Summary

In this paper, the authors propose a mixed initiative approach for fact checking that combine both human knowledge and experience with the automated information retrieval and machine learning. The paper discusses the challenges of the massive amount of information available today on the internet that some of them might not be accurate which introduce a risk to the information consumers. The proposed system retrieve relevant information about a certain topic and use machine learning and natural language processing to assess the factuality of the information and provide a confidence level to the users and let the user decide wither to use the information or do a manual research to validate the claims. The partnership between the artificial intelligence system and human interaction can offer effective fact checking that can support the human decision in a salable and effective way.

Reflection

I found the approach used by the authors to be very interesting. I personally had a discussion with a friend recently about a certain topic that was mentioned in Wikipedia, and I thought the numbers and facts mentioned were accurate but it turns out the information were wrong and he asked me to check an accreditable source. If I had the opportunity to use the proposed system on the paper, then accredited source could have ranked higher than Wikipedia.

The proposed system is very important in our digital age where so much information is generated on a daily bias and we are not only searching for information, but we are also receiving so much information through social media related to current events and some of these events have high impact on our life and we need to assess the factuality of these information and the proposed system can help a lot on that.

The proposed system is like a search engine that not only rank document based on relevance to the search query but also based on the fact-checking assessment of the information. The human interaction is like the relevance feedback in search engine which can improve the retrieval of the information which leads to a better ranking.

Questions

• The AI systems can be biased because the training data can be biased. How can we make the system unbiased?
• The proposed system use information retrieval to retrieve relevant articles about a certain topic and then use the machine learning to validate the source of the information and then present the confidence level of each article. Do you think the system should filter out articles with poor accuracy as they might confuse the user? Or they might be very valuable?

• With the increase usage of social networking, many individuals write or share fake news intentionally or unintentionally. Millions of people post information every day. Can we use the proposed system to assess the fake news? if yes, then can we scale the system to assess millions or billions of tweets or posts?

Read More

In today’s era of fake news where new information is constantly spawning everywhere, the great importance of fact checking cannot be understated. The public has a right to remain informed and be able to obtain true information from accurate, reputable sources. But all too often, people are inundated with too much information and the cognitive load of fact checking everything would be too much. Automated fact checking has made strides but previous work has focused primarily on model accuracy and not on the people who need to use them. This paper is the first to study an interface for humans to use a fact checking tool. The tool is pretrained on the Emergent dataset of annotated articles and sources and uses two models, one that predicts article stance on a claim and the other that calculates the accuracy of the claim based on the reputation of the sources. The application works by taking a claim and retrieving articles that talk about the claim. It uses the article stance model to classify if the articles are for or against the given claim, and then predicts the claim’s accuracy based on the collective reputation of its sources. It conveys that its models are not accurate and provides confidence levels for its accuracy claims. It also provides sliders for the human verifiers to adjust the predicted stance of the articles and also to adjust the source reputation according to their beliefs or new information. The authors run three experiments to test the efficacy of the tool for human fact checkers. They find that the users tend to trust the system, which can be problematic when the system is inaccurate.

I find it interesting that for the first experiment, the System group’s error rate somewhat follows the stance classifiers error rate. The crowd workers are probably not going through the process of independently verifying the stance of the articles and simply trust the predicted stance they are shown. Potentially this could be mitigated by adding incentives (like extra reward) to have them actually read the articles in full. But on the flip side, we can see that their accuracy (supposedly) becomes better when they are given the sliders to modify the stances and reputation. Maybe that interactivity was the clue they needed to understand that the predicted values aren’t set in stone and could potentially be inaccurate. Though I find it strange that the Slider group in the second experiment did not adjust the sliders if they were questioning the sources. What I find even stranger though is the fact that the authors decided to keep the claim that allowing the users to use the sliders made them more accurate. This claim is what most readers would take away unless they were carefully reading the experiments and the riders. And I don’t like that they kept the second experiment results despite them not showing any useful signal. Ultimately, I don’t buy into their push that this tool is something that is useful for the general user as it stands now. And I also don’t really see how this tool could serve as a technological mediator for people with opposing views, at least not the way they described it. I find that this could serve as a useful automation tool for expert fact checkers as part of their work but not for the ordinary user, which is what they model by using crowdworkers. I like the ideas that the paper is going for, of having automated fact checking that helps for the ordinary user and I’m glad they acknowledge the drawbacks. But I think there are too many drawbacks that prevent me from fully buying into the claims of this paper. It’s poetic that I have my doubts about the claims of a paper describing a system that asks you to question claims.

1. Do you think this tool would actually be useful in the hands of an ordinary user? Or would it serve better in the hands of an expert fact checker?
2. What would you like to see added to the interface, in addition to what they already have?
3. This is a larger question, but is there value in having the transparency of the machine learning models in the way they have done (by having sliders that we can manipulate to see the final value change)? How much detail is too much? What about for more complex models where you can’t have that instantaneous feedback (like style transfer) how do you provide explainability there?
4. Do you find the experiments rigorous enough and conclusions significant enough to back up the claims they are making?

Read More

In this paper, the authors state that the current fully automatic fact-checking systems are not good enough for three reasons: model transparency, taking world facts into consideration, and model uncertainty communication. So, the authors went on and built a system including humans in the loop. Their proposed system uses two classifiers that each predict the reliability of a supporting document of a claim and the veracity of the document. Using these weighted classifications, the confidence of the system’s prediction about a claim is shown to the user. The users can further manipulate the system by modifying the weights of the system. The authors conducted a user study of their system with Mturk workers. The authors found that their approach was effective, but also noted that too much information or misleading predictions can lead to big user errors.

First off, it was hilarious that the authors cited Wikipedia to introduce Information Literacy in a paper about evaluating information. I personally took it as a subtle joke left by the authors. However, it also led me to a question about the system. If I did not miss it, the authors did not explain where the relevant sources or articles came from that supported a claim. I was a little concerned if some of the articles used in the study were not reliable sources.

Also, the users conducted the user study using their own defined set of claims. While I understand this was needed for efficient study, I wanted to know how the system would work in the wild. If a user searched a claim that he or she knows is true, would the system agree with high confidence? If not, would the user have been able to correct the system using their interface? It seemed that some portion of the users were confused, especially with the error correction part of the system. I think these would have been valuable to know and would seriously need to be addressed if the system were to become a commercial product.

These are the questions that I had while reading the paper:

1. How much user intervention do you think is enough for these kinds of systems? I personally think if the users are given too much power over the system, users will apply their bias to the correction and get false positives.

2. What would be a good way for the system to only retrieve ‘reliable’ sources to reference? Stating that a claim is true based on a Wikipedia article would obviously not be so assuring. Also, academic papers cannot address all claims, especially if they are social claims. What would be a good threshold? How could this be detected?

3. Given the current system, would you believe the results that the system gives? Do you think the system addresses the three requirements that the authors introduced which all fact-checking systems should possess? I personally think that system transparency is still lacking. The system shows a lot about what kind of sources it used and how much weight it’s putting into them, but it does not really explain how it made the decision.

Read More

Word count: 567

Summary of the Reading

This paper presents the design and evaluation of a system that is designed to help people check the validity of claims. The system starts with a user entering a claim into the system. The system then shows the user a list of articles related to the claim with a prediction that is based on the articles, about whether or not the claim is true. The system will give a percentage chance for whether the article is true. Each article that the system shows also shows a reputation score for the source of the article and a support score for the article. The user can then adjust these if they don’t think that the system has accurate information. 

The system seemed to help users come to the right conclusions when it had the right data but also seemed to make human judgements worse when the system had inaccurate information. This shows the usefulness of such a system and also gives a reason to be careful about implementing it.

Reflections and Connections

I think that this article tackles a very real and very important problem. Misinformation is more prevalent now than ever, and it is getting harder and harder to find the truth. This can have real effects on people’s lives. If people get the wrong information about a medication or a harmful activity, they may experience a preventable injury or even death. Misinformation can also have a huge impact on politics and may get people to vote in a way they might not otherwise.

The paper brings up the fact that people may over rely on a system like this, just blindly believing the results of the system without putting more thought into it and I think that is the paradox of this system. People want correct information and we want it to be easy to find out if something is correct or not, but the fact of the matter is that it’s just not easy to find out if something is true or not. A system like this would be great when it worked and told people the truth, but it would make the problem worse when it came to the wrong conclusion and then made more people more confident in their wrong answer. No matter how good a system is, it will still fail. Even the best journalists in the world, writing for the most prestigious newspapers in the world, will get things wrong. And, a system like this one will get things wrong even more often. The fact of the matter is that people should always be skeptical and they should always do research before believing that something is true or not, because no easy answer like this can ever be 100% right and if it can’t be 100% right, we shouldn’t trick ourselves into trusting it any more than we should. This is a powerful tool, but we should not rely on it or anything like it.

Questions

1. Should we even try to make systems like this if they will be wrong some of the time?
2. How can we make sure that people don’t over rely on systems like this? Can we still use them without only using them?
3. What’s the best way to check facts? How do you check your facts?

Read More

This work aims to design and evaluate a mixed-initiative approach to fact-checking by blending human knowledge and experience with the efficiency and scalability of automated information retrieval and ML. The paper positions automatic fact-checking systems as an assistive technology to augment human decision making. The proposed system fulfills three key properties namely model transparency, support for integrating user knowledge, and quantification and communication of model uncertainty. Three experiments were conducted using MTurk workers to measure the participants’ performance in predicting the veracity of given claims using the system developed. The first experiment compared users who perform the task with and without seeing ML predictions. The second compared a static interface with an interactive interface where the users were provided options to mend or override the predictions of the AI system. Results showed that the users were generally able to use the interface but this was of little use when the predictions were accurate. The last experiment compared a gamified task design with a non-gamified one, but no significant differences in performance were found. The paper also discusses the limitations of the proposed system and explores further research opportunities.

I liked the fact that the focus of the paper was more on designing automated systems that were user-friendly rather than focussing on improving prediction accuracy. The paper takes into consideration the human element of the human-AI interaction and focuses on making the system better and more meaningful. The proposed system aims to learn from the user and provide a personalized prediction based on the opinions and inputs from the user.

I liked the focus on transparency and communication. Transparent models help users better understanding the internal workings of the system and hence helps build trust. Regarding communication, I feel that conveying the confidence of the prediction helps the users make an informed decision. This is much better than a system that might have high precision but does not communicate the confidence scores. In cases where this system makes an error, the consequences are likely to be precarious since the user might blindly follow the prediction of the system.

The side-effect of making the system transparent was interesting. Not only would transparency lead to higher trust levels, but it would also help teach and structure the user’s own information literacy skills regarding the logical process to follow for assessing claim validity. Thereby, the system proposed truly leveraged the complementary strengths of the human and the AI.

• Apart from the three properties incorporated in the study (transparency, support for integrating user knowledge, and communication of model uncertainty), what are some other properties that can be incorporated to better the AI systems?
• The study aims to leverage the complementary strengths of humans and AI but certain results were inconclusive as noted in the paper. Besides the limitations enumerated in the paper, what are other potential drawbacks of the proposed system?
• Given that the study presented is in the context of automated fact-checking systems, what other AI systems can these principles be applied to?

Read More

Summary:
This paper discusses the issue of fact-checking, i.e., the estimation of the credibility of a given statement, which is extremely pertinent in today’s climate. With the generation and exploration of information becoming increasingly simpler, the task of judging the trustworthiness of the found information is becoming increasingly challenging. Researchers have attempted to address this issue with a plethora of AI based fact-checking tools. However, a majority of these are based on neural network models that provide no description, whatsoever, of how they have arrived at a particular verdict. It has not yet been addressed how this lack of transparency, in addition to a poor user interface, affects the usability or even, the perceived authenticity of such tools. This tends to breed mistrust in the minds of the user towards the tool, which ends up being counter-productive to the original goal. This paper tackles this important issue head-on by designing a transparent system that mixes the scope and scalability of an AI system with inputs from the user. While this transparency increases the user’s trust in the system, the user’s input also improves the predictive accuracy of the system. On a very important note, there is an additional side-benefit that this user interaction addresses the deeper issue of information illiteracy in our society, by cultivating the skill of questioning the veracity of a claim and the reliability of a source. The researchers build a model that uses NLP tools to aggregate and assess various articles related to a certain statement and assign a stance to each of the articles. The user can modify the weights associated with each of these stances and influence the verdict generated by the model on the veracity of the claim. They, further, conduct three studies, wherein they judge the usability, the effectiveness, and flaws of this model. They compare the assessments of multiple statements by the participants, before and after being exposed to the model’s prediction. They also verify if the interaction with the model provides additional support as compared to simply displaying the result. A third task estimates if the gamification of this task has any effects. While the third task seems inconclusive, the first two tasks lead the researchers to conclude that interaction with the system causes an increase in the user’s trust in the model’s results, albeit even in the case of a false prediction of the model. However, this interaction also helps improve the prediction of the model, for the claims tested herein.

Reflections:
The paper brings up an interesting point about the transparency of the model. When people talk about an AI system normally a binary system comes to the mind, which takes in a statement, assesses it in a black box, and returns a binary answer. What is interesting is that the ability to interact with the predictions of the models, enables the user to improve their own judgment and even compensate for the lacking of the model. The aspect of user interaction aspect in AI systems has been grossly overlooked. While there are some clear undeniable benefits of this model, there is a very dangerous issue that this human modifiable fact-checking could lead to. Using the slider to modify the reputation of a given source, can potentially lead to user’s inducing their own biases into the system, and effectively creating echo chambers of their own views. This could nefariously impact the verdict of the ML system and thus reinforcing the possible prejudices of the user. I would suggest that the model should assign x% weightage to its own assessment of the source and (100-x)% to the user’s assessment. This would be a step in ensuring that the user’s prejudices do not suppress the model’s judgment completely. However, without doubt, the advantage of this interaction, inadvertently, helping a user learn how to tackle misinformation and check the reputation of sources, is highly laudable. This is definitely worth considering in future models along these lines. From the human perspective, the bayesian or linear approaches adopted by these models make them very intuitive to understand. However, one must not underestimate the effectiveness of neural networks in being more powerful in aggregating relevant information and assessing its quality. A simple linear approach is bound to have its fallacies, and hence, it would be interesting to see a model with uses the power of neural networks in addition to these techniques with help with the transparency aspect. On a side note, it would have been useful to have more information on the NLP and ML methods used. The related work regarding these models is also insufficient to provide a clear background about the existing techniques. One glaring issue with the paper is their circumventing of the insignificance of their result in task #2. They mention that the p-value is just below the threshold. However, statistics teaches us that the exact value is not of importance, it’s the threshold that is set before conducting the experiment that matters. Thus, the statement “..slightly larger than the 0.05..” is simply careless.

Questions:
1. Why is there no control group in task 2 and task 3 as well?
2. What are your general thoughts on the paper? Do you approve of the methodology?

Read More