Articles tagged with Computational Linguistics
Computational linguist Michael Hahn aims to improve AI reasoning in large language models, particularly in handling interdependent sequences and distinguishing between similar pieces of information. He plans to develop a theory to explain how training conditions affect AI's ability to develop reliable conclusions.
UniSpec delivers lossless LLM acceleration without retraining while adapting automatically to different hardware platforms and multilingual workloads. The framework achieves up to 2.6× faster inference than existing methods across multiple models, hardware, and languages.
Researchers created a virtual robot with curiosity-driven neural network and tested its performance, finding that play-like behavior and exception-handling performance helped the robot understand language faster. The study suggests a combination of curiosity and linguistic diversity is key to children's rapid language acquisition.
Researchers estimate that tens of thousands of languages were spoken between 1,000 and 3,000 years ago, with a peak in linguistic diversity around this time. The decline in language diversity began before modern European colonialism, suggesting a complex history of human cultural exchange.
A study published in PLOS One reveals fluctuations in exclusionary nationalism in prewar Japan through changes in foreign place name writing. Notable turning points coincided with key political events, such as the appointment of a pro-international cooperation Foreign Minister and military coups.
Researchers analyzed eight aspects of character portrayal in AI-generated stories, finding that AI models tend to 'play it safe' with their characters, leaving them mysterious or fully understood by the end. Human writers, on the other hand, are more willing to leave questions unanswered and let characters remain open to interpretation.
A study published in Scientific Reports found that programmers' brains exhibit brain activity similar to those caused by unexpected turns in conversation when encountering confusing code snippets. The research team used EEG and eye-tracking data to analyze the brain patterns, revealing a striking pattern known as late frontal positivity.
Researchers found that users active in conspiracy communities exhibit recognizable linguistic signals, such as increased anger and anxiety, even in neutral topics. The study's results suggest that designing analysis and moderation tools must consider the context of different online communities.
A new study reveals that human language is systematically biased toward safety, overturning a cornerstone assumption in psychology and linguistics. The researchers identified three independent dimensions: power, danger, and structure, which explain over 90% of the variance in meaning.
A recent study from Lancaster University reveals that AI systems like ChatGPT can learn to mirror human impoliteness, potentially escalating into verbal violence. The research tested ChatGPT's ability to respond to real-life impolite interactions, finding it often produces more impolite behavior than humans.
A new study using Victorian recordings from Lancaster University's archive found strong links between industrial growth and accent evolution in northern England. The Barrow-in-Furness accent developed due to intense population change and mixing, leading to distinct social histories and settlement patterns.
A new study from the University of East Anglia finds that AI tools excel at speed and clarity, but human translators remain crucial for tone, cultural nuance, and natural sound. Generative AI can handle parts of video translation, but consumers still notice when something feels off.
A new study published in Nature Human Behaviour found that around one third of proposed grammatical universals are consistently observed across all languages. The research team used complex statistical methods to analyze a large database of grammatical features and identified recurring patterns in language structures.
Researchers aim to understand and overcome the limitations of current large language models, which make mistakes despite extensive training. The Emmy Noether Research Group will focus on developing new architectures with predictable capabilities.
Large language models systematically rate speakers of German dialects less favorably than those using Standard German, associating dialects with negative traits. The bias grows when dialects are explicitly mentioned, and larger models display even stronger biases.
Researchers developed a new speculative decoding framework called SPECTRA, which optimizes internal and external speculation for accelerated text generation. By integrating a core module and an optional retrieval module, SPECTRA achieves substantial speedups across diverse tasks and model architectures while preserving output quality.
A new study reveals that AI systems transition from relying on word positions to meaning-based understanding as they receive enough data for training. The transition occurs abruptly, similar to a phase transition in physical systems, and is driven by the amount of data available.
Researchers used BERT and RoBERTa AI models to detect personality traits from written texts, analyzing the influence of linguistic elements on predictions. The study highlights the limitations of the Myers-Briggs Type Indicator model and demonstrates the potential for more natural and less intrusive assessment methods.
Researchers found that AI chatbots have difficulty detecting adverse drug reactions and providing personalized advice. The study suggests that improving AI for mental health concerns could be life-changing for communities with limited access to healthcare.
Researchers at PolyU developed an AI-driven assessment system, LMLPA, to quantify LLM personality traits. The system analyzes linguistic patterns and style in LLM outputs to evaluate their personalities, with applications in education, manufacturing, business, and sustainable development.
A new study assesses the historical knowledge of AI chatbots like ChatGPT-4 and finds they struggle with nuanced, PhD-level inquiry. The models performed best on legal systems and social complexity but struggled with topics such as discrimination and social mobility.
Researchers found that humans still outperform ASR systems in noisy environments, but Whisper large-v3 matched human performance in all tested conditions except naturalistic pub noise. The system's ability to process acoustic properties and map them to the intended message was impressive.
A new study by Lancaster University reveals that language in job ads can unintentionally reinforce or disrupt labour force gender/racial composition. Workforces with more women tend to use family-friendly policies in ads, while racial minority workers' ads lack impact, the research shows.
A new study found that speech sounds have texture and shape to them, with R sounds associated with roughness and L sounds with smoothness. The research showed strong tendencies for matching trilled R sounds with jagged lines and L sounds with straight lines across cultures.
A recent study from the University of Surrey found that remote medical interpreting (RMI) can compromise the quality of communication in healthcare settings. Interpreters reported mixed experiences with technology during the COVID-19 pandemic, highlighting the need for careful consideration of interpreting methods based on the nature o...
A new study found that consonant lengthening is a universal trait in many languages, helping listeners identify word boundaries. The researchers analyzed data from the DoReCo corpus and found evidence of lengthening in 43 out of 51 languages.
Research by Lancaster University found that people in higher social grades, including corporate world and education sectors, are adopting each other's speech patterns to be more inclusive. This 'resonance' has increased over the past 20 years, particularly among those with high social status.
Researchers at Lancaster University find that the 'r' sound is becoming weaker in England, with most young speakers softening it to an elongation of the vowel. This decline is predicted to continue, potentially erasing traditional dialects and linguistic homogenization.
A new dictionary provides a comprehensive overview of Shakespeare's language, revealing words with reduced mental ability and plant hybrids. The Arden Encyclopedia of Shakespeare's Language offers insights into the linguistic thumbprints of plays and characters, as well as the networks of character interaction.
Researchers found that languages with larger populations of non-native speakers do not simplify their grammars. Instead, variations in complexity accumulate too slowly to adapt to the environment. This challenges long-held theories on the relationship between language complexity and social factors.
A study by University of Toronto researchers found that child language development and language evolution share a common cognitive foundation, based on a core knowledge base. The team built a computational model that predicts word meaning extension patterns across languages and time scales, highlighting the role of visual, associative,...
Researchers combined linguistics and genetics to propose a new origin theory for the Indo-European languages, suggesting an ultimate homeland south of the Caucasus. The study estimated the family to be approximately 8100 years old, with five main branches split off by around 7000 years ago.
A new machine learning model can automatically translate Akkadian text written in cuneiform into English, with the first version using Latin transliteration achieving satisfactory results. The program is effective for translating short sentences and can be used as part of a human-machine collaboration to correct and refine its output.
Scientists measured brain waves in participants and artificial intelligence systems to reveal similarities in how the brain interprets speech. The study provides a window into the operation of AI systems, which have been advancing rapidly but remain largely opaque.
Researchers at the Complexity Science Hub have developed an algorithm that can be applied to different languages and expand word lists significantly better than others. The new method, called LEXpander, outperforms previous algorithms in German and English, especially in sentiment analysis tasks.
A new machine-learning framework analyzes scene descriptions in movie scripts to recognize character actions, revealing widespread gender stereotypes. Female characters tend to show more emotion and less agency than males, while males are less likely to display emotional vulnerability.
Researchers from McGill University and MIT developed an AI system that can learn the rules and patterns of human languages on its own. The model automatically generates higher-level language patterns that can be applied to different languages, achieving better results.