Tag: llm

How Large Language Models Are Changing Relation Extraction in NLP

When you type a question into a search engine like “Who wrote Hamlet?” it does more than match keywords. It connects the dots between “Shakespeare” and “Hamlet,” identifying the relationship between a person and their work. This process of finding and labelling relationships in text is called relation extraction (RE). It powers everything from knowledge graphs to fact-checking systems.

In the past, relation extraction systems were built with hand-crafted rules or required thousands of annotated examples to train. Now, large language models (LLMs) such as GPT, T5, and LLaMA are making it possible to do relation extraction with far less data and more flexibility. In this post, I want to explore what relation extraction is, how LLMs are transforming it, and why this matters for anyone interested in the future of language technology.

The Old Way Versus the New Way

Relation extraction used to rely heavily on feature engineering. Researchers would first hand-craft linguistic or statistical features from the text, such as part-of-speech tags, dependency parses, named entity types, and the words appearing between two entities. They often built lexical patterns like “X, the capital of Y” or “X, founded by Y” to capture specific relationships. They also extracted syntactic paths (for example, the shortest path between two entities in a dependency tree) or used windowed word embeddings around the entities to encode local context. These features were then fed into classifiers like Support Vector Machines, logistic regression models, or decision trees to decide which relationship best described the entity pair. Later, neural networks like BERT and RoBERTa replaced much of this manual feature design by learning their own representations directly from text, but they still needed large, labeled datasets to achieve good performance.

Today, LLMs can often handle relation extraction with just a prompt. Instead of training a model from scratch, you can ask an LLM directly: “Identify the relationship between ‘Einstein’ and ‘the theory of relativity.’” With the right instructions, the model can output the relation (“Einstein proposed the theory of relativity”) even if it has never seen that exact pair in training.

See the diagram showing “classic relation extraction pipeline” versus “LLM-based pipeline” below.

How LLMs Tackle Relation Extraction

LLMs have been trained on massive amounts of text, which gives them a broad sense of how entities and relationships co-occur. This makes them powerful at zero-shot and few-shot learning. In other words, you do not need to show them hundreds of examples of a relation type before they can identify it.

For example, a simple prompt could be:

Prompt:“Read the sentence: ‘Marie Curie discovered radium.’ What is the relationship between the entities?”
Model Output from ChatGPT-5 and Grok-4:

[ChatGPT-5] In the sentence “Marie Curie discovered radium”, the relationship between the entities is:

  • Marie Curie (the subject / agent) performed the action.
  • radium (the object / theme) is what was discovered.

This is a “discoverer–discovery” relationship (also called “person–invention/discovery” or “agent–object” in relation extraction). In other words, Marie Curie is the discoverer and radium is the thing discovered.

[Grok-4] In the sentence “Marie Curie discovered radium,” the relationship between the entities “Marie Curie” and “radium” is that of discoverer and discovered. Marie Curie, a scientist, is the person who identified or found the chemical element radium.

This ability to generalize makes LLMs more versatile, especially for relations in domains where labeled data is scarce, like medicine or law. However, it also introduces risks. LLMs may hallucinate relationships that are not actually in the text or mislabel subtle ones, so careful evaluation is still necessary.

Recent Research Highlights

A major paper, A Survey on Cutting-Edge Relation Extraction Techniques Based on Language Models (Diaz-Garcia & López, 2024), reviews 137 recent ACL papers (2020-2023) that use language models for relation extraction. It shows that BERT-based methods still lead many benchmarks while models like T5 are rising in few-shot and unseen-relation settings.

Other papers from ACL 2024 and 2025 explore how well LLMs handle unseen relation types, cross-domain relation extraction, and low-resource settings. These studies show steady improvements but also highlight open questions about factuality, bias, and consistency.

Why This Matters Beyond Academia

Relation extraction sits at the core of knowledge-driven applications. Building or updating a knowledge graph for a company’s internal documents, mapping patient histories in healthcare, or connecting laws to court cases in legal tech all depend on accurately identifying relationships between entities.

LLMs make it possible to automate these tasks more quickly. Instead of spending months labeling data, organizations can draft knowledge structures with an LLM, then have humans verify or refine the results. This speeds up research and decision-making while expanding access to insights that would otherwise stay hidden in text.

Challenges and Open Questions

While LLMs are powerful, they are not flawless. They may infer relationships that are plausible but incorrect, especially if the prompt is ambiguous. Evaluating relation extraction at scale is also difficult, because many relations are context-specific or only partially expressed. Bias in training data can also skew the relationships a model “sees” as likely or normal.

Researchers are now working on ways to add uncertainty estimates, retrieval-augmented methods (i.e., combining information retrieval with generative models to improve response accuracy and relevance), and better benchmarks to test how well models extract relations across different domains and languages.

My Take as a High Schooler Working in NLP

As someone who has built a survey analysis platform and published research papers about sentiment classification, I find relation extraction exciting because it can connect scattered pieces of information into a bigger picture. Specifically, for projects like my nonprofit Student Echo, a future system could automatically link student concerns to policy areas or resources.

At the same time, I am cautious. Seeing how easily LLMs generate answers reminds me that relationships in text are often subtle. Automating them risks oversimplifying complex realities. Still, the idea that a model can find and organize connections that would take a person hours to spot is inspiring and worth exploring.

Conclusion

Relation extraction is moving from hand-built rules and large labeled datasets to flexible, generalist large language models. This shift is making it easier to build knowledge graphs, extract facts, and understand text at scale. But it also raises new questions about reliability, fairness, and evaluation.

If you want to dig deeper, check out A Survey on Cutting-Edge Relation Extraction Techniques Based on Language Models (arXiv link) or browse ACL 2024–2025 papers on relation extraction. Watching how this field evolves over the next few years will be exciting, and I plan to keep following it for future blog posts.

— Andrew

5,286 hits

0

Latest Applications of NLP to Recommender Systems at RecSys 2025

Introduction

The ACM Conference on Recommender Systems (RecSys) 2025 took place in Prague, Czech Republic, from September 22–26, 2025. The event brought together researchers and practitioners from academia and industry to present their latest findings and explore new trends in building recommendation technologies.

This year, one of the most exciting themes was the growing overlap between natural language processing (NLP) and recommender systems. Large language models (LLMs), semantic clustering, and text-based personalization appeared everywhere, showing how recommender systems are now drawing heavily on computational linguistics. As someone who has been learning more about NLP myself, it is really cool to see how the research world is pushing these ideas forward.

Paper Highlights

A Language Model-Based Playlist Generation Recommender System

Paper Link

Relevance:
Uses language models to generate playlists by creating semantic clusters from text embeddings of playlist titles and track metadata. This directly applies NLP for thematic coherence and semantic similarity in music recommendations.

Abstract:
The title of a playlist often reflects an intended mood or theme, allowing creators to easily locate their content and enabling other users to discover music that matches specific situations and needs. This work presents a novel approach to playlist generation using language models to leverage the thematic coherence between a playlist title and its tracks. Our method consists in creating semantic clusters from text embeddings, followed by fine-tuning a transformer model on these thematic clusters. Playlists are then generated considering the cosine similarity scores between known and unknown titles and applying a voting mechanism. Performance evaluation, combining quantitative and qualitative metrics, demonstrates that using the playlist title as a seed provides useful recommendations, even in a zero-shot scenario.

An Off-Policy Learning Approach for Steering Sentence Generation towards Personalization

Paper Link

Relevance:
Focuses on off-policy learning to guide LLM-based sentence generation for personalized recommendations. Involves NLP tasks like controlled text generation and personalization via language model fine-tuning.

Abstract:
We study the problem of personalizing the output of a large language model (LLM) by training on logged bandit feedback (e.g., personalizing movie descriptions based on likes). While one may naively treat this as a standard off-policy contextual bandit problem, the large action space and the large parameter space make naive applications of off-policy learning (OPL) infeasible. We overcome this challenge by learning a prompt policy for a frozen LLM that has only a modest number of parameters. The proposed Direct Sentence Off-policy gradient (DSO) effectively propagates the gradient to the prompt policy space by leveraging the smoothness and overlap in the sentence space. Consequently, DSO substantially reduces variance while also suppressing bias. Empirical results on our newly established suite of benchmarks, called OfflinePrompts, demonstrate the effectiveness of the proposed approach in generating personalized descriptions for movie recommendations, particularly when the number of candidate prompts and reward noise are large.

Enhancing Sequential Recommender with Large Language Models for Joint Video and Comment Recommendation

Paper Link

Relevance:
Integrates LLMs to enhance sequential recommendations by processing video content and user comments. Relies on NLP for joint modeling of multimodal text (like comments) and semantic user preferences.

Abstract:
Nowadays, reading or writing comments on captivating videos has emerged as a critical part of the viewing experience on online video platforms. However, existing recommender systems primarily focus on users’ interaction behaviors with videos, neglecting comment content and interaction in user preference modeling. In this paper, we propose a novel recommendation approach called LSVCR that utilizes user interaction histories with both videos and comments to jointly perform personalized video and comment recommendation. Specifically, our approach comprises two key components: sequential recommendation (SR) model and supplemental large language model (LLM) recommender. The SR model functions as the primary recommendation backbone (retained in deployment) of our method for efficient user preference modeling. Concurrently, we employ a LLM as the supplemental recommender (discarded in deployment) to better capture underlying user preferences derived from heterogeneous interaction behaviors. In order to integrate the strengths of the SR model and the supplemental LLM recommender, we introduce a two-stage training paradigm. The first stage, personalized preference alignment, aims to align the preference representations from both components, thereby enhancing the semantics of the SR model. The second stage, recommendation-oriented fine-tuning, involves fine-tuning the alignment-enhanced SR model according to specific objectives. Extensive experiments in both video and comment recommendation tasks demonstrate the effectiveness of LSVCR. Moreover, online A/B testing on KuaiShou platform verifies the practical benefits of our approach. In particular, we attain a cumulative gain of 4.13% in comment watch time.

LLM-RecG: A Semantic Bias-Aware Framework for Zero-Shot Sequential Recommendation

Paper Link

Relevance:
Addresses domain semantic bias in LLMs for cross-domain recommendations using generalization losses to align item embeddings. Employs NLP techniques like pretrained representations and semantic alignment to mitigate vocabulary differences across domains.

Abstract:
Zero-shot cross-domain sequential recommendation (ZCDSR) enables predictions in unseen domains without additional training or fine-tuning, addressing the limitations of traditional models in sparse data environments. Recent advancements in large language models (LLMs) have significantly enhanced ZCDSR by facilitating cross-domain knowledge transfer through rich, pretrained representations. Despite this progress, domain semantic bias arising from differences in vocabulary and content focus between domains remains a persistent challenge, leading to misaligned item embeddings and reduced generalization across domains.

To address this, we propose a novel semantic bias-aware framework that enhances LLM-based ZCDSR by improving cross-domain alignment at both the item and sequential levels. At the item level, we introduce a generalization loss that aligns the embeddings of items across domains (inter-domain compactness), while preserving the unique characteristics of each item within its own domain (intra-domain diversity). This ensures that item embeddings can be transferred effectively between domains without collapsing into overly generic or uniform representations. At the sequential level, we develop a method to transfer user behavioral patterns by clustering source domain user sequences and applying attention-based aggregation during target domain inference. We dynamically adapt user embeddings to unseen domains, enabling effective zero-shot recommendations without requiring target-domain interactions.

Extensive experiments across multiple datasets and domains demonstrate that our framework significantly enhances the performance of sequential recommendation models on the ZCDSR task. By addressing domain bias and improving the transfer of sequential patterns, our method offers a scalable and robust solution for better knowledge transfer, enabling improved zero-shot recommendations across domains.

Trends Observed

These papers reflect a broader trend at RecSys 2025 toward hybrid NLP-RecSys approaches, with LLMs enabling better handling of textual side information (like reviews, titles, and comments) for cold-start problems and cross-domain generalization. This aligns with recent surveys on LLMs in recommender systems, which note improvements in semantic understanding over traditional embeddings.

Final Thoughts

As a high school student interested in computational linguistics, reading about these papers feels like peeking into the future. I used to think of recommender systems as black boxes that just show you more videos or songs you might like. But at RecSys 2025, it is clear the field is moving toward systems that actually “understand” language and context, not just click patterns.

For me, that is inspiring. It means the skills I am learning right now, from studying embeddings to experimenting with sentiment analysis, could actually be part of real-world systems that people use every day. It also shows how much crossover there is between disciplines. You can be into linguistics, AI, and even user experience design, and still find a place in recommender system research.

Seeing these studies also makes me think about the responsibility that comes with more powerful recommendation technology. If models are becoming better at predicting our tastes, we have to be careful about bias, fairness, and privacy. This is why conferences like RecSys are so valuable. They are a chance for researchers to share ideas, critique each other’s work, and build a better tech future together.

— Andrew

5,286 hits

0

Rethinking AI Bias: Insights from Professor Resnik’s Position Paper

I recently read Professor Philip Resnik’s thought-provoking position paper, “Large Language Models Are Biased Because They Are Large Language Models,” published in Computational Linguistics 51(3), which is available via open access. This paper challenges conventional perspectives on bias in artificial intelligence, prompting a deeper examination of the inherent relationship between bias and the foundational design of large language models (LLMs). Resnik’s primary objective is to stimulate critical discussion by arguing that harmful biases are an inevitable outcome of the current architecture of LLMs. The paper posits that addressing these biases effectively requires a fundamental reevaluation of the assumptions underlying the design of AI systems driven by LLMs.

What the paper argues

  • Bias is built into the very goal of an LLM. A language model tries to predict the next word by matching the probability patterns of human text. Those patterns come from people. People carry stereotypes, norms, and historical imbalances. If an LLM learns the patterns faithfully, it learns the bad with the good. The result is not a bug that appears once in a while. It is a direct outcome of the objective the model optimizes.
  • Models cannot tell “what a word means” apart from “what is common” or “what is acceptable.” Resnik uses a nurse example. Some facts are definitional (A nurse is a kind of healthcare worker). Other facts are contingent but harmless (A nurse is likely to wear blue clothing at work). Some patterns are contingent and harmful if used for inference (A nurse is likely to wear a dress to a formal occasion). Current LLMs do not have an internal line that separates meaning from contingent statistics or that flags the normative status of an inference. They just learn distributions.
  • Reinforcement Learning from Human Feedback (RLHF) and other mitigations help on the surface, but they have limits. RLHF tries to steer a pre-trained model toward safer outputs. The process relies on human judgments that vary by culture and time. It also has to keep the model close to its pretraining, or the model loses general ability. That tradeoff means harmful associations can move underground rather than disappear. Some studies even find covert bias remains after mitigation (Gallegos et al. 2024; Hofmann et al. 2024). To illustrate this, consider an analogy: The balloon gets squeezed in one place, then bulges in another.
  • The root cause is a hard-core, distribution-only view of language. When meaning is treated as “whatever co-occurs with what,” the model has no principled way to encode norms. The paper suggests rethinking foundations. One direction is to separate stable, conventional meaning (like word sense and category membership) from contextual or conveyed meaning (which is where many biases live). Another idea is to modularize competence, so that using language in socially appropriate ways is not forced to emerge only from next-token prediction. None of this is easy, but it targets the cause rather than only tuning symptoms.

Why this matters

Resnik is not saying we should give up. He is saying that quick fixes will not fully erase harm when the objective rewards learning whatever is frequent in human text. If we want models that reason with norms, we need objectives and representations that include norms, not only distributions.

Conclusion

This paper offers a clear message. Bias is not only a content problem in the data. It is also a design problem in how we define success for our models. If the goal is to build systems that are both capable and fair, then the next steps should focus on objectives, representations, and evaluation methods that make room for norms and constraints. That is harder than prompt tweaks, but it is the kind of challenge that can move the field forward.

Link to the paper: Large Language Models Are Biased Because They Are Large Language Models

— Andrew

5,286 hits

0

Summary: “Large Language Models Are Improving Exponentially”

I recently read an article on IEEE Spectrum titled “Large Language Models Are Improving Exponentially”. Here is a summary of its key points.

Benchmarking LLM Performance

Benchmarking large language models (LLMs) is challenging because their main goal is to produce text indistinguishable from human writing, which doesn’t always correlate with traditional processor performance metrics. However, it remains important to measure their progress to understand how much better LLMs are becoming over time and to estimate when they might complete substantial tasks independently.

METR’s Findings on Exponential Improvement

Researchers at Model Evaluation & Threat Research (METR) in Berkeley, California, published a paper in March called Measuring AI Ability to Complete Long Tasks. They concluded that:

  • The capabilities of key LLMs are doubling every seven months.
  • By 2030, the most advanced LLMs could complete, with 50 percent reliability, a software-based task that would take humans a full month of 40-hour workweeks.
  • These LLMs might accomplish such tasks much faster than humans, possibly within days or even hours.

Potential Tasks by 2030

Tasks that LLMs might be able to perform by 2030 include:

  • Starting up a company
  • Writing a novel
  • Greatly improving an existing LLM

According to AI researcher Zach Stein-Perlman, such capabilities would come with enormous stakes, involving both potential benefits and significant risks.

The Task-Completion Time Horizon Metric

At the core of METR’s work is a metric called “task-completion time horizon.” It measures the time it would take human programmers to complete a task that an LLM can complete with a specified reliability, such as 50 percent.

Their plots (see graphs below) show:

  • Exponential growth in LLM capabilities with a doubling period of around seven months (Graph at the top).
  • Tasks that are “messier” or more similar to real-world scenarios remain more challenging for LLMs (Graph at the bottom).

Caveats About Growth and Risks

While these results raise concerns about rapid AI advancement, METR researcher Megan Kinniment noted that:

  • Rapid acceleration does not necessarily result in “massively explosive growth.”
  • Progress could be slowed by factors such as hardware or robotics bottlenecks, even if AI systems become very advanced.

Final Summary

Overall, the article emphasizes that LLMs are improving exponentially, potentially enabling them to handle complex, month-long human tasks by 2030. This progress comes with significant benefits and risks, and its trajectory may depend on external factors like hardware limitations.

You can read the full article here.

— Andrew

0

Caring Machines, Centered Humans: Lessons from Ai4 2025

At Ai4 2025 (August 11–13, Las Vegas), two of the most influential voices in artificial intelligence expressed strikingly different visions for the future. Geoffrey Hinton, often called the “Godfather of AI,” suggested that AI should be designed with something like “maternal instincts.” He argued that as AI becomes smarter than humans, we cannot realistically control it through traditional dominance strategies. The only model we have of a less intelligent being guiding a more intelligent one is the relationship between a baby and its mother. A mother cares for her child not because she is weaker, but because she is built to protect and nurture. Hinton believes this kind of protective orientation is what could keep humanity safe in the long run.

Fei-Fei Li, sometimes called the “Godmother of AI,” offered a different perspective in a CNN interview. She disagrees with parental analogies for AI. Instead, she emphasizes designing human-centered AI, systems that uphold human dignity, promote agency, and avoid emotional metaphors that could mislead how we understand AI.

Summary Comparison of Views

When I first read about these contrasting views, I found myself agreeing with both in different ways. On one hand, Hinton’s maternal metaphor captures the seriousness of what could happen if superintelligence arrives sooner than many expect. If AI truly surpasses human intelligence, relying solely on control may fail. On the other hand, Li’s approach feels grounded and practical. She reminds us that the ethical choices we make today will set the trajectory for future systems.

The best answer may not lie in choosing between them, but in combining their strengths. I think about this as a layered model. The foundation should be Li’s human-centered AI: respect, fairness, transparency, and agency. On top of that we need what Hinton calls protective alignment. These would be structural safeguards that ensure highly intelligent systems still act in ways that preserve human well-being.

Hybrid Framework Diagram
Here is how I visualize this combination of perspectives: Li’s human-centered AI forms the core, while Hinton’s protective alignment provides the outer safeguard.

Practical Integration

  • Development Phase (Near-Term, Li):
    Apply human-centered AI frameworks to today’s large language models, robotics, and decision-support systems.
    Focus on privacy, bias reduction, explainability, and giving users agency over their interactions with AI.
  • Safety Research Phase (Mid- to Long-Term, Hinton):
    Begin embedding structural safeguards that mimic “caring instincts.”
    Example: AI systems with hard-coded prohibitions against harming humans, but reinforced by higher-order goals like proactively ensuring human thriving.
  • Governance and Oversight:
    Combine Li’s push for international, human-centered AI policy with Hinton’s insistence on global collaboration to avoid runaway dominance races.

In other words, AI should be designed to treat humanity as worth protecting, while being anchored in the principles of dignity.

As a high school student exploring AI and computational linguistics, I believe this hybrid vision is the most realistic path forward. It addresses the near-term challenges of fairness, transparency, and accountability while also preparing for the long-term risks of superintelligence. For me, this is not just an abstract debate. Thinking about how we embed values and safety into AI connects directly to my own interests in language models, hate speech detection, robotics, and how technology interacts with human society.

The future of AI is not predetermined. It will be shaped by the principles we choose now. By combining Hinton’s call for protective instincts with Li’s insistence on human-centered design, we have a chance to build AI that both cares for us and respects us.

For readers interested in the original coverage of this debate, see the CNN article here.

— Andrew

0

The AI Gap: How Socioeconomic Status Shapes Language Technology Use — A Perspective from Best Social Impact Paper at ACL 2025

The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025) recently finished in Vienna, Austria from July 27 to August 1. The conference announced a few awards, one of which is Best Social Impact Paper. This award was given to two papers:

  1. AfriMed-QA: A Pan-African, Multi-Specialty, Medical Question-Answering Benchmark Dataset (by Charles Nimo et al.)
  2. The AI Gap: How Socioeconomic Status Affects Language Technology Interactions (by Elisa Bassignana, Amanda Cercas Curry, and Dirk Hovy).

In this blog post, I’ll talk about the second paper and share the findings from the paper and my thoughts on the topic. You can read the full paper here: https://aclanthology.org/2025.acl-long.914.pdf

What the Paper is About

This paper investigates how socioeconomic status (SES) influences interactions with language technologies, particularly large language models (LLMs) like ChatGPT, highlighting an emerging “AI Gap” that could exacerbate social inequalities. Drawing from the Technology Acceptance Model and prior work on digital divides, the authors argue that SES shapes technology adoption through factors like access, digital literacy, and linguistic habits, potentially biasing LLMs toward higher-SES patterns and underrepresenting lower-SES users.

Methods

The study surveys 1,000 English-speaking participants from the UK and US via Prolific, stratified by self-reported SES using the MacArthur scale (binned as low: 1-3, middle: 4-7, upper: 8-10). It collects sociodemographic data, usage patterns of language technologies (e.g., spell checkers, AI chatbots), and 6,482 real prompts from prior LLM interactions. Analysis includes statistical tests (e.g., chi-square for usage differences), linguistic metrics (e.g., prompt length, concreteness via Brysbaert et al.’s word ratings), topic modeling (using embeddings, UMAP, HDBSCAN, and GPT-4 for cluster descriptions), and markers of anthropomorphism (e.g., phatic expressions like “hi” and politeness markers like “thank you”).

Key Findings

  • Usage Patterns: Higher-SES individuals access more devices daily (e.g., laptops, smartwatches) and use LLMs more frequently (e.g., daily vs. rarely for lower SES). They employ LLMs for work/education (e.g., coding, data analysis, writing) and technical contexts, while lower-SES users favor entertainment, brainstorming, and general knowledge queries. Statistically significant differences exist in frequency (p < 0.001), contexts (p < 0.001), and tasks (p < 0.001).
  • Linguistic Differences in Prompts: Higher-SES prompts are shorter (avg. 18.4 words vs. 27.0 for low SES; p < 0.05) and more abstract (concreteness score: 2.57 vs. 2.66; p < 0.05). Lower-SES prompts show higher anthropomorphism (e.g., more phatic expressions) and concrete language. A bag-of-words classifier distinguishes SES groups (Macro-F1 39.25 vs. baseline 25.02).
  • Topics and Framing: Common topics (e.g., translation, mental health, medical advice, writing, text editing, finance, job, food) appear across groups, but framing varies—e.g., lower SES seeks debt reduction or low-skill jobs; higher SES focuses on investments, travel itineraries, or inclusivity. About 45% of prompts resemble search-engine queries, suggesting LLMs are replacing traditional searches.
  • User Perceptions: Trends indicate lower-SES users anthropomorphize more (e.g., metaphorical verbs like “ask”), while higher-SES use jargon (e.g., “generate”), though not statistically significant.

Discussion and Implications

The findings underscore how SES stratifies LLM use, with higher-SES benefiting more in professional/educational contexts, potentially widening inequalities as LLMs optimize for their patterns. Benchmarks may overlook lower-SES styles, leading to biases. The authors advocate the development of inclusive NLP technologies to accommodate different SES needs and habitus and mitigate the existing AI Gap.

Limitations and Ethics

Limited to Prolific crowdworkers (skewed middle/low SES, tech-savvy), subjective SES measures, and potential LLM-generated responses. Ethical compliance includes GDPR anonymity, opt-outs, and fair compensation (£9/hour).

Overall, the paper reveals SES-driven disparities in technology interactions, urging NLP development to address linguistic and habitual differences for equitable access and reduced digital divides.

My Takeaway

As a high school student who spends a lot of time thinking about fairness in AI, I find this paper important because it reminds us that bias is not just about language or culture, it can also be tied to socioeconomic status. This is something I had not thought much about before. If AI systems are trained mostly on data from higher SES groups, they might misunderstand or underperform for people from lower SES backgrounds. That could affect how well people can use AI for education, job searching, or even just getting accurate information online.

For me, the takeaway is that AI researchers need to test their models with SES diversity in mind, just like they do with gender or language diversity. And as someone interested in computational linguistics, it is inspiring to see that work like this is getting recognized with awards at ACL.

— Andrew

1

Speeding Up AI for Everyone: The PaPaformer Model Making Language Tech Work on Phones and Low-Power Devices

AI has become more capable than ever, but many of the most advanced tools still require massive cloud servers to run. That means if you want ChatGPT-level performance, you usually need a reliable internet connection and a lot of computing power behind the scenes. But what if you could have that kind of AI right on your phone, even without Wi‑Fi?

That’s where the PaPaformer model comes in.

What is the PaPaformer Model?
PaPaformer is a new AI architecture developed to train large language models more efficiently and make them small enough to run smoothly on low-power devices like smartphones, tablets, or even embedded systems. You can read more about it in the original paper here: PaPaformer: Language Model from Pre-trained Parallel Paths.

Unlike most large models today that require powerful cloud servers to process requests, PaPaformer is designed so the model can be stored and run directly on your device. This means you can use advanced language technology without a constant internet connection. It also helps protect privacy, since your data stays local instead of being sent to the cloud for processing.

Why It Matters
By making AI lighter and more portable, PaPaformer could bring powerful language tools to more people around the world, including those with limited internet access or older devices. It could also make AI faster to respond, since it does not have to constantly send data back and forth to the cloud.

Examples in Action
Imagine using ChatGPT-style features on a budget smartphone in a remote area. With most current apps, like the regular ChatGPT app, you still need a strong internet connection because the AI runs on servers, not your device. But with a PaPaformer-powered tool, the AI would actually run locally, meaning you could:

  • Translate between languages instantly, even without Wi‑Fi
  • Use a speech-to-text tool for endangered languages that works entirely on your device
  • Let teachers translate lessons in real time for students in rural schools without relying on an internet connection
  • Help students write essays in multiple languages privately, without sending drafts to a remote server

This offline capability is the big difference. It is not just accessing AI through the cloud, it is carrying the AI with you wherever you go.

Looking Ahead
If PaPaformer and similar approaches keep improving, we could see a future where advanced AI is available to anyone, anywhere, without needing expensive devices or constant internet access. For someone like me, interested in computational linguistics, this could also open up new possibilities for preserving languages, creating translation tools, and making language technology more inclusive worldwide.

— Andrew

0

Is the Increasing Trend of Leveraging LLMs like ChatGPT in Writing Research Papers Concerning?

On August 4, 2025, Science published a tech news piece titled “One-fifth of computer science papers may include AI content,” written by Phie Jacobs, a general assignment reporter at Science. The article reports on a large-scale analysis conducted by researchers at Stanford University and the University of California, Santa Barbara. They examined over 1 million abstracts and introductions and found that by September 2024, 22.5% of computer science papers showed signs of input from large language models such as ChatGPT. The researchers used statistical modeling to detect common word patterns linked to AI-generated writing.

This caught my attention because I was surprised at how common AI-generated content has already become in academic research. I agree with the concern raised in the article, particularly this point:

Although the new study primarily looked at abstracts and introductions, Dmitry Kobak (University of Tübingen data scientist) worries authors will increasingly rely on AI to write sections of scientific papers that reference related works. That could eventually cause these sections to become more similar to one another and create a “vicious cycle” in the future, in which new LLMs are trained on content generated by other LLMs.

From my own experience writing research papers over the past few years, I can see why this concern is valid. If you have followed my blog, you know I have published two research papers and am currently working on a third. While working on my papers, I occasionally used ChatGPT (including its Deep Research) to help find peer-reviewed sources for citations instead of relying solely on search engines like Google Scholar. However, I quickly realized that depending on ChatGPT for this task can be risky. In my case, about 30% of the citations it provided were inaccurate, which meant I had to verify each one manually. For reliable academic sourcing, I found Google Scholar much more trustworthy because current LLMs are still prone to “hallucinations.” You may have encountered other AI tools like Consensus AI, a search engine tailored for scientific research and limited to peer-reviewed academic papers only. Compared to ChatGPT Deep Research, it’s faster and more reliable for academic queries, but I strongly recommend always verifying AI outputs, as both tools can occasionally produce inaccuracies.

The Science article also highlights that AI usage varies significantly across disciplines. “The amount of artificial intelligence (AI)-modified sentences in scientific papers had surged by September 2024, almost two years after the release of ChatGPT, according to an analysis.” The table below shows estimates of AI usage by field, with certain disciplines adopting AI much faster than others. James Zou, a computational biologist at Stanford University, suggests these differences may reflect varying levels of familiarity with AI technology.

While the study from Stanford and UCSB is quite solid, Data Scientist Kobak pointed out that the estimates above could be underreported. One reason for this is that some authors may have started removing “red flag” words from manuscripts to avoid detection. For example, the word “delve” became more common right after ChatGPT launched, but its usage dropped sharply once it became widely recognized as a hallmark of AI-generated text.

If you want to read the full article, you can find it here: Science – One-fifth of computer science papers may include AI content.

— Andrew

Update: Here is another more recent report from Nature.

1

How NLP Helps Robots Handle Interruptions: A Summary of JHU Research

I recently came across an awesome study from Johns Hopkins University describing how computational linguistics and NLP can make robots better conversational partners by teaching them how to handle interruptions, a feature that feels basic for humans but is surprisingly hard for machines.

What the Study Found

Researchers trained a social robot powered by a large language model (LLM) to manage real-time interruptions based on speaker intent. They categorized interruptions into four types: Agreement, Assistance, Clarification, and Disruption.

By analyzing human conversations from interviews to informal discussions, they designed strategies tailored to each interruption type. For example:

  • If someone agrees or helps, the robot pauses, nods, and resumes speaking.
  • When someone asks for clarification, the robot explains and continues.
  • For disruptive interruptions, the robot can either hold the floor to summarize its remaining points before yielding to the human user, or it can stop talking immediately.

How NLP Powers This System

The robot uses an LLM to:

  1. Detect overlapping speech
  2. Classify the interrupter’s intent
  3. Select the appropriate response strategy

In tests involving tasks and conversations, the system correctly interpreted interruptions about 89% of the time and responded appropriately 93.7% of the time.

Why This Matters in NLP and Computational Linguistics

This work highlights how computational linguistics and NLP are essential to human-robot interaction.

  • NLP does more than generate responses; it helps robots understand nuance, context, and intent.
  • Developing systems like this requires understanding pause cues, intonation, and conversational flow, all core to computational linguistics.
  • It shows how multimodal AI, combining language with behavior, can enable more natural and effective interactions.

What I Found Most Interesting

The researchers noted that users didn’t like when the robot “held the floor” too long during disruptive interruptions. It reminded me how pragmatic context matters. Just like people expect some rules in human conversations, robots need these conversational skills too.

Looking Ahead

This research expands what NLP can do in real-world settings like healthcare, education, and social assistants. For someone like me who loves robots and language, it shows how computational linguistics helps build smarter, more human-friendly AI systems.

If you want to dive deeper, check out the full report from Johns Hopkins:
Talking robots learn to manage human interruptions

— Andrew

0

WAIC 2025: What Geoffrey Hinton’s “Tiger” Warning Taught Me About AI’s Future

At the end of July (7/26 – 7/28), Shanghai hosted the 2025 World Artificial Intelligence Conference (WAIC), drawing over 1,200 participants from more than 40 countries. Even though I wasn’t there, I followed the conference closely, especially the keynote from Geoffrey Hinton, the so-called “Godfather of AI.” His message? AI is advancing faster than we expect, and we need global cooperation to make sure it stays aligned with human values.

Hinton’s talk was historic. It was his first public appearance in China, and he even stood throughout his address despite back pain, which was noted by local media. One quote really stuck with me: “Humans have grown accustomed to being the most intelligent species in the world – what if that’s no longer the case?” That’s a big question, and as someone who’s diving deeper into computational linguistics and large language models, I felt both amazed and a little uneasy.

His warning compared superintelligent AI to a tiger we’re raising as a pet. If we’re not careful, he said, “the tiger” might one day turn on us. The point wasn’t to scare everyone, but to highlight why we can’t rely on simply pulling the plug if AI systems surpass human intelligence. Hinton believes we need to train AI to be good from the beginning because shutting it down later might not be an option.

WAIC 2025 wasn’t all doom and gloom though. Hinton also talked about the huge potential of AI to accelerate science. For example, he highlighted DeepMind’s AlphaFold as a breakthrough that solved a major biology challenge, predicting protein structures. That shows how powerful AI can be when guided properly.

What stood out the most was the recurring theme of cooperation. Hinton and others, like former Google CEO Eric Schmidt, emphasized the need for global partnerships on AI safety and ethics. Hinton even signed the “Shanghai AI Safety Consensus” with other experts to support international collaboration. The message was clear: no single country can or should handle AI’s future alone.

As a high school student passionate about AI and language, I’m still learning how these pieces fit together. But events like WAIC remind me that the future of AI isn’t just about building smarter systems, it’s also about making sure they work for everyone.

For those interested, here’s a more detailed summary of Hinton’s latest speech: Pandaily Report on WAIC 2025

You can also explore the official WAIC website here: https://www.worldaic.com.cn/

— Andrew

0

Blog at WordPress.com.

Up ↑