分享

Large language models (LLMs) can suggest missing elements from items listed in a prompt, which can be used for list completion or recommendations based on users' history. However, their performance degrades when presented with too many items, as they start to suggest items already included in the input list. This occurs at around 100 items for mid-2024 flagship LLMs. We evaluate this phenomenon on both synthetic problems (e.g., finding missing numbers in a given range of shuffled integers) and realistic movie recommendation scenarios. We refer to this issue as \textit{attention overflow}, as preventing repetition requires attending to all items simultaneously. Although iterative loops can mitigate this problem, their costs increase with the repetition rate, affecting the language models' ability to derive novelty from lengthy inputs.

Error correction (EC) based on large language models is an emerging technology to enhance the performance of automatic speech recognition (ASR) systems. Generally, training data for EC are collected by automatically pairing a large set of ASR hypotheses (as sources) and their gold references (as targets). However, the quality of such pairs is not guaranteed, and we observed various types of noise which can make the EC models brittle, e.g. inducing overcorrection in out-of-domain (OOD) settings. In this work, we propose two fundamental criteria that EC training data should satisfy: namely, EC targets should (1) improve linguistic acceptability over sources and (2) be inferable from the available context (e.g. source phonemes). Through these criteria, we identify low-quality EC pairs and train the models not to make any correction in such cases, the process we refer to as conservative data filtering. In our experiments, we focus on Japanese ASR using a strong Conformer-CTC as the baseline and finetune Japanese LLMs for EC. Through our evaluation on a suite of 21 internal benchmarks, we demonstrate that our approach can significantly reduce overcorrection and improve both the accuracy and quality of ASR results in the challenging OOD settings.

The mainstream automatic speech recognition (ASR) technology usually requires hundreds to thousands of hours of annotated speech data. Three approaches to low-resourced ASR are phoneme or subword based supervised pre-training, and self-supervised pre-training over multilingual data. The Iu Mien language is the main ethnic language of the Yao ethnic group in China and is low-resourced in the sense that the annotated speech is very limited. With less than 10 hours of transcribed Iu Mien language, this paper investigates and compares the three approaches for Iu Mien speech recognition. Our experiments are based on the recently released, three backbone models pretrained over the 10 languages from the CommonVoice dataset (CV-Lang10), which correspond to the three approaches for low-resourced ASR. It is found that phoneme supervision can achieve better results compared to subword supervision and self-supervision, thereby providing higher data-efficiency. Particularly, the Whistle models, i.e., obtained by the weakly-supervised phoneme-based multilingual pre-training, obtain the most competitive results.

Although engagement in self-directed wellness exercises typically declines over time, integrating social support such as coaching can sustain it. However, traditional forms of support are often inaccessible due to the high costs and complex coordination. Large Language Models (LLMs) show promise in providing human-like dialogues that could emulate social support. Yet, in-depth, in situ investigations of LLMs to support behavior change remain underexplored. We conducted two randomized experiments to assess the impact of LLM agents on user engagement with mindfulness exercises. First, a single-session study, involved 502 crowdworkers; second, a three-week study, included 54 participants. We explored two types of LLM agents: one providing information and another facilitating self-reflection. Both agents enhanced users' intentions to practice mindfulness. However, only the information-providing LLM, featuring a friendly persona, significantly improved engagement with the exercises. Our findings suggest that specific LLM agents may bridge the social support gap in digital health interventions.

Language models are transforming the ways that their users engage with the world. Despite impressive capabilities, over-consumption of language model outputs risks propagating unchecked errors in the short-term and damaging human capabilities for critical thinking in the long-term, particularly in knowledge-based tasks. How can we develop scaffolding around language models to curate more appropriate use? We propose selective frictions for language model experiences, inspired by behavioral science interventions, to dampen misuse. Frictions involve small modifications to a user's experience, e.g., the addition of a button impeding model access and reminding a user of their expertise relative to the model. Through a user study with real humans, we observe shifts in user behavior from the imposition of a friction over LLMs in the context of a multi-topic question-answering task as a representative task that people may use LLMs for, e.g., in education and information retrieval. We find that frictions modulate over-reliance by driving down users' click rates while minimally affecting accuracy for those topics. Yet, frictions may have unintended effects. We find marked differences in users' click behaviors even on topics where frictions were not provisioned. Our contributions motivate further study of human-AI behavioral interaction to inform more effective and appropriate LLM use.

Online courses have significantly lowered the barrier to accessing education, yet the varying content quality of these videos poses challenges. In this work, we focus on the task of automatically evaluating the quality of video course content. We have constructed a dataset with a substantial collection of video courses and teaching materials. We propose three evaluation principles and design a new evaluation framework, \textit{VCEval}, based on these principles. The task is modeled as a multiple-choice question-answering task, with a language model serving as the evaluator. Our method effectively distinguishes video courses of different content quality and produces a range of interpretable results.

In human-computer interaction (HCI), Speech Emotion Recognition (SER) is a key technology for understanding human intentions and emotions. Traditional SER methods struggle to effectively capture the long-term temporal correla-tions and dynamic variations in complex emotional expressions. To overcome these limitations, we introduce the PCQ method, a pioneering approach for SER via \textbf{P}rogressive \textbf{C}hannel \textbf{Q}uerying. This method can drill down layer by layer in the channel dimension through the channel query technique to achieve dynamic modeling of long-term contextual information of emotions. This mul-ti-level analysis gives the PCQ method an edge in capturing the nuances of hu-man emotions. Experimental results show that our model improves the weighted average (WA) accuracy by 3.98\% and 3.45\% and the unweighted av-erage (UA) accuracy by 5.67\% and 5.83\% on the IEMOCAP and EMODB emotion recognition datasets, respectively, significantly exceeding the baseline levels.

Geological borehole descriptions contain detailed textual information about the composition of the subsurface. However, their unstructured format presents significant challenges for extracting relevant features into a structured format. This paper introduces GEOBERTje: a domain adapted large language model trained on geological borehole descriptions from Flanders (Belgium) in the Dutch language. This model effectively extracts relevant information from the borehole descriptions and represents it into a numeric vector space. Showcasing just one potential application of GEOBERTje, we finetune a classifier model on a limited number of manually labeled observations. This classifier categorizes borehole descriptions into a main, second and third lithology class. We show that our classifier outperforms both a rule-based approach and GPT-4 of OpenAI. This study exemplifies how domain adapted large language models enhance the efficiency and accuracy of extracting information from complex, unstructured geological descriptions. This offers new opportunities for geological analysis and modeling using vast amounts of data.

This paper describes the submissions of team TalTech-IRIT-LIS to the DISPLACE 2024 challenge. Our team participated in the speaker diarization and language diarization tracks of the challenge. In the speaker diarization track, our best submission was an ensemble of systems based on the pyannote.audio speaker diarization pipeline utilizing powerset training and our recently proposed PixIT method that performs joint diarization and speech separation. We improve upon PixIT by using the separation outputs for speaker embedding extraction. Our ensemble achieved a diarization error rate of 27.1% on the evaluation dataset. In the language diarization track, we fine-tuned a pre-trained Wav2Vec2-BERT language embedding model on in-domain data, and clustered short segments using AHC and VBx, based on similarity scores from LDA/PLDA. This led to a language diarization error rate of 27.6% on the evaluation data. Both results were ranked first in their respective challenge tracks.

In this paper, we provide a large audio-visual speaker recognition dataset, VoxBlink2, which includes approximately 10M utterances with videos from 110K+ speakers in the wild. This dataset represents a significant expansion over the VoxBlink dataset, encompassing a broader diversity of speakers and scenarios by the grace of an optimized data collection pipeline. Afterward, we explore the impact of training strategies, data scale, and model complexity on speaker verification and finally establish a new single-model state-of-the-art EER at 0.170% and minDCF at 0.006% on the VoxCeleb1-O test set. Such remarkable results motivate us to explore speaker recognition from a new challenging perspective. We raise the Open-Set Speaker-Identification task, which is designed to either match a probe utterance with a known gallery speaker or categorize it as an unknown query. Associated with this task, we design concrete benchmark and evaluation protocols. The data and model resources can be found in http://voxblink2.github.io.

This report describes the submission of HYU ASML team to the IEEE Signal Processing Cup 2024 (SP Cup 2024). This challenge, titled "ROBOVOX: Far-Field Speaker Recognition by a Mobile Robot," focuses on speaker recognition using a mobile robot in noisy and reverberant conditions. Our solution combines the result of deep residual neural networks and time-delay neural network-based speaker embedding models. These models were trained on a diverse dataset that includes French speech. To account for the challenging evaluation environment characterized by high noise, reverberation, and short speech conditions, we focused on data augmentation and training speech duration for the speaker embedding model. Our submission achieved second place on the SP Cup 2024 public leaderboard, with a detection cost function of 0.5245 and an equal error rate of 6.46%.

Sparse Mixture of Expert (SMoE) models have emerged as a scalable alternative to dense models in language modeling. These models use conditionally activated feedforward subnetworks in transformer blocks, allowing for a separation between total model parameters and per-example computation. However, large token-routed SMoE models face a significant challenge: during inference, the entire model must be used for a sequence or a batch, resulting in high latencies in a distributed setting that offsets the advantages of per-token sparse activation. Our research explores task-specific model pruning to inform decisions about designing SMoE architectures, mainly modulating the choice of expert counts in pretraining. We investigate whether such pruned models offer advantages over smaller SMoE models trained from scratch, when evaluating and comparing them individually on tasks. To that end, we introduce an adaptive task-aware pruning technique UNCURL to reduce the number of experts per MoE layer in an offline manner post-training. Our findings reveal a threshold pruning factor for the reduction that depends on the number of experts used in pretraining, above which, the reduction starts to degrade model performance. These insights contribute to our understanding of model design choices when pretraining with SMoE architectures, particularly useful when considering task-specific inference optimization for later stages.