The Impact of Large Language Models on Social Media Communication

This article explores the impact of large language models (LLMs) on social media communication, with a focus on the spread of misinformation and cyberbullying. As social media becomes an integral part of modern life, challenges such as the rapid spread of misinformation and unethical online behavior continue to escalate. In this paper, the lab's main research delves into how large language models can improve the accuracy of information dissemination on platforms such as Twitter with their advanced capabilities and larger parameters. It also highlights the application of LLMs in identifying and filtering misinformation, as well as potential ethical and privacy considerations associated with their use. The studies mentioned here also explore the impact of LLMs in shaping social media communications, addressing technological advancements, and attendant social responsibilities.


INTRODUCTION
Social networking has become an important symbol of maintaining social progress and civilization development, and this means that social media has completely changed people's social lives and communication methods because it activates and evokes the value of interpersonal relationships in human communication.In today's society, people are usually inseparable from social media.In other words, social media has become a part of everyone's life.In an era when social media is very developed, while TikTok and YouTube short videos have begun to develop crazily, X (Twitter) is also the leader in social media communication, leading the advancement of social communication to all people in the world.However, while developing, there are many challenges and problems in social media that every social media developer and social media user needs to face.To give a controversial example, the spread of false information and cyberbullying has become serious, and usually, we can find that false information spreads faster than correct or truthful information because people usually believe what they want to believe.In an article written by Kee in 2022 about cyberbullying under the influence of the COVID-19 pandemic, it was mentioned that the increase in the use of social media has led to an increasing number of cyberbullying incidents [4].Regarding cyberbullying, people will lash out and vent their anger on social media because of the Internet, regardless of cultural differences and moral issues.This paper will mainly focus on the research and discussion of the spread of false information, and then mainly mention the impact and help of large language models on social media.
Because of the continuous advancement of large language model technology and the expansion of application scenarios, we can foresee that its role in the media field will be broader and deeper.The application of large language models in the media will inevitably have a profound impact on society, so social media needs to assume corresponding social responsibilities, pay attention to the authenticity and credibility of information, and avoid spreading false information and irresponsible content.Törnberg (2023) and his laboratory members pointed out that large language models have shown great potential in different social media platforms [8].Moreover, in Törnberg's research, they also used natural language models to create specific algorithms based on them to reveal the potential impact of algorithm changes on social media communication.Overall, large language models can help understand and improve the state of social media, as well as effectively study the broader complex social media environment.The development of large language models is rapid, which can help social media to control the accuracy of information.Because large language models have larger parameters than machine learning models, large language models can achieve better results in training and prediction than machine learning models.This article will also describe in detail how to use large language models to train (fine-tuning) and predict whether information in social media is accurate or incorrect.This essay will next describe in detail the working principle of large language models, the current status of social media, the application of large language models in social media, the challenges and opportunities of large language models in social media, ethical and privacy considerations, and future work for large language models.Language model development and debugging.

OVERVIEW OF LARGE LANGUAGE MODELS
A large language model is a model trained based on a large amount of language data as a large number of parameters, and this model can be used to analyze, parse, and predict language in different fields.The working principle of the large language model is mainly divided into two stages.One is to directly use the existing large language model, such as Llama 2, to directly ask the questions you want to ask and use the officially trained model to Answer the question directly.Roumeliotis (2023) and the team mainly explained in a discussion article about the Llama 2 major language model that in terms of fine-tuning functions, the model provides a wide range of fine-tuning functions [10].In terms of domain development, large language models can also be adapted to various tasks and domains.Although there is no need to participate in training or fine-tuning artificially, the questions generally answered are like a hundred flowers blooming, and the way and results of the answers are usually not what we expect.The other stage is also the most important and challenging because it includes the fine-tuning model that is widely discussed in the AI industry.After fine-tuning the model efficiently and effectively, the model will usually return the format and results we expect.For example, you will get a dictionary format or a fixed-format sentence after excellent fine-tuning: "Label:…, explanation:…".
There are various methods of training or fine-tuning.The project "LLaMA Factory: Training and Evaluating Large Language Models with Minimal Effort" by hiyouga user on GitHub in 2023 explains in detail how to use Llama2 to fine-tune the model [7].Moreover, he/she gave methods for using parameters and checking results, which largely helped the laboratory fine-tune the Llama2 large language model.One method generally recognized in the industry where large language models are used is to require three parameters in a JSON file.The first parameter is instruction, which means to explain the question you need GPT to answer.For example, "The relationship between A's statement and B's tweet has the same meaning" or "This tweet is toxic.".The second parameter is input.The meaning of this is that when you mention the words in the instructions, you can specify them in the input details.To give an example related to the previous one, if A and B are mentioned in the instruction, then in the input you can explain what A and B are respectively.The third parameter is output, which is also the most important part of the fine-tuning file.The meaning of this is what format of results you expect the large language model to return to you.For example, you only want a label (True or False), an explanation, a probability, or a collection of them.After completing the fine-tuned JSON file and including the instruction, input, and output mentioned earlier, you are ready to fine-tune it.
Examples of fine-tuning training will be shown and demonstrated to clearly explain the preparation and operation.In the original data acquisition, the laboratory mainly provided some tweets sent by people on X (Twitter) during the virus pandemic for specific text analysis.The original data mainly includes tweets and retweets posted by Twitter users in different regions during the virus pandemic.Most of these tweets were determined to be real tweets sent by real users and had not been altered in any way.For privacy protection and ethical reasons, the user's real username is also not included in the raw data, and each tweet is accompanied by a time.Next, through the embedding vector matching algorithm and the help of many network volunteers, the public certification and recognition of the fact that the claim and the tweet are matched are completed, and the relationship between the claim and the tweet is voted out as positive, negative, or neutral.The voting data mentioned here are used to determine the matching relationship between a claim and a tweet through the mturk service platform provided by Amazon, allowing online volunteers to help us provide a label for the data.After getting hundreds of labels from volunteers, we got the most popular label from the options we gave (positive, negative, or neutral) as a claim and a tweet label.As a mode, in terms of data, although many things are subjective because everyone's judgment of facts is different, but because of the many aspects of subjectivity, it can also help us use it as a reference to the objective existence of the data.Then, to achieve better fine-tuning effects, the laboratory used ChatGPT's API to generate forward tweets, reverse tweets, and neutral tweets based on the claim.Then, the fine-tuning data and test data are divided into 80% and 20% through machine learning model division to prepare for future model testing.
As a team member of the experimental group, I have made many attempts to fine-tune the model, but the data on model performance is not very impressive, so here we mainly show the fine-tuning attempts of the model rather than the data display.First, in the first attempt, there are three labels on the available data, and they represent the matching relationship between claim and tweet respectively.This means that if the label relationship between A and B is correct, it means that the other two label relationships are wrong.Such training data will cause a problem of data imbalance, because one-third of the training data is correct and two-thirds of the training data is wrong, which is very poor for fine-tuning the model.The performance of this model was both logically and practically very poor, resulting in the logical derivation score (AUC) being lower than 0.5.In the second attempt, the problem of data imbalance was solved as a classification problem.First, if you want to make the number of true labels and false labels consistent, you need to make some changes in the false label because the true label is stable and serves as a reference for the false label to make modifications.The solution here is that when the instruction of the true label mentions one of the three labels, then the instruction in the false label randomly selects another one of the three labels as the label in the false label.For example, in the true label, if the matching relationship of the instruction is positive, then in the false label, one can be randomly selected from negative and neutral as the matching relationship in the instruction.This attempt gave many subsequent attempts a solid foundation to make more changes and self-modifications.In the fine-tuning model, because the performance of the model is much higher than 0.5 in the AUC score, this means that the model begins to try Use the training data we give you to make logical deductions rather than random guesses.

Tables
After several fine-tuning in the laboratory, the optimal format of the three parameters mentioned above was summarized.The instrumentation is "The Relationship between the claim and the tweet is the label (positive, negative, or neutral)", and the input is the claim.and the specific explanation of the tweet, the output is a boolean (true or false) and a judgment explanation to ensure that the large language model gives reasonable results based on our fine-tuning.The following will also specifically show different model performances by sklearn.metricsthrough different fine-tuning formats.One point worth mentioning about the experiment here lies in the two fine-tuning methods and wording.Although there is not much difference in the output (true or false, with an explanation), the method of asking is also completely different.The method with the agree label mainly asks about the relationship between agreement or disagreement between a claim and a tweet, while the method with three clear labels mainly asks about the relationship between a claim and the three labels of a tweet.
On the test data, the best fine-tuning model voted by the project team got the results we wanted.Its format is "Label (true or false).Reason: ".If the label is true, then in terms of reason, you only need to mention the specific logical reason.If the label is false, then in terms of reason, it is necessary to mention the correct matching relationship between claim and tweet, and also mention why such a matching relationship is correct.In the summary of model performance, the advantage of the model is that it gives us the format we want, and successfully predicts the relationship between a claim and a tweet, and tells the specific reasons, which further proves that the model does not Random guess, but a logical derivation based on the training data combined with the internal model.But in terms of shortcomings, the challenges of the model also come one after another.Among the results of large language model prediction derivation, some results are contradictory.For example, although the label prediction is correct, the reason given by the model is completely inconsistent and irrelevant.This is also the reason why big language models are being controversial, because sometimes to please users, big language models will return the results they want (results look correct but are misleading).This challenge is also a problem faced by many developers today.So, every time after encountering this kind of problem, developers need to give feedback to the large language model and re-fine tune it to avoid similar problems.

CURRENT STATE OF SOCIAL MEDIA
Social media has gained great popularity and repercussions among both the younger and older generations to a large extent.As Esteban (2023) mentioned, Facebook already had nearly 2.3 billion users at the beginning of 2019, and it became the most popular social media platform that year [1].The number of users of other platforms is also extraordinary because their numbers are also in a total of 1.5 billion user clusters.They are familiar with YouTube, Instagram, WeChat, and TikTok.From the data, we can see that from 2019 to now, our social platforms have penetrated each of our lives, and we have become inseparable from them and even use them frequently every day.
Since the beginning of 2019, social media has begun to encounter great challenges to fake news and misinformation, and it has reached its peak during the COVID-19 period (2020 to now).Ravichandran and his team published a paper on the Classification of Covid-19 misinformation in 2023.In the paper, they mentioned that many organizations are accustomed to using a lot of misinformation to spread to achieve certain goals or intentions [2].And because a lot of misinformation and misleading information is easy for people to believe without delving into whether the information is wrong or not, a lot of this information tends to spread quickly and makes misinformation or inaccurate information become so-called "correct" information forwarded on various platforms.Therefore, many researchers, including laboratories, have begun to use large language models to try to solve similar problems so that people can get a good experience and environment to discuss and communicate on social media.The continuation of this article will thoughtfully discuss how to use large language models in social media as a solution.

APPLICATION OF LARGE LANGUAGE MODELS IN SOCIAL MEDIA
After the large language model has been fine-tuned and achieved a relatively outstanding effect, this model can usually be used to detect and filter misinformation or abnormal information in social media.If companies like First of all, the first plan can be targeted at teenagers or children.From the article by Kobiruzzaman (2018) on the impact of social media on teenagers, it can be seen that a large number of 210 million teenagers are addicted to using social media every day, and they will uncontrollably and unknowingly accept many mistakes made by others information and bad information [3].Furthermore, social media allows false information and rumors to spread quickly, leading to defamation and legal issues.Going back to the use of large language models, social media can use large language models to identify which information is wrong and which information is bad, and then classify and categorize it.Social media can then help tell teens what information is wrong, for example by displaying "The information in this Tweet is false or misleading" beneath each false tweet.This ensures that every teenager has the right to see every tweet, and ensures that teenagers are not influenced by wrong and inappropriate tweets to make bad decisions in their lives and communications on social media.Another solution is used by decision-makers in various industries and even politicians in the political world.
Although compared with teenagers, decision-makers and politicians in various industries have more of their ideas and make decisions based on past experiences and after reading information on social media, they can also be misled by some wrong information.Making wrong decisions will lead their companies in the wrong direction, and even lead the country in the wrong direction.The article "ChatGPT and the hospitality and tourism industry: an overview of current trends and future research directions" published by Gursoy in 2023 shows that real-time data management updates and privacy issues are hidden dangers and challenges that require attention from policymakers [9].For data management updates, because the dynamic nature of tourism information requires real-time updates to adapt to market changes, decision-makers can use this factual information to make corresponding decisionmaking changes the first time.However, dynamic real-time updating of information is the biggest problem, because whenever the real-time updated information is incorrect and misunderstood, decision-makers may make wrong decisions due to wrong information, leading to the company's failure.In other words, because the company's decision-makers want to be the first in the market to be the leader in the corresponding industry, they will have to get first-hand information to make decisions.This first-hand information may often not be verified, but directly used as reference information to help decision-making.Just like the youth plan, the tweet display of decision-makers can also inform them of the information accuracy of some tweets, thereby helping them understand the current social needs and make good decisions to promulgate new products or policies.To adapt to a society that is updating and changing every day.This also ensures that everyone has the right to tweet and retweet and that everyone has the right to know what information is wrong or misleading before making any decisions.

CHALLENGES AND OPPORTUNITIES
The challenge of large language models must be inevitable because laboratories usually avoid overfitting to achieve 100% accuracy of the model.Therefore, large language models cannot achieve 100% prediction and classification of whether each tweet is misinformation or normal information.From the first direction, some erroneous tweets are difficult to identify and classify by large language models, so such tweets that are difficult to distinguish will not have labels.Furthermore, because people already know that there is a large language model mechanism in social media, if a tweet is not labeled as misinformation, people will think that the tweet is normal.This will create new problems and challenges because misinformation will use this mechanism to spread widely to prove that they are normal.After all, the large language model does not give a label, and then more people will believe these tweets.From the second direction, that is, conversely, because the error of the large language model is not just classifying a wrong tweet into the wrong category, the large language model also has a chance of classifying the correct information into the wrong category.information, and then gives an error label on top of the correct tweet.This will also cause new problems because correct information should be promoted and disseminated, but this will cause correct information to be covered up, reported, and blocked by people.If these two directions occur at the same time, then multiple times the number of problems will ensue in the social media environment.
There are solutions to the problems mentioned above because we can test a large number of tweets after training the large language model to see the performance of the model.Then, when the problem of classification errors is discovered, the misclassified tweets are manually classified into the correct classification to prepare for the next model fine-tuning.After a large number of manual corrections to the labels, you can return to the laboratory for fine-tuning, which can greatly and effectively solve the problems mentioned above.
Although large language models are mostly used in articles and tweets, this model can also be used on video websites.For example, YouTube already has the function of subtitles in the current development of YouTube, so we can use the language analysis and summary functions of large language models to separately extract the subtitles in a video and do topic classification and summary, such as tweet classification.Same preparation.Then, YouTube can use large language models to classify misinformation in YouTube videos, and this can also improve people's environmental experience on the video website to make social communication effective and correct.In an article called "Offensive language detection in Tamil YouTube comments by adapters and cross-domain knowledge transfer" written by Malliga Subramanian and other team members in 2022, it was highlighted that YouTube is used in large language models.The summary of their laboratory means that because YouTube has many toxic properties in both comments and videos, after the large language model can be fine-tuned, YouTube can detect these toxic comments and Videos and process, so that the communication environment on YouTube will be better than the previous experience [6].

ETHICAL AND PRIVACY CONSIDERATIONS
In terms of ethics and personal privacy, the handling of large language models is often a controversial topic because there are many unconfirmed rumors that large language models will use data provided by users or developers to train and improve themselves to be Used by a user or developer.In an article called "ChatGPT for good?On opportunities and challenges of large language models for education" written by Enkelejda Kasneci in 2023, it was emphasized that data security is important for large language models.His summary means that large language models need to be recognized and accepted by users before they can be collected and trained [5].However, this controversial topic can be avoided because the laboratory can ensure and declare that the data used by the large language model when fine-tuning the model is obtained with permission.Then, the company and the developer can sign a contract to explain that when using a large language model, the company and the developer will not extract the user's data for training without the user's knowledge.Users on various social media should have the right to know that large language models have been used to give each tweet or video a label.In this way, such controversial issues of morality and personal privacy will be alleviated, and the anxiety and questions brought by users and developers will be solved accordingly.Another ethical issue is the harm of misinformation, as large language models may spread false or misleading information to users.This results in less informed users and erodes trust in the information being shared.In sensitive fields, such as law or medicine, misinformation from large language models can also lead users to perform unethical or illegal actions [12] (Weidinger, 2021).

CONCLUSION AND FUTURE WORK
In summary, large language models have great potential in improving the integrity and reliability of social media communication.By fine-tuning and deploying these models, social media platforms can be found to be more effective in combating the spread of misinformation and unethical behavior online.However, in terms of the challenges and limitations of large language models, there is still a long way to go to improve large language models, such as the risk of overfitting and potential information misclassification.Additionally, ethics and privacy issues remain key issues, so transparent and responsible use of large language models is something every developer needs to do.Looking forward, big language models, if they can be subject to continuous improvement and ethical oversight, will be able to maximize their benefits while also minimizing potential harm.If managed properly, the integration of LLMs with social media can lead to a more informed and ethical digital communication environment.Liu and Yiheng also specifically discussed the current status of large language models in 2023 and what they believe are the prospects of large language models in different industries and academic scopes [11].In terms of ChatGPT, after large-scale pre-training of human factors, instruction finetuning, and feedback reinforcement learning, ChatGPT can be even more powerful and has been significantly improved in various wellknown fields.For industries, the ChatGPT model can effectively interpret the requirements put forward by users, such as reading a large number of articles and analyzing real-time feedback from customers, which immediately improves user efficiency.However, the potential ethical implications are still worthy of attention and need to be addressed and challenged in the future because privacy and data security concerns are issues worthy of attention in every industry and academic field.

Table 1 :
Comparative Analysis of Instruction Methods: This table displays the ROC AUC, Precision, and F1 Scores for four distinct instruction methods, namely 'Relationship (positive/negative, neutral)' and 'Agree/disagree/nor', both with and without explanation.

Table 2 :
Performance Trends of Instruction Methods: The graph illustrates ROC AUC, Precision, and F1 Scores for four instructional approaches, highlighting the performance trade-offs between methods with and without explanations.