AI in between online and offline discourse - and what has ChatGPT to do with all of that?
0 likes36 views
The document discusses the interdependencies between artificial intelligence (AI), online and offline discourse, and the role of models like ChatGPT within this framework. It highlights the challenges of AI in understanding and influencing discourse, including biases, misinformation, and the need for transparency and reproducibility. Additionally, it examines advancements in AI language models and their implications for social sciences and discourse analysis.
![26
What is driving the recent AI advances?
Deep & representation learning for language understanding
Percentage of deep learning papers in major NLP conferences
(Source: Young et al., Recent Trends in Deep Learning Based Natural Language Processing)
• Pretraining of embeddings: predicting
low-dimensional vector representations
of words & text, e.g. Word2Vec
[Mikolov et al., 2013]
• RNN/CNN architectures in
encoder/decoder settings (e.g. for
machine translation) [Vaswani et al.,
2017]
• Pretraining language models for task-
specific transfer learning, e.g., BERT -
Bidirectional Encoder Representations
from Transformers [Devlin et al., 2018]
T. Mikolov et al., Distributed Representations of Words and Phrases and their Compositionality, NIPS (2013)
J. Devlin et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018)
A. Vaswani et al. Attention is all you need, NIPS (2017)](https://image.slidesharecdn.com/dietze-onlinediscourse-ai-chatgpt-230517091816-a8c9ccfd/85/AI-in-between-online-and-offline-discourse-and-what-has-ChatGPT-to-do-with-all-of-that-21-320.jpg)
![AI biases: learned and elevated by AI/deep learning models
Source: https://techcrunch.com/2016/03/24/microsoft-silences-its-new-a-i-bot-tay-after-twitter-users-teach-it-racism/
35
[N-word]
• Biases in human interactions can be learned and elevated by ML models => fairness and ethics of AI, data
quality and representativeness
• Crucial: research into understanding of AI models](https://image.slidesharecdn.com/dietze-onlinediscourse-ai-chatgpt-230517091816-a8c9ccfd/85/AI-in-between-online-and-offline-discourse-and-what-has-ChatGPT-to-do-with-all-of-that-30-320.jpg)
AI in between online and offline discourse - and what has ChatGPT to do with all of that?
- 1. AI in between online and offline discourse - and what is the role of ChatGPT in all of that? Denkreise, Universität Bonn Stefan Dietze, 10.05.2023
- 2. 2 Discourse Interactions Algorithms/AI Discourse, interactions, AI: complex interdependencies Society, Media, Politics & Policies (Offline & Online)
- 3. 3 Discourse, interactions, AI: complex interdependencies AI-based methods to understand & influence online discourse & interactions (PART I) ▪ Recommendation & ranking in Web search (e.g. Google), shopping or social media ▪ Social media feeds & timelines ▪ Fact checking, detection of misinformation (e.g. news & social media) ▪ Classification of users (e.g. into political leanings, experts/lay people etc) ▪ Hate-speech detection ▪ Stance & sentiment detection ▪ … AI-based methods trained on & influencing data from discourse & interactions Discourse Interactions Algorithms/AI
- 4. 4 Discourse, interactions, AI: complex interdependencies Contemporary and future AI challenges (PART II) ▪ Self-reinforcing cycle between AI & online/offline behaviour: AI influenced by / trained on data from discourse/interactions that are in turn influenced by AI and so on ▪ Biases (in models & behaviour) => echo chambers, filter bubbles etc ▪ Transparency, reproducibility, decay ▪ …. Discourse Interactions Algorithms/AI
- 5. 5 AI for mining claims and stances in online discourse stance, claim trustworthiness? stance, claim trustworthiness?
- 6. 6 Detecting stances towards claims/opinions in online discourse Motivation ▪ Problem: detecting stance of documents (e.g. Web pages, scientific publication) towards a given claim (unbalanced class distribution) ▪ Motivation: stance of documents (in particular disagreement) useful (a) as signal for truthfulness (fake news detection) and (b) document or source classification (PLDs, publishers) Approach ▪ Cascading binary classifiers: addressing individual issues (e.g. misclassification costs) per step ▪ Features, e.g. textual similarity (Word2Vec etc), sentiments, LIWC, etc. ▪ Best-performing models: 1) SVM with class-wise penalty, 2) CNN, 3) SVM with class-wise penalty ▪ Experiments on FNC-1 dataset (and FNC baselines) Results ▪ Minor overall performance improvement ▪ Improvement on disagree class by 27% (but still far from robust) A. Roy, A. Ekbal, S. Dietze, P. Fafalios, Exploiting stance hierarchies for cost-sensitive stance detection of Web documents, IJIS2022
- 7. 7 ClaimsKG: a knowledge graph of fact-checked claims (to feed AI) Motivation ▪ Fact-checked claims spread across various (unstructured) fact-checking sites ▪ Example: finding claims about / made by US republican politicians across the Web? Approach ▪ Harvesting claims & metadata from fact- checking/news sites (e.g. snopes.com, Politifact.com etc); currently approx. 70.000 claims ▪ Information extraction & linking o Linking mentioned entities eg POTUS and Trump to „wikipedia:DonaldTrump“ o Normalisation of ratings (true, false, mixture, other); coreference resolution of claims o Exposing data through established vocabularies and W3C standards and APIs https://data.gesis.org/claimskg/ A. Tchechmedjiev, P. Fafalios, K. Boland, S. Dietze, B. Zapilko, K. Todorov, ClaimsKG – A Live Knowledge Graph of fact-checked Claims, ISWC2019
- 8. 8 ClaimsKG as training corpus for AI models for fact-checking ▪ Ground truth data for AI-based methods for fact- checking ▪ Initial step for any fact-checker when „check-worthy claims“ appear: „Claim Retrieval“ i.e. identifying if an arbitrary statement (eg from a news article) has already been fact-checked (ClaimsKG as knowledge base) ▪ Shared AI task: „CLEF2022-CheckThat! Lab“ ▪ Unsupervised approach „SimBA“ achieving state-of- the-art performance (P@1 = 94%) Hövelmeyer, A., Boland, K., Dietze, S., SimBa at CheckThat! 2022: Lexical and Semantic Similarity Based Detection of Verified Claims in an Unsupervised and Supervised Way, CLEF Working Notes 2022, CLEF 2022- Conference and Labs of the Evaluation Forum.
- 9. http://dbpedia.org/resource/Tim_Berners-Lee wna:positive-emotion onyx:Intensity "0.75" onyx:Intensity "0.0" http://dbpedia.org/resource/Solid wna:negative-emotion AI for mining opinions: the case of Twitter Challenges ▪ Heterogenity: multimodal, multilingual, informal, “noisy” language ▪ Context dependence: interpretation of tweets/posts (entities, sentiments) requires consideration of context (e.g. time, linked content), “Dusseldorf” => City or Football team ▪ Dynamics & scale: e.g. 6000 tweets per second, plus interactions (retweets etc) and context (e.g. 25% of tweets contain URLs) ▪ Evolution and temporal aspects: evolution of interactions over time crucial for many social sciences questions ▪ Representativity and bias: demographic distributions not known a priori in archived data collections
- 10. TweetsKB – a large-scale research corpus of societal opinions ▪ Harvesting & archiving of 14 Billion tweets (permanent collection from Twitter 1% sample since 2013) ▪ Information extraction pipeline to build a KG of entities, interactions & sentiments (distributed Map/Reduce batch processing) o Entity linking with knowledge graph/DBpedia (“president”/“potus”/”trump” => dbp:DonaldTrump) o Sentiment analysis/annotation o Geotagging o Lifting into knowledge graph schema Dimitrov, D., Baran, E., Fafalios, P., Yu, R., Zhu, X., Zloch, M., Dietze, S., TweetsCOV19 – A Knowledge Base of Semantically Annotated Tweets about the COVID-19 Pandemic, CIKM2020 https://data.gesis.org/tweetskb
- 11. https://data.gesis.org/tweetskb ▪ Harvesting & archiving of 14 Billion tweets (permanent collection from Twitter 1% sample since 2013) ▪ Information extraction pipeline to build a KG of entities, interactions & sentiments (distributed Map/Reduce batch processing) o Entity linking with knowledge graph/DBpedia (“president”/“potus”/”trump” => dbp:DonaldTrump) o Sentiment analysis/annotation o Geotagging o Lifting into knowledge graph schema ▪ Public, privacy-aware, large-scale research corpus of public opinions and their evolution => interdisciplinary research Dimitrov, D., Baran, E., Fafalios, P., Yu, R., Zhu, X., Zloch, M., Dietze, S., TweetsCOV19 – A Knowledge Base of Semantically Annotated Tweets about the COVID-19 Pandemic, CIKM2020 TweetsKB – a large-scale research corpus of societal opinions
- 12. Social science research using TweetsKB https://dd4p.gesis.org Investigating Vaccine Hesitancy in DACH countries
- 13. Germany suspends vaccinations with Astra Zeneca Twitter discourse zu “Impfbereitschaft” Social science research using TweetsKB Investigating Vaccine Hesitancy in DACH countries https://dd4p.gesis.org
- 14. Social science research using TweetsKB Vaccine Hesitancy– key topics in “safety” category „Schwangerschaft“ „Kimmich“ „Alter“ „Nebenwirkungen“ „Herzinfarkt“ „Zulassung“ https://dd4p.gesis.org
- 15. 17 ▪ AI4Sci project: understanding and classification of science discourse online (news, social Web) ▪ Percentage of tweets containing links to scientific articles (journals, publishers, science blogs etc) ▪ Uses list of > 30 K science web domains ▪ Data source: TweetsKB (https://data.gesis.org/tweetskb/), > 10 bn tweets archived since 2013 https://ai4sci-project.org/ How about scientific discourse on the Web? Example: Twitter
- 16. How about scientific discourse on the Web? Example: Twitter 18 https://ai4sci-project.org/ Science claim Science reference Science relevance No science Science reference Hafid, S., Schellhammer, S., Bringay, S., Todorov, K., Dietze, S., "SciTweets - A Dataset and Annotation Framework for Detecting Scientific Online Discourse", CIKM2022
- 17. SciTweets dataset & classifier 20 ▪ Ground truth dataset, heuristics-based sampling strategy and annotation framework for testing classification models ▪ 1261 expert-labeled tweets across all classes/labels ▪ Baseline classifiers based on SciBERT transformer model (fine-tuned/tested on SciTweets) ▪ Ongoing: analysis of large-scale science discourse and its evolution Hafid, S., Schellhammer, S., Bringay, S., Todorov, K., Dietze, S., SciTweets - A Dataset and Annotation Framework for Detecting Scientific Online Discourse, CIKM2022 https://ai4sci-project.org/
- 18. 21 But: do we actually need dedicated NLP methods after ChatGPT? • ChatGPT does not outperform specialized models in most NLP tasks • ChatGPT’s impressiveness stems from its capacity to do many things reasonably well • ChatGPT does not mark the end of AI/NLP research Source: http://opensamizdat.com/posts/chatgpt_survey/
- 19. 24 Discourse, interactions, AI: complex interdependencies Contemporary and future AI challenges (PART II) ▪ Self-reinforcing cycle between AI & online/offline behaviour: AI influenced by / trained on data from discourse/interactions that are in turn influenced by AI and so on ▪ Biases (in models & behaviour) => echo chambers, filter bubbles etc ▪ Transparency, reproducibility, decay ▪ …. Discourse Interactions Algorithms/AI
- 20. 25 What‘s the role of ChatGPT et al in all this? • ChatGPT didn’t come out of nowhere but is a continuation of trends throughout the last decades • Mostly powered by progress in (deep) representation & transfer learning
- 21. 26 What is driving the recent AI advances? Deep & representation learning for language understanding Percentage of deep learning papers in major NLP conferences (Source: Young et al., Recent Trends in Deep Learning Based Natural Language Processing) • Pretraining of embeddings: predicting low-dimensional vector representations of words & text, e.g. Word2Vec [Mikolov et al., 2013] • RNN/CNN architectures in encoder/decoder settings (e.g. for machine translation) [Vaswani et al., 2017] • Pretraining language models for task- specific transfer learning, e.g., BERT - Bidirectional Encoder Representations from Transformers [Devlin et al., 2018] T. Mikolov et al., Distributed Representations of Words and Phrases and their Compositionality, NIPS (2013) J. Devlin et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018) A. Vaswani et al. Attention is all you need, NIPS (2017)
- 22. 27 Transfer learning for language understanding Goal: Language Model (“reusable language representation”) (Paris:France, Tokyo:Japan) Approach: un-/self-supervised Machine Learning (e.g. „Masked Language Modeling“) Data: the more the merrier, e.g. Wikipedia, Google Book Corpus (200 Bn words), Common Crawl, Google crawl of indexed Web pages etc Data cheap, computation costly Goal: dedicated model for downstream task (e.g. dialog system, Tweet classification => Tweet X: RightWing) Approach: supervised machine learning, reinforcement learning Data: labelled data (e.g. expert annotated tweets, weak labels) Data costly, computation cheap Pretraining of foundational language models Finetuning/supervised ML („the old ways“)
- 23. 28 Transfer learning for language understanding Goal: Language Model (“reusable language representation”) (Paris:France, Tokyo:Japan) Approach: un-/self-supervised Machine Learning (e.g. „Masked Language Modeling“) Data: the more the merrier, e.g. Wikipedia, Google Book Corpus (200 Bn words), Common Crawl, Google crawl of indexed Web pages etc Data cheap, computation costly Goal: dedicated model for downstream task (e.g. dialog system, Tweet classification => Tweet X: RightWing) Approach: supervised machine learning, reinforcement learning Data: labelled data (e.g. expert annotated tweets, weak labels) Data costly, computation cheap Pretraining of foundational language models Finetuning/supervised ML („the old ways“) • Language understanding capabilities of models established during pretraining stage • Paradigm shift in NLP / NLU due to transferable foundational models • Little/no finetuning required for specific down-stream tasks („Zero-Shot Learning“) • Requires no costly data annotation but can be fed with massive volumes of text • „The larger the better“ (GPT3 = 175 Bn parameters)
- 24. 29 AI progress & societal impact/challenges ▪ Stopping the „arms race“? The bigger the better? ▪ Concentration of power: platform providers (gatekeeper to data), AI companies (gatekeeper to models, infrastructure) ▪ Environmental cost (Bender et al., On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?) ▪ Misinformation from generative models (e.g. deep fakes) ▪ But it‘s not only about AI: data, infrastructure, information has been controlled by large commercial platform providers for quite some time
- 25. 30 Crucial AI challenges that we/AI research is trying to tackle ▪ Transparency of AI models ▪ Access to data, infrastructure, models ▪ Reproducibility & explainability ▪ Bias of data & AI models ▪ Decay of AI models ▪ …
- 26. 31 AI transparency Source: https://motherboard.vice.com/en_us/article/j5npeg/why-is-google-translate-spitting-out-sinister-religious-prophecies/ • Neural machine translation as black box system => hard to pin down causes of „learned“ translations • Potential reasons: adversarial data fed into NN; NN learned to generate syntactically well formed sentences/statements even on nonsensical input
- 27. 32 AI transparency: how to evaluate AI if it outperforms humans? • Saturation of NLP/image processing benchmarks over time. Initial performance at -1, human performance at 0. • Challenges for performance evaluation • Creating novel kinds of AI benchmarks • Note: humans still outperform AI in many (less specialized) tasks/respects, e.g. “common-sense reasoning” Kiela, D. et al., Dynabench: Rethinking Benchmarking in NLP, NAACL2021
- 28. 33 AI decay: evolving language and discourse context ▪ LMs trained on large volumes of text ▪ Yet: strong vocabulary shift, over- /underrepresentation of topics/vocabulary in particular time periods (e.g. Twitter COVID19- discourse 2020 vs 2019) ▪ AI models for online discourse analysis require frequent training and updates (access to data/infrastructure?) ▪ Efficient methods for updating large language models Source: Hombaiah et al., “Dynamic Language Models for continuously evolving Content”, SIGKDD2021
- 29. 34 AI reproducibility: „A worrying analysis of neural recommender approaches” Dacrema, M. F., Cremonesi, P., Jannach, D., 2019. Are we really making much progress? A worrying analysis of recent neural recommendation approaches. In ACM RecSys2019. DOI:https://doi.org/10.1145/3298689.3347058 Are DL-based methods reproducible? Do they actually beat simple baselines? • Reproducibility crisis (due to lack of transparency) • State-of-the-art crisis (driven by peer review crisis)
- 30. AI biases: learned and elevated by AI/deep learning models Source: https://techcrunch.com/2016/03/24/microsoft-silences-its-new-a-i-bot-tay-after-twitter-users-teach-it-racism/ 35 [N-word] • Biases in human interactions can be learned and elevated by ML models => fairness and ethics of AI, data quality and representativeness • Crucial: research into understanding of AI models
- 31. 36 Model understanding: do LLMs actually „understand“ language? ▪ Language model (e.g. GPT) prompting as a way to assess its capacity to comprehend language („intelligence“) ▪ How does prompt syntax influence „knowledge retrieval“? Linzbach, S., Tressel, T., Kallmeyer, L., Dietze, S., Jabeen, H., 2023. Decoding Prompt Syntax: Analysing its Impact on Knowledge Retrieval in Large Language Models. In Companion Proceedings of the ACM Web Conference 2023 The role of syntax in question-answering through LLMs
- 32. 37 ▪ Experiments on BERT-based (Transformer) models ▪ Testing for sentence typology (simple, compound, etc) and active/passive/nominalised verb morphology Linzbach, S., Tressel, T., Kallmeyer, L., Dietze, S., Jabeen, H., 2023. Decoding Prompt Syntax: Analysing its Impact on Knowledge Retrieval in Large Language Models. In Companion Proceedings of the ACM Web Conference 2023 Model understanding: do LLMs actually „understand“ language? The role of syntax in question-answering through LLMs
- 33. 38 Model understanding: how to measure actual „intelligence“ of AI? • AI benchmark datasets/tests, eg to probe their commonsense reasoning skills, e.g. AI2Reasoning (ARC) • Models learn to perform well on a task while struggling to actually understand • Example: ChatGPT produces reasonable-sounding answers that are actually wrong • Constructing benchmark datasets for evaluation of AI models remains a huge challenge Branco et al., Shortcutted Commonsense: Data Spuriousness in Deep Learning of Commonsense Reasoning, EMNLP 2021
- 34. 39 To sum up: how to beat the cycle? ▪ AI research = creating understanding AI models, e.g. what and how do LLMs learn? (e.g. SFB Deep Linguistic Modeling) ▪ FAIR access to AI data & models (e.g. GESIS Methods Hub) ▪ Values-oriented ML: e.g. explainability > accuracy ▪ Realistic and fair benchmark datasets & setups for AI evaluation ▪ Maintainability: efficient ways on how to train/update foundational models (e.g. DD4P) ▪ Interdisiciplinary research into complex interdependencies between AI & online / offline discourse (e.g. NEWORDER) ▪ Governance & incentives: monopolies, ecological aspects and access to data/infrastructure are not only technical but societal & political issues Discourse Interactions Algorithms/AI Society, Media, Politics & Policies (Offline & Online)
- 36. 41 References ▪ Branco et al., Shortcutted Commonsense: Data Spuriousness in Deep Learning of Commonsense Reasoning, EMNLP 2021 ▪ Linzbach, S., Tressel, T., Kallmeyer, L., Dietze, S., Jabeen, H., 2023. Decoding Prompt Syntax: Analysing its Impact on Knowledge Retrieval in Large Language Models. In Companion Proceedings of the ACM Web Conference 2023 ▪ Dacrema, M. F., Cremonesi, P., Jannach, D., 2019. Are we really making much progress? A worrying analysis of recent neural recommendation approaches. In ACM RecSys2019. DOI:https://doi.org/10.1145/3298689.3347058 ▪ Kiela, D. et al., Dynabench: Rethinking Benchmarking in NLP, NAACL2021 ▪ Hombaiah et al., “Dynamic Language Models for continuously evolving Content”, SIGKDD2021 ▪ Hafid, S., Schellhammer, S., Bringay, S., Todorov, K., Dietze, S., SciTweets - A Dataset and Annotation Framework for Detecting Scientific Online Discourse, CIKM2022 ▪ Dimitrov, D., Baran, E., Fafalios, P., Yu, R., Zhu, X., Zloch, M., Dietze, S., TweetsCOV19 – A Knowledge Base of Semantically Annotated Tweets about the COVID-19 Pandemic, CIKM2020 ▪ A. Tchechmedjiev, P. Fafalios, K. Boland, S. Dietze, B. Zapilko, K. Todorov, ClaimsKG – A Live Knowledge Graph of fact-checked Claims, ISWC2019 ▪ Hövelmeyer, A., Boland, K., Dietze, S., SimBa at CheckThat! 2022: Lexical and Semantic Similarity Based Detection of Verified Claims in an Unsupervised and Supervised Way, CLEF Working Notes 2022, CLEF 2022- Conference and Labs of the Evaluation Forum. ▪ A. Roy, A. Ekbal, S. Dietze, P. Fafalios, Exploiting stance hierarchies for cost-sensitive stance detection of Web documents, IJIS2022