Search

Multimodal sentiment analysis (MSA) systems leverage information from different modalities to predict human sentiment intensities. Incomplete modality is an important issue that may cause a significant performance drop in MSA systems. By generative imputation, i.e., recovering the missing data from available data, systems may achieve robust performance but will lead to high computational costs. This paper introduces a knowledge distillation method, called `Multi-Modal Contrastive Knowledge Distillation' (MM-CKD), to address the issue of incomplete modality in video sentiment analysis with lower computation cost, as a novel non-imputation-based method. We employ Multi-view Supervised Contrastive Learning (MVSC) to transfer knowledge from a teacher model to student models. This approach not only leverages cross-modal knowledge but also introduces cross-sample knowledge with supervision, jointly improving the performance of both teacher and student models through online learning. Our method gives competitive results with significantly lower computational costs than state-of-the-art imputation-based methods.

We discuss the stabilization of multiple moduli by utilizing Siegel modular forms in the framework of $Sp(2g,\mathbb{Z})$ modular invariant theories. We derive the stationary conditions at CP-conserving fixed points for a generic modular- and CP-invariant scalar potential. The stabilization of multiple moduli is explicitly demonstrated in $Sp(4,\mathbb{Z})$ and $Sp(6,\mathbb{Z})$ modular invariant scalar potentials. Furthermore, it turns out that there exists a light axion when the moduli are stabilized nearby a fixed point., Comment: 31 pages, 3 figures

In this study, we leveraged Channel State Information (CSI), commonly utilized in WLAN communication, as training data to develop and evaluate five distinct machine learning models for recognizing human postures: standing, sitting, and lying down. The models we employed were: (i) Linear Discriminant Analysis, (ii) Naive Bayes-Support Vector Machine, (iii) Kernel-Support Vector Machine, (iv) Random Forest, and (v) Deep Learning. We systematically analyzed how the accuracy of these models varied with different amounts of training data. Additionally, to assess their spatial generalization capabilities, we evaluated the models' performance in a setting distinct from the one used for data collection. The experimental findings indicated that while two models -- (ii) Naive Bayes-Support Vector Machine and (v) Deep Learning -- achieved 85% or more accuracy in the original setting, their accuracy dropped to approximately 30% when applied in a different environment. These results underscore that although CSI-based machine learning models can attain high accuracy within a consistent spatial structure, their performance diminishes considerably with changes in spatial conditions, highlighting a significant challenge in their generalization capabilities.

Maximum a posteriori decoding, a commonly used method for neural machine translation (NMT), aims to maximize the estimated posterior probability. However, high estimated probability does not always lead to high translation quality. Minimum Bayes Risk (MBR) decoding (\citealp{kumar2004minimum}) offers an alternative by seeking hypotheses with the highest expected utility. In this paper, we show that Quality Estimation (QE) reranking (\citealp{fernandes-etal-2022-quality}), which uses a QE model as a reranker, can be viewed as a variant of MBR. Inspired by this, we propose source-based MBR (sMBR) decoding, a novel approach that utilizes synthetic sources (generated via back-translation or paraphrasing) as ``support hypotheses'' and a reference-free quality estimation metric as the utility function, marking the first work to solely use sources in MBR decoding. Experiments show that sMBR outperforms QE reranking and the standard MBR decoding. Our findings suggest that sMBR is a promising approach for NMT decoding.

Speech emotion recognition (SER) has garnered increasing attention due to its wide range of applications in various fields, including human-machine interaction, virtual assistants, and mental health assistance. However, existing SER methods often overlook the information gap between the pre-training speech recognition task and the downstream SER task, resulting in sub-optimal performance. Moreover, current methods require much time for fine-tuning on each specific speech dataset, such as IEMOCAP, which limits their effectiveness in real-world scenarios with large-scale noisy data. To address these issues, we propose an active learning (AL)-based fine-tuning framework for SER, called \textsc{After}, that leverages task adaptation pre-training (TAPT) and AL methods to enhance performance and efficiency. Specifically, we first use TAPT to minimize the information gap between the pre-training speech recognition task and the downstream speech emotion recognition task. Then, AL methods are employed to iteratively select a subset of the most informative and diverse samples for fine-tuning, thereby reducing time consumption. Experiments demonstrate that our proposed method \textsc{After}, using only 20\% of samples, improves accuracy by 8.45\% and reduces time consumption by 79\%. The additional extension of \textsc{After} and ablation studies further confirm its effectiveness and applicability to various real-world scenarios. Our source code is available on Github for reproducibility. (https://github.com/Clearloveyuan/AFTER)., Comment: Accepted by Journal of Natural Language Processing. arXiv admin note: text overlap with arXiv:2310.00283

Dataset distillation aims to compress a training dataset by creating a small number of informative synthetic samples such that neural networks trained on them perform as well as those trained on the original training dataset. Current text dataset distillation methods create each synthetic sample as a sequence of word embeddings instead of a text to apply gradient-based optimization; however, such embedding-level distilled datasets cannot be used for training other models whose word embedding weights are different from the model used for distillation. To address this issue, we propose a novel text dataset distillation approach, called Distilling dataset into Language Model (DiLM), which trains a language model to generate informative synthetic training samples as text data, instead of directly optimizing synthetic samples. We evaluated DiLM on various text classification datasets and showed that distilled synthetic datasets from DiLM outperform those from current coreset selection methods. DiLM achieved remarkable generalization performance in training different types of models and in-context learning of large language models. Our code will be available at https://github.com/arumaekawa/DiLM., Comment: Accepted by Findings of NAACL 2024

We introduce height-bounded LZ encodings (LZHB), a new family of compressed representations that are variants of Lempel-Ziv parsings with a focus on bounding the worst-case access time to arbitrary positions in the text directly via the compressed representation. An LZ-like encoding is a partitioning of the string into phrases of length $1$ which can be encoded literally, or phrases of length at least $2$ which have a previous occurrence in the string and can be encoded by its position and length. An LZ-like encoding induces an implicit referencing forest on the set of positions of the string. An LZHB encoding is an LZ-like encoding where the height of the implicit referencing forest is bounded. An LZHB encoding with height constraint $h$ allows access to an arbitrary position of the underlying text using $O(h)$ predecessor queries. While computing the smallest LZHB encoding efficiently seems to be difficult [Cicalese \& Ugazio 2024, arxiv], we give the first linear time algorithm for strings over a constant size alphabet that computes the greedy LZHB encoding, i.e., the string is processed from beginning to end, and the longest prefix of the remaining string that can satisfy the height constraint is taken as the next phrase. Our algorithms significantly improve both theoretically and practically, the very recently and independently proposed algorithms by Lipt\'ak et al. (arxiv, to appear at CPM 2024). We also analyze the size of height bounded LZ encodings in the context of repetitiveness measures, and show for some constant $c$, the size $z_{HB}$ of the optimal LZHB encoding with height bound $c\log n$ is $O(g_{rl})$, where $g_{rl}$ is the size of the smallest run-length grammar. We also show $z_{HB} = o(g_{rl})$ for some family of strings, making $z_{HB}$ one of the smallest known repetitiveness measures for which $O({\sf polylog} n)$ time access is possible using linear space., Comment: abstract shortened to fit arxiv requirements

We investigate the compression sensitivity [Akagi et al., 2023] of lex-parse [Navarro et al., 2021] for two operations: (1) single character edit and (2) modification of the alphabet ordering, and give tight upper and lower bounds for both operations. For both lower bounds, we use the family of Fibonacci words. For the bounds on edit operations, our analysis makes heavy use of properties of the Lyndon factorization of Fibonacci words to characterize the structure of lex-parse.

A string $w$ is said to be a minimal absent word (MAW) for a string $S$ if $w$ does not occur in $S$ and any proper substring of $w$ occurs in $S$. We focus on non-trivial MAWs which are of length at least 2. Finding such non-trivial MAWs for a given string is motivated for applications in bioinformatics and data compression. Fujishige et al. [TCS 2023] proposed a data structure of size $\Theta(n)$ that can output the set $\mathsf{MAW}(S)$ of all MAWs for a given string $S$ of length $n$ in $O(n + |\mathsf{MAW}(S)|)$ time, based on the directed acyclic word graph (DAWG). In this paper, we present a more space efficient data structure based on the compact DAWG (CDAWG), which can output $\mathsf{MAW}(S)$ in $O(|\mathsf{MAW}(S)|)$ time with $O(\mathsf{e}_\min)$ space, where $\mathsf{e}_\min$ denotes the minimum of the sizes of the CDAWGs for $S$ and for its reversal $S^R$. For any strings of length $n$, it holds that $\mathsf{e}_\min < 2n$, and for highly repetitive strings $\mathsf{e}_\min$ can be sublinear (up to logarithmic) in $n$. We also show that MAWs and their generalization minimal rare words have close relationships with extended bispecial factors, via the CDAWG., Comment: Accepted for IWOCA 2024

Studying the nature of spiral arms is essential for understanding the formation of the intricate disc structure of the Milky Way. The European Space Agency's Gaia mission has provided revolutionary observational data that have uncovered detailed kinematical features of stars in the Milky Way. However, so far the nature of spiral arms continues to remain a mystery. Here we present that the stellar kinematics traced by the classical Cepheids around the Perseus and Outer spiral arms in the Milky Way shows strikingly different kinematical properties from each other: the radial and azimuthal velocities of Cepheids with respect to the Galactic centre show positive and negative correlations in the Perseus and Outer arms, respectively. We also found that the dynamic spiral arms commonly seen in an N-body/hydrodynamics simulation of a Milky Way-like galaxy can naturally explain the observed kinematic trends. Furthermore, a comparison with such a simulation suggests that the Perseus arm is being disrupted while the Outer arm is growing. Our findings suggest that two neighbouring spiral arms in distinct evolutionary phases - growing and disrupting phases - coexist in the Milky Way., Comment: 10 pages, 10 figures; accepted for publication in MNRAS

Multimodal emotion recognition aims to recognize emotions for each utterance of multiple modalities, which has received increasing attention for its application in human-machine interaction. Current graph-based methods fail to simultaneously depict global contextual features and local diverse uni-modal features in a dialogue. Furthermore, with the number of graph layers increasing, they easily fall into over-smoothing. In this paper, we propose a method for joint modality fusion and graph contrastive learning for multimodal emotion recognition (Joyful), where multimodality fusion, contrastive learning, and emotion recognition are jointly optimized. Specifically, we first design a new multimodal fusion mechanism that can provide deep interaction and fusion between the global contextual and uni-modal specific features. Then, we introduce a graph contrastive learning framework with inter-view and intra-view contrastive losses to learn more distinguishable representations for samples with different sentiments. Extensive experiments on three benchmark datasets indicate that Joyful achieved state-of-the-art (SOTA) performance compared to all baselines.

Speech emotion recognition (SER) has drawn increasing attention for its applications in human-machine interaction. However, existing SER methods ignore the information gap between the pre-training speech recognition task and the downstream SER task, leading to sub-optimal performance. Moreover, they require much time to fine-tune on each specific speech dataset, restricting their effectiveness in real-world scenes with large-scale noisy data. To address these issues, we propose an active learning (AL) based Fine-Tuning framework for SER that leverages task adaptation pre-training (TAPT) and AL methods to enhance performance and efficiency. Specifically, we first use TAPT to minimize the information gap between the pre-training and the downstream task. Then, AL methods are used to iteratively select a subset of the most informative and diverse samples for fine-tuning, reducing time consumption. Experiments demonstrate that using only 20\%pt. samples improves 8.45\%pt. accuracy and reduces 79\%pt. time consumption.

Conventional automatic evaluation metrics, such as BLEU and ROUGE, developed for natural language generation (NLG) tasks, are based on measuring the n-gram overlap between the generated and reference text. These simple metrics may be insufficient for more complex tasks, such as question generation (QG), which requires generating questions that are answerable by the reference answers. Developing a more sophisticated automatic evaluation metric, thus, remains an urgent problem in QG research. This work proposes PMAN (Prompting-based Metric on ANswerability), a novel automatic evaluation metric to assess whether the generated questions are answerable by the reference answers for the QG tasks. Extensive experiments demonstrate that its evaluation results are reliable and align with human evaluations. We further apply our metric to evaluate the performance of QG models, which shows that our metric complements conventional metrics. Our implementation of a GPT-based QG model achieves state-of-the-art performance in generating answerable questions.

We present highly sampled photometry of the supernova (SN) 2023ixf, a Type II SN in M101, beginning 2 days before its first known detection. To gather these data, we enlisted the global Unistellar Network of citizen scientists. These 252 observations from 115 telescopes show the SN's rising brightness associated with shock emergence followed by gradual decay. We measure a peak $M_{V}$ = -18.18 $\pm$ 0.09 mag at 2023-05-25 21:37 UTC in agreement with previously published analyses., Comment: 4 pages, 1 figure

Since April 2020, the SuperKEKB has been operating with the crab waist scheme. The luminosity record achieved in June 2022 was $4.71 \times 10^{34} \text{ cm}^{-2}\text{s}^{-1}$, which overtook its predecessor KEKB by more than a factor of 2. The beam-beam interaction plays a key role in causing vertical blowup and consequently limiting the luminosity performance of SuperKEKB. In this paper, we examine luminosity tunings under the influence of beam-beam effects and review the luminosity performance of SuperKEKB with the crab-waist operation from 2020 to 2022., Comment: 14 pages, 6 figures

The beam-beam interaction is one of the most critical factors determining the luminosity performance of colliders. As a circular collider utilizing the crab-waist scheme, multiple factors, such as beam-beam, crab waist, impedances, etc., interact to determine the luminosity of SuperKEKB. The interplay of these factors makes it challenging to predict luminosity via simulations. This paper presents recent advances in understanding the luminosity performance of SuperKEKB from beam-beam simulations and experiments. The key aspects affecting the luminosity of SuperKEKB, as well as the areas where further research is needed, are highlighted., Comment: 16 pages, 19 figures

LZ-End is a variant of the well-known Lempel-Ziv parsing family such that each phrase of the parsing has a previous occurrence, with the additional constraint that the previous occurrence must end at the end of a previous phrase. LZ-End was initially proposed as a greedy parsing, where each phrase is determined greedily from left to right, as the longest factor that satisfies the above constraint~[Kreft & Navarro, 2010]. In this work, we consider an optimal LZ-End parsing that has the minimum number of phrases in such parsings. We show that a decision version of computing the optimal LZ-End parsing is NP-complete by showing a reduction from the vertex cover problem. Moreover, we give a MAX-SAT formulation for the optimal LZ-End parsing adapting an approach for computing various NP-hard repetitiveness measures recently presented by [Bannai et al., 2022]. We also consider the approximation ratio of the size of greedy LZ-End parsing to the size of the optimal LZ-End parsing, and give a lower bound of the ratio which asymptotically approaches $2$.

The suffix trees are fundamental data structures for various kinds of string processing. The suffix tree of a text string $T$ of length $n$ has $O(n)$ nodes and edges, and the string label of each edge is encoded by a pair of positions in $T$. Thus, even after the tree is built, the input string $T$ needs to be kept stored and random access to $T$ is still needed. The \emph{linear-size suffix tries} (\emph{LSTs}), proposed by Crochemore et al. [Linear-size suffix tries, TCS 638:171-178, 2016], are a "stand-alone" alternative to the suffix trees. Namely, the LST of an input text string $T$ of length $n$ occupies $O(n)$ total space, and supports pattern matching and other tasks with the same efficiency as the suffix tree without the need to store the input text string $T$. Crochemore et al. proposed an \emph{offline} algorithm which transforms the suffix tree of $T$ into the LST of $T$ in $O(n \log \sigma)$ time and $O(n)$ space, where $\sigma$ is the alphabet size. In this paper, we present two types of \emph{online} algorithms which "directly" construct the LST, from right to left, and from left to right, without constructing the suffix tree as an intermediate structure. Both algorithms construct the LST incrementally when a new symbol is read, and do not access the previously read symbols. Both of the right-to-left construction algorithm and the left-to-right construction algorithm work in $O(n \log \sigma)$ time and $O(n)$ space. The main feature of our algorithms is that the input text string does not need to be stored., Comment: 27 pages, 14 figures. arXiv admin note: text overlap with arXiv:1901.10045