๐ Exploring Large Language Models for Communication Games: An Empirical Study on Werewolf ๐
"Keeps LLMs frozen, and relies on retrieval and reflection on past communications and experiences for improvement, which is inspired by the human learning process of "reflection-in-action"." [gal30b+] ๐ค #CL
๐ Can NLP Models 'Identify', 'Distinguish', and 'Justify' Questions That Don't Have a Definitive Answer? ๐
"QnotA Dataset is constructed with the aim to test the ability of QA models in identifying, distinguishing and justifying questions without answers (QnotA)." [gal30b+] ๐ค #CL
๐ When Less Is More: Investigating Data Pruning for Pretraining LLMs at Scale ๐๐ง
"Performs a rigorous comparison at scale of the simple data quality estimator of perplexity, as well as more sophisticated and computationally intensive estimates of the Error L2-Norm and memorization." [gal30b+] ๐ค #CL #LG
๐ MoEController: Instruction-Based Arbitrary Image Manipulation with Mixture-of-Expert Controllers ๐ญ๐
"Leverages large language models (ChatGPT) and image synthesis models (ControlNet) to generate a large number of image-text pairs that can be used for global and local image manipulation datasets." [gal30b+] ๐ค #CV #CL
๐ Evaluation and Mitigation of Agnosia in Multimodal Large Language Models ๐ญ๐
"Proposes EMMA, an evaluation-mitigation framework that automatically creates fine-grained and diverse visual question answering examples to assess the extent of agnosia in Multimodal Pre-trained Language Models (MLLMs) comprehensively." [gal30b+] ๐ค #CV #CL
๐ Beyond Static Datasets: A Deep Interaction Approach to LLM Evaluation ๐๐พ
"Based on the deep interaction between large language models, which can help us evaluate large language models in real-world scenarios such as machine translation and code generation." [gal30b+] ๐ค #CL #AI
๐ Encoding Multi-Domain Scientific Papers by Ensembling Multiple CLS Tokens ๐๐ง
"Proposes Multi2SPE -- it encourages each of multiple CLS tokens to learn diverse ways of aggregating token embeddings, then sums them up together to create a single vector representation." [gal30b+] ๐ค #CL #DL #LG
๐ Fuzzy Fingerprinting Transformer Language-Models for Emotion Recognition in Conversations ๐๐พ
"We feed the utterances and their previous conversational turns to a pre-trained RoBERTa, obtaining contextual embedding utterance representations, that are then supplied to an adapted Fuzzy Fingerprint classification module." [gal30b+] ๐ค #CL #AI
๐ From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting ๐
"Fine-tunes GPT-4 on CNN DailyMail and generate entity-centric summaries by iteratively incorporating missing salient entities without increasing the length of the summary, using a chain of density prompt." [gal30b+] ๐ค #CL
๐ GLS-CSC: A Simple but Effective Strategy to Mitigate Chinese STM Models' Over-Reliance on Superficial Clue ๐
"Proposes a novel resampling training strategy called Gradually Learn Samples Containing Superficial Clue (GLS-CSC) to mitigate STM models' over-reliance on superficial clues." [gal30b+] ๐ค #CL
๐ RST-style Discourse Parsing Guided by Document-Level Content Structures ๐
"The proposed pipeline for RST-DP incorporates structure-aware news content sentence representations derived from the task of News Discourse Profiling via only a few additional layers in the neural network model architecture." [gal30b+] ๐ค #CL
๐ Unsupervised Multi-Document Summarization with Holistic Inference ๐
"Incorporates the holistic beam search inference method associated with the holistic measurements, named Subset Representative Index (SRI), which balances the importance and diversity of a subset of sentences from the source documents and can be calculated in unsupervised and adaptive manners." [gal30b+] ๐ค #CL
๐ ConDA: Contrastive Domain Adaptation for AI-generated Text Detection ๐๐พ๐ง
"Develops a contrastive domain adaptation framework, called ConDA, which learns domain-invariant feature representations via a contrastive loss in conjunction with standard domain adaptation techniques such as DANN and CDAN." [gal30b+] ๐ค #CL #AI #LG
โ๏ธ https://github.com/AmritaBh/ConDA-gen-text-detection
๐ https://arxiv.org/abs/2309.03992v1 #arxiv
๐ Multiple Representation Transfer From Large Language Models to End-to-End ASR Systems ๐๐
"Transferring multiple representations of large language models improves the end-to-end ASR performance by up to 15% relative CER compared to transferring only a single representation." [gal30b+] ๐ค #CL #SD
๐ A Function Interpretation Benchmark for Evaluating Interpretability Methods ๐๐พ๐ง
"We procedurally construct a suite of functions resembling components of trained neural networks, and use them to evaluate interpretability methods that use language models (LMs) to propose descriptions in language or code." [gal30b+] ๐ค #CL #AI #LG
โ๏ธ https://github.com/multimodal-interpretability/FIND
๐ https://arxiv.org/abs/2309.03886v1 #arxiv
๐ Implicit Design Choices and Their Impact on Emotion Recognition Model Development and Evaluation ๐ง ๐๐
"Emotion recognition from visual and audio data is accomplished by using a convolutional neural network (CNN) and a recurrent neural network (RNN), respectively, to encode the two modalities." [gal30b+] ๐ค #LG #CL #SD
๐ Zero-Shot Audio Captioning via Audibility Guidance ๐๐
"A caption is generated by combining a large pre-trained language model, such as GPT-2, with a multimodal matching model, which scores how well a text matches the input audio and a text classifier which provides the guidance for audibility." [gal30b+] ๐ค #SD #CL
๐ Introducing "Forecast Utterance" for Conversational Data Science ๐๐
"Introduces a new concept called a Forecast Utterance and then focus on automatically interpreting users' prediction goals from these Utterances by casting it as a slot-filling problem." [gal30b+] ๐ค #CL #HC
๐ The Daunting Dilemma with Sentence Encoders: Success on Standard Benchmarks, Failure in Capturing Basic Semantic Properties ๐
"Presents a study to explore the performance of popular sentence encoders on downstream tasks and their capability to capture basic semantic properties such as paraphrasing, synonym replacement, antonym replacement, and jumbling." [gal30b+] ๐ค #CL
๐ Word Segmentation Granularity in Korean ๐
"The process of splitting a sentence into words for further processing, for instance, part-of-speech tagging, chunking, dependency parsing, etc." [gal30b+] ๐ค #CL