Di Jiang

The empirical variance estimation and its corresponding top- k query is a fundamental problem in the data mining and data analytics and serves as an inherent building block for many clustering and feature selection algorithms. Since the exact computation requires scanning the whole dataset which will be prohibitively expensive for many real-time applications, all existing studies in the literature are dedicated to find the approximate solutions by using the sampling techniques. For the top- k query processing, we observe that all existing studies analyze the error of the estimated variance of each selected attribute independently by using the traditional centrality inequality (e.g., Chernoff bounds/Hoeffding’s inequality) and then adopt the traditional union bound to estimate the aggregate error of the k selected attributes. As such, the bound is significantly loose and renders their algorithm sensitive to the parameter k . Motivated by this, in this paper, we propose a once-for-all progressive sampling algorithm, namely Top- k E mpirical V ariance Computation with Rademacher A verage (EVA) , for jointly analyzing the aggregate error of the variances of all attributes one-for-all. In our algorithm, the tuples are sampled iteratively in batches. In each iteration, we estimate the accuracy currently achieved with the tuples already being sampled and derive the error bound by adopting a key concept called Rademacher Average from the statistical machine learning theory. Our error estimation algorithm enjoys two features. Firstly, it is data-dependent which fully makes use of the tuples already sampled and help us terminate the algorithm earlier once our desired accuracy is achieved. Secondly and more importantly, it estimates the aggregate error of the selected attributes simultaneously once-for-all which considers their inter-relation and as such, it is tighter than the traditional union bound-based method. Our empirical study shows that our algorithm outperforms the state-of-the-art algorithms by orders of magnitudes in terms of the efficiency with the same accuracy guarantee.

Latent Dirichlet Allocation (LDA) is a prominent generative probabilistic model used for uncovering abstract topics within document collections. In this paper, we explore the effectiveness of augmenting topic models with Large Language Models (LLMs) through integration into two key phases: Initialization and Post-Correction. Since the LDA is highly dependent on the quality of its initialization, we conduct extensive experiments on the LLM-guided topic clustering for initializing the Gibbs sampling algorithm. Interestingly, the experimental results reveal that while the proposed initialization strategy improves the early iterations of LDA, it has no effect on the convergence and yields the worst performance compared to the baselines. The LLM-enabled post-correction, on the other hand, achieved a promising improvement of 5.86% in the coherence evaluation. These results highlight the practical benefits of the LLM-in-the-loop approach and challenge the belief that LLMs are always the superior text mining alternative.

This technical report introduces innovative optimizations for Kaldi-based Automatic Speech Recognition (ASR) systems, focusing on acoustic model enhancement, hyperparameter tuning, and language model efficiency. We developed a custom Conformer block integrated with a multistream TDNN-F structure, enabling superior feature extraction and temporal modeling. Our approach includes advanced data augmentation techniques and dynamic hyperparameter optimization to boost performance and reduce overfitting. Additionally, we propose robust strategies for language model management, employing Bayesian optimization and $n$-gram pruning to ensure relevance and computational efficiency. These systematic improvements significantly elevate ASR accuracy and robustness, outperforming existing methods and offering a scalable solution for diverse speech recognition scenarios. This report underscores the importance of strategic optimizations in maintaining Kaldi's adaptability and competitiveness in rapidly evolving technological landscapes.

The single-source personalized PageRank (SSPPR) is a fundamental problem in the social works which find a lot of applications in information retrieval, recommender system, etc. Given a source node $s$ in a networks, it returns an importance score called PPR value of each node in the network w.r.t. the source node $s$. Despite that SSPPR is extensively studied in the traditional offline social networks, no research effort has been devoted to the problem under the emerging online social networks (OSN) where the network is online given and the access to the network is restricted. In an OSN, the network topology is not known apriori and the users are not allowed to randomly access the vertices and edges. Instead, the network provider only allows the users to explore the network from a given source vertex by using an API step by step in an crawling fashion. As such, all existing algorithms for SSPPR which requires the full knowledge of the global or local topology can not be applied to OSN. We observe that the only possible method for SSPPR on OSN is the Monte Carlo (MC) simulation of random walks. But the traditional MC algorithm employs the static sampling for MC simulation which calculate the total number of random walks needed and obtain all of them in one lump. But this static sampling renders them suffer from the lack of interactiveness and flexibility which forces the users to wait until the end of algorithm without any feedback/indicator regarding the progress and thus, largely degrades the user experience a lot. Besides, they always overestimate the total number of random walks needed due to the nature of the worse-case analysis which is sample-independent. Motivated by this, in this paper, we propose a progressive sampling algorithm, namely Single-Source Personalized PageRank on Online Social Networks with Rademacher Average (PANDA), to boost the processing of SSPPR on the Online Social Networks. In our algorithm, the random walks are sampled iteratively in batches. In each iteration, we estimate the accuracy currently achieved with the random walks already being sampled and derive the error bound by adopting a key concept called Rademacher Average from the statistical machine learning theory. As such, our proposed algorithm is data-dependent which fully make sense of the random walks already sampled and help us terminate the algorithm earlier once our desired accuracy is achieved. We also develop a dynamic sampling schedule which in each iteration derives the sample size for the next iteration. Our empirical study shows that our algorithm outperforms the state-of-the-art algorithms by orders of magnitudes in terms of the efficiency with the same accuracy guarantee.

Mengze Hong, Chen Jason Zhang, Chaotao Chen, Rongzhong Lian, Di Jiang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track). 2025.

Automatic speech Recognition (ASR) is a fundamental and important task in the field of speech and natural language processing. It is an inherent building block in many applications such as voice assistant, speech translation, etc. Despite the advancement of ASR technologies in recent years, it is still inevitable for modern ASR systems to have a substantial number of erroneous recognition due to environmental noise, ambiguity, etc. Therefore, the error correction in ASR is crucial. Motivated by this, this paper studies ASR error correction in the Chinese language, which is one of the most popular languages and enjoys a large number of users in the world. We first create a benchmark dataset named \emph{ASR-EC} that contains a wide spectrum of ASR errors generated by industry-grade ASR systems. To the best of our knowledge, it is the first Chinese ASR error correction benchmark. Then, inspired by the recent advances in \emph{large language models (LLMs)}, we investigate how to harness the power of LLMs to correct ASR errors. We apply LLMs to ASR error correction in three paradigms. The first paradigm is prompting, which is further categorized as zero-shot, few-shot, and multi-step. The second paradigm is finetuning, which finetunes LLMs with ASR error correction data. The third paradigm is multi-modal augmentation, which collectively utilizes the audio and ASR transcripts for error correction. Extensive experiments reveal that prompting is not effective for ASR error correction. Finetuning is effective only for a portion of LLMs. Multi-modal augmentation is the most effective method for error correction and achieves state-of-the-art performance.

Recurrent neural network transducer (RNN-T) is an important branch of current end-to-end automatic speech recognition (ASR). Various promising approaches have been designed for boosting RNN-T architecture; however, few studies exploit the effectiveness of pretrained methods in this framework. In this paper, we introduce the pretrained acoustic extractor (PAE) and the pretrained linguistic network (PLN) to enhance the Conformer long short-term memory (Conformer-LSTM) transducer. First, we construct the input of the acoustic encoder with two different latent representations: one extracted by PAE from the raw waveform, and the other obtained from filter-bank transformation. Second, we fuse an extra semantic feature from the PLN into the joint network to reduce illogical and homophonic errors. Compared with previous works, our approaches are able to obtain pretrained representations for better model generalization. Evaluation on two large-scale datasets has demonstrated that our proposed approaches yield better performance than existing approaches.

Due to the rising awareness of privacy protection and the voluminous scale of speech data, it is becoming infeasible for Automatic Speech Recognition (ASR) system developers to train the acoustic model with complete data as before. For example, the data may be owned by different curators, and it is not allowed to share with others. In this paper, we propose a novel paradigm to solve salient problems plaguing the ASR field. In the first stage, multiple acoustic models are trained based upon different subsets of the complete speech data, while in the second phase, two novel algorithms are utilized to generate a high-quality acoustic model based upon those trained on data subsets. We first propose the Genetic Merge Algorithm (GMA), which is a highly specialized algorithm for optimizing acoustic models but suffers from low efficiency. We further propose the SGD-Based Optimizational Merge Algorithm (SOMA), which effectively alleviates the efficiency bottleneck of GMA and maintains superior model accuracy. Extensive experiments on public data show that the proposed methods can significantly outperform the state-of-the-art. Furthermore, we introduce Shapley Value to estimate the contribution score of the trained models, which is useful for evaluating the effectiveness of the data and providing fair incentives to their curators.