which hybrid training and
strong sampling worked best for fact-checking applications, and wanted
to share our interpretation of the results we got. Finally, we want to
explain our recommendations on data augmentations. All of the above
allowed us to set our baseline in fact-checking in the CLEF Checkthat!
2020 Task 2 competition.
Keywords: Fact-Checking, Veracity, Evidence-based Verification, Fake News
Detection, Computational Journalism, Natural Language Processing, Deep Learn-
ing, Language Model, Sentence Similarity.
Introduction
In the run-up to elections in many economic powers, being able to discern de-
ception from real news has never been more critical. Moreover, in recent years,
the actual cost of inaccurate news has been assessed more thoroughly. With oc-
currences like the Bloomberg dubious report almost wiping up 6 billion euros of
Vinci market value, resulting in a 5 million euros fine for the media conglomerate
recently [13] [4], it is also manifest that fake news can have a significant impact
on economic and financial markets. As we are observing during the 2019/2020
coronavirus crisis [8], false health information, potentially causes grave harm to
public safety and societies as a whole, from the economic to the political stability
of entire countries. While most of the effort to combat them is handmade, we
believe automation is vital to ensure propelling this fight to a broader impact.
Copyright c 2020 for this paper by its authors. Use permitted under Creative Com-
mons License Attribution 4.0 International (CC BY 4.0). CLEF 2020, 22-25 Septem-
ber 2020, Thessaloniki, Greece.
In CheckThat! 2020 Task 2 [10], as descibed in [9], the intended goal was to
effectively rank claims based on their relevancy to an input tweet. We had to
output a list of numerical scores for sentences in the database, the higher the
individual score, the more accurately related the match with the input text is.
It became natural that we judged our results with the Mean Average Precision
metric (MAP), evaluating if the valid claim is in the top-scored evidences. We
will expand in the background section on the several approaches that have been
tried. Figure 1 explains our strategy: after preprocessing the data and filtering
out potentially unsuitable claims, we can finally estimate the matching probabil-
ities with our neural network on acceptable claims. This filtering enables faster
computations while preserving most of the performance that we have when com-
puting scores for all documents of the database. We consequently trained our
transformer model to learn sentence similarity on a semantic level, which we will
detail in the methodology section.
Solving such a task will assist human fact-checkers in increasing their check-
ing throughput drastically. A desirable algorithm for this task will permit ob-
taining source candidates automatically, diminishing proof searching times for
human fact-checkers, as explained in Figure 1. It is noteworthy that sentence
correspondence analysis only serves as a guide for humans, and cannot replace
them for the ultimate purpose of fact-checking. Indeed, the necessity for human
settlements can further ensure trust from everyone on the exactness of the fact-
checking process, confidence, which is also strengthened by the associated facts
from cited references so anyone can verify them.
Fig. 1. Fact-checking pipeline
Furthermore, grand strides have been made for general interpretable artificial
intelligence, and resolving such a task benefits in achieving this in the fact-
checking setting. Being able to find matches on an evidence base allows the
algorithm to provide its sources for individuals to gauge its prediction quality
and exhibits where an automatic entailment model built on top would acquire
its reference material from.
For all those reasons, this aim has a practical significance and usefulness.
After focusing on what has been done already to solve such a problem, we will
succinctly present our competition-winning submission. We will outline the un-
derlying architecture and how we trained it, which will be a strong starting point
to describe the insights we gained from the competition. The first recommenda-
tion we share is how multi-modal knowledge can be leveraged to enhance score
quality. Then we want to discuss the use of external datasets, which helped us
propelling our algorithm farther and our interpretation of why specific datasets
appeared to help more than others. Building on these observations, we want to
highlight how proper sampling is required to train models efficiently. Finally, we
will give a concise overview of other data augmentations that have impacted our
model both for training and inference.
Background
In Task 2, we recall the goal is to rank claims based on their correspondence
with the input statement. So, chiefly our goal is to build a sentence similarity
model to learn semantics and hidden meanings. The latest advances in Natu-
ral Language Processing (NLP) suggest to use Transformer models and atten-
tion models. The Transformer [14] idea placed the latest paradigm for solving
sequence-to-sequence tasks while handling long-range dependencies with ease.
This architecture is based on self-attention, which allows the model to look at
the other words in the input sequence to understand a particular word in the
sequence better. Such attention is instrumental when the sentence is too long, or
it contains a hidden context, this is why research has devised different variants
of Transformer models, of which Bidirectional Encoder Representations from
Transformers (BERT) is the most notable and constitutes state of the art [2].
BERT’s introduction caused a massive surprise in the NLP community, as it
outperformed previous models in so many distinct tasks. This considerable suc-
cess comes from BERT being one of the first models that were non-directional.
Previous models based on recurrent neural nets (RNNs) or convolutional neu-
ral nets (CNNs) implied a notion of direction, insofar as they implied a specific
sequentiality when reading the input. For example, these models will predict a
missing word in a sentence using either previous words or the following words,
but can solely combine both approaches as an ensembling method. Nevertheless,
with the BERT approach based on transformers, this prediction can be made
using all available words simultaneously through the attention-based architec-
ture, making BERT much more potent in creating context-based embeddings.
These context-aware representations are of interest for the task at hand, since
they allow elegant intertwined semantic links between query input and potential
candidate evidence. In addition to a new non-directional approach, BERT-like
models also introduced a new way of encoding input data, which makes BERT
able to receive a single sentence as input or a two-sentence input, making it suit-
able for such a large field of different tasks. Moreover, the pre-trained weights
on vast amounts of data were released, leading to a sizable interest by the com-
munity and a substantial amount of analysis, understanding, and improvement
of the architecture, creating a series of BERT based papers with several adjust-
ments and specifications [6].
We also deem noteworthy to draw parallels between the CLEF Task 2 natural
pipeline and that of some of the pipelines proposed in Open Domain Question
Answering, like has been posted recently by HuggingFace labs [3], since both
exhibit the same kind of filtering/classify pipeline. This blog post has not inspired
our work, as it just was published at the time of writing, but it is nonetheless
an interesting source to reframe our current work.
Methodology: Our submission overview
Task Definition and dataset
Task 2 aims to build tools that verify tweets based on a database of genuine
claims. When we feed a tweet to the model, the comparison made with the
claims database will produce a ranking from the most to the least relevant claims
extracted from the database.
Fig. 2. Task 2 definition
Pairs of tweet and related claim constitute the samples of the available
dataset; the tweet is the input that we want to verify, and the related claim
can be retrieved from the database of verified claims.
Proposed approach
A rank-based method is an obvious thinking starting point to solve this task: a
method, where the model will have to learn patterns to rank pairs based on how
related they are. Nevertheless, for this specific task, we want to exploit the fact
that a single claim can be related to the input, simplifying the ranking task to
either being related, i.e., top-ranked, or not related, i.e., ranked second or more.
A rank-based model can be the right approach if the rank scores are smoothed
and well distributed over all claims, which is hard to achieve, since we do not
have this information in the provided dataset.
Training: Therefore, we instead opted for a binary classification approach,
as follows:
– Using genuine examples (pair tweet/claim) provided in the dataset.
– Generating wrong examples using several sampling strategies.
– Using a binary cross-entropy loss.
Inference: For all our submissions, once the model is trained properly, we
pair each tweet in the test set with all the verified claims in the database, and
we then apply the neural network with a composite score s function for ranking,
that we defined as in Equation 1, with pp being the probability of being related,
and pn being the probability of not being related to the query. It is important to
mention that we replace the last sigmoid activation by a softmax activation for
the last layer, even though in most cases sigmoid and softmax perform similarly,
since the values sum up to roughly one already on the right data. Nevertheless,
softmax will prove a more reliable activation on unknown data.
(
pp if pp ≥ pn
s (pn , pp ) = (1)
1 − pn otherwise
Network architecture
Our network relies upon on a RoBERTa checkpoint [7] to begin training. There-
fore, the underlying architecture is a basic BERT architecture with a 12-layer,
768-hidden, 12-heads configuration, which amounts to 125 million parameters
[17]. Those parameters result from the need to have both robust and fast infer-
ence, which it both achieved in practice, as it achieved the top score of around
93% MAP at depth 5, with an inference time below 0.1 seconds per pair. In
absolute fairness, we shall point out that the complexity of inference, with m
being the number of facts in our database, and n being the number of queries,
is in O(nm), which means our approach requires a pre-filtering to scale with a
broader knowledge base.
Augmentations and dataset
We must now disclose the external data we tried to provide as a supplement for
training. We first tried using SciFact [15], a scientific claim verification corpus,
comprising 1 409 claims fact-checked against 5 183 abstracts with entailment
labels. Secondly, we tried augmenting our corpus with FEVER [12], a quite
substantially large dataset with 185 445 short claims, a few orders of magnitude
larger than SciFact and CLEF datasets. Finally, we also used Liar [16], which
constituted around 12 800 claims, placing it as a middle-ground in terms of size
and featuring short claims and evidence.
Results: Our insights
For all comparisons here, we used our checkpoints obtained at early training to
avoid overfitting issues to falsify any results, and we will mention whenever we
use a model checkpoint for our submissions for reference. We used Whoosh [1]
as a baseline instead of elastic search, and it provides a clean Python API, and
obtained decent metrics compared to ElasticSearch in our experiments.
Multi-modal data utilization
We define multi-modal information in this context as any information coming
from either external links, pictures, videos, or audio information contained in
the textual data; information, which we eventually translated into text data
that can be fed to the model alongside the original tweet, as described in Figure
3. The extraction process mentioned can be applied on many kinds of data; from
image reverse search to get a title for pictures, to fancier image/video description
networks, or even speech-to-text algorithms, the possibilities are varied. For our
submission, we mainly converted links to the page name associated with the
linked page, as well as getting a title for images referred to in tweets, with an
image reverse search. Table 1 sums up the results we have in terms of inference,
and we can notice a significant uplift in performance, since, with the same model
checkpoint, and no retraining, we achieve around 3% better performance, which
amounts to more than half of the potential performance left to gain in this case.
The importance of such preprocess arises from the inherent multimedia essence of
Twitter data, which can also be found in most news outlets nowadays. Hence, we
preconize easing up the learning and inference process by adding those additional
pieces of information.
Fig. 3. Multi-modal pipeline
Metric Depth Whoosh baseline score Without multi-modal With multi-modal
MAP 1 0.804 0.96 0.99
MAP 3 0.838 0.967 0.992
MAP 5 0.843 0.97 0.992
MAP 10 0.846 0.97 0.992
MAP 20 0.847 0.97 0.992
MAP all 0.847 0.97 0.992
Table 1. Metrics table on dev set. Results comparison between baseline, with and
without multi-modal information.
Hybrid training
This section will explain how we used hybrid training with different external
data and their effect on the overall results.
FEVER: Sampling pairs from the FEVER dataset during training seems to
help the model improve its performance. One reason is that those pairs have
similar syntactic structures and lengths to the ones available in Task 2. The
performance was stable on the dev set; the results on the test set, where we had
better MAP at depth k values, confirmed this.
SciFact: Sampling pairs from the SciFact dataset during training did not
improve the model performance. One reason is that examples from Scifact have
varying numbers of tokens, which are ampler compared to our core data, making
it hard for the model to converge and generalize well.
Liar: Same as SciFact, the Liar dataset did not improve the performance of
the model, since the structure of the examples does not match with the ones we
have in Task 2.
Metric Depth without FEVER with FEVER
MAP 1 0.8970 0.9070
MAP 3 0.9260 0.9370
MAP 5 0.9290 0.9380
MAP 10 0.9290 0.9380
MAP 20 0.9290 0.9380
MAP all 0.9290 0.9380
Table 2. Metrics table on test set. Results comparison between baseline, with and
without sampling from FEVER dataset.
Sampling scheme
We also advocate for sensible sampling schemes to ensure training on the opti-
mal information signal. When scrutinizing the CLEF dataset, we observe that
naive sampling will often sample what we call negative samples, i.e., samples not
related to the query we pair it with. Those negative samples are utterly unre-
lated to the said query, leading to learning a more trivial solution of not pairing
sentences which do not share anything meaningful. However, this naive approach
leads to subpar MAP, as it will have a hard time distinguishing close evidence
and rank them in an unsatisfiable manner as a consequence. We can illustrate
this effect with the following example: with the query ”The Pope is blue with
yellow stripes.” and most evidence talking about age and place of birth of vari-
ous people, plus a few sentences about the Pope, a non-optimal sampling could
lead to the algorithm only giving a high score to anything with ”The Pope” con-
tained in it, and therefore, poorly scoring the other sentences about the Pope
potentially. We solved this, inspiring ourselves from a reinforcement learning
principle called experience replay [5], in which we use past predictions to learn
more efficiently. In our case, we went towards this idea by sampling according to
indexed searches, which produced relatively good results that could fool our al-
gorithm trained with naive sampling, forcing it to learn the proper semantics to
distinguish between syntactically and lexically close sentences. Ideally, we would
take the top ranked results of our model to generate hard examples, however
we chose to avoid the incurred computationnal cost, while using ElasticSearch
results as a proxy for our model ranking. This sampling is a fast and accurate
enough approximation of the reinforced mechanism of sampling new data based
on our algorithm’s error.
We used a balanced sampling, and we fully acknowledge that doing this will
introduce a bias in the distribution of our dataset, which can have negative
impact on the learning; nevertheless, we advocate that it is worth leveraging
towards the fact-checking goal. In this context, we sample pairs of query and
evidence, therefore balancing means we sample as much matching pairs than
non-matchihng pairs. In Table 3, we have evidence it can be hugely beneficial to
the training process, as it helps to gain a tremendous amount of performance,
using the same network as our submission, making the difference between a
usable and unusable solution.
Metric Depth With proposed sampling With naive sampling
MAP 1 0.96 0.1
MAP 3 0.967 0.1
MAP 5 0.97 0.1
MAP 10 0.97 0.1
MAP 20 0.97 0.1
MAP all 0.97 0.1
Table 3. Metrics table on dev set. Results comparison between with and without
proper sampling.
Additional augmentations
In this section, we will present some other strategies that we tried but did not
seem to impact the improvement of the model’s performance significantly.
Named entity recognition (NER) based augmentation : Given that
Task 2 data contains a lot of named entities, which are a crucial indicator of
similarity between sentences, we tried to run a named entity recognition model
on pairs and augment them with the detected labels. The model is a fine-tuned
version of the bert-base multilingual cased model weights from HuggingFace’s
transformers [17]. The results did not improve the overall performance, maybe
because our augmentation strategy was not robust since we only add a separator
and the detected labels at the end of each sentence, which is not very efficient
since many sentences will have common labels and this will prevent the model
from a better convergence. A better strategy to try would be to add the label next
to the detected entity so that the model fits well on this kind of augmentation,
and will make more sense of the relation between the words and the added labels.
Another strategy would be to use the one-hot encoding of the named entities
and stack them with BERT embeddings before the classifier layers; that way,
the model will learn how to use those added encodings without having much
negative impact on the semantics learned from the original sentence.
ConceptNet [11] based augmentation : ConceptNet, a general knowl-
edge graph, treats words and phrases of natural language as nodes that it con-
nects with labeled edges, therein representing the general knowledge involved in
understanding language, facilitating comprehension of the meanings hidden be-
hind everyday use of a word. The intuition behind using ConceptNet for our task
is that combining it with words/sentence embeddings empowers an understand-
ing that is not accessible solely from distributional semantics. As with NER, we
augmented pairs using ConceptNet API, with semantic relations it produces.
Whenever we detect a part of speech in a sentence, i.e., verbs, nouns, or ad-
jectives, we append to sentences the relations retrieved from the graph (e.g.,
”chanting is synonym of shouting”). When adding those augmentations, the re-
sults did not vary significantly, since we gained around 0.5% in MAP at depth 1
on dev set, and almost stable for the other depths. Again, the strategy of adding
a separator and the augmentation at the end of the sentence may not be the
most pertinent. Indeed, ConceptNet supplies plenty of distinct relations such as
”synonym of”, ”antonym of”, ”related to”, ”can be done with”..., so choosing
the right relation is of great importance, and cannot be done straightforwardly,
as the correct relation to use may depend on context.
Metric Depth without ConceptNet with ConceptNet
MAP 1 0.96 0.965
MAP 3 0.967 0.967
MAP 5 0.967 0.968
MAP 10 0.967 0.969
MAP 20 0.967 0.970
MAP all 0.967 0.970
Table 4. Metrics table on dev set. Results comparison between with and without using
ConceptNet.
Back and forth translation : In order to augment the positive examples
in the dataset, one can think of using back and forth translation strategy, which
means : given a source S (English in our case) and target T (anything but En-
glish) language, we want to feed the example to translation algorithms following
two steps:
– Step one: the example will be fed to a translation algorithm from S to T,
that way we will get a new sentence in the target language.
– Step two: the new sentence will be fed to a translation algorithm from T to
S, that way we will get a new sentence in the source language, semantically
the same as the first one but with a different formulation.
Metric Depth without translation augmentation with augmentation
MAP 1 0.96 0.95
MAP 3 0.967 0.96
MAP 5 0.967 0.965
MAP 10 0.967 0.965
MAP 20 0.967 0.965
MAP all 0.967 0.965
Table 5. Metrics table on dev set. Results comparison between with and without using
translation augmentations.
Doing so, the dataset will be more diverse in terms of syntax, and will help
the model to generalize better. The model are MarianMT models from Hugging-
Face’s transformers [17].
This strategy did not help the baseline model to improve because the results
were stable (Figure 5). We believe that this is due to the fact that the dataset
is not big enough, so the augmentations did not have much effect on the overall
performance.
Conclusion
In conclusion, our first recommendation for fact-checking was to ensure any
piece of information, no matter the medium, is leveraged, to avoid as much as
possible, relying on hidden information/context or implicit knowledge. Then,
we wanted to stress the importance of consistent extra datasets, and how they
helped training, even when dealing with unrelated subjects, as it was the case
for us with FEVER, but also how they would impair learning if their size are not
consistent. We also explained the importance of proper sampling. Last but not
least, we presented a few augmentations that we experimented with, primarily
semantic and lexical related augmentations, and hopefully gave inspiration for
more augmentations to help fact-checking efforts in the future.
References
1. Chaput, M. Whoosh documentation, 2020.
2. Devlin, J., Chang, M., Lee, K., and Toutanova, K. BERT: pre-training of
deep bidirectional transformers for language understanding. CoRR abs/1810.04805
(2018).
3. Jernite, Y. Explain anything like i’m five: A model for open domain long form
question answering, 2020.
4. Kadri, F., and de Presse, A. F. Le faux vinci ou l’éloge de la prudence., 2016.
5. Lin, L.-J. Reinforcement learning for robots using neural networks. Tech. rep.,
Carnegie-Mellon Univ Pittsburgh PA School of Computer Science, 1993.
6. Liu, Q., Kusner, M. J., and Blunsom, P. A survey on contextual embeddings.
ArXiv abs/2003.07278 (2020).
7. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis,
M., Zettlemoyer, L., and Stoyanov, V. Roberta: A robustly optimized BERT
pretraining approach. CoRR abs/1907.11692 (2019).
8. Nations, U. During this coronavirus pandemic, ‘fake news’ is putting lives at risk:
Unesco, 2020.
9. Shaar, S., Babulkov, N., Da San Martino, G., and Nakov, P. That is a
known lie: Detecting previously fact-checked claims. In Proceedings of the 58th
Annual Meeting of the Association for Computational Linguistics (Online, July
2020), ACL ’20, Association for Computational Linguistics, pp. 3607–3618.
10. Shaar, S., Nikolov, A., Babulkov, N., Alam, F., Barrón-Cedeño, A., El-
sayed, T., Hasanain, M., Suwaileh, R., Haouari, F., Da San Martino, G.,
and Nakov, P. Overview of the CLEF-2020 CheckThat! lab on automatic identifi-
cation and verification of claims in social media: English tasks. In Working Notes of
CLEF 2020—Conference and Labs of the Evaluation Forum (Thessaloniki, Greece,
2020), CLEF ’2020.
11. Speer, R., Chin, J., and Havasi, C. Conceptnet 5.5: An open multilingual graph
of general knowledge. CoRR abs/1612.03975 (2016).
12. Thorne, J., Vlachos, A., Christodoulopoulos, C., and Mittal, A. FEVER:
a large-scale dataset for fact extraction and VERification. In Proceedings of the
2018 Conference of the North American Chapter of the Association for Com-
putational Linguistics: Human Language Technologies, Volume 1 (Long Papers)
(New Orleans, Louisiana, June 2018), Association for Computational Linguistics,
pp. 809–819.
13. Times, B. French hoax costs bloomberg 5m euros in fines., 2019.
14. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez,
A. N., Kaiser, L., and Polosukhin, I. Attention is all you need. CoRR
abs/1706.03762 (2017).
15. Wadden, D., Lin, S., Lo, K., Wang, L. L., van Zuylen, M., Cohan, A., and
Hajishirzi, H. Fact or fiction: Verifying scientific claims, 2020.
16. Wang, W. Y. “liar, liar pants on fire”: A new benchmark dataset for fake news
detection. In Proceedings of the 55th Annual Meeting of the Association for Com-
putational Linguistics (Volume 2: Short Papers) (Vancouver, Canada, July 2017),
Association for Computational Linguistics, pp. 422–426.
17. Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A.,
Cistac, P., Rault, T., Louf, R., Funtowicz, M., and Brew, J. Huggingface’s
transformers: State-of-the-art natural language processing. ArXiv abs/1910.03771
(2019).