Most graph-to-text works are built on the encoder-decoder framework with cross-attention mechanism. Recent studies have shown that explicitly modeling the input graph structure can significantly improve the performance. However, the vanilla structural encoder cannot capture all specialized information in a single forward pass for all decoding steps, resulting in inaccurate semantic representations. Meanwhile, the input graph is flatted as an unordered sequence in the cross attention, ignoring the original graph structure. As a result, the obtained input graph context vector in the decoder may be flawed. To address these issues, we propose a Structure-Aware Cross-Attention (SACA) mechanism to re-encode the input graph representation conditioning on the newly generated context at each decoding step in a structure aware manner. We further adapt SACA and introduce its variant Dynamic Graph Pruning (DGP) mechanism to dynamically drop irrelevant nodes in the decoding process. We achieve new state-of-the-art results on two graph-to-text datasets, LDC2020T02 and ENT-DESC, with only minor increase on computational cost.
The field of natural language processing (NLP) has made significant strides in recent years, particularly in the development of large-scale vision-language models (VLMs). These models aim to bridge the gap between text and visual information, enabling a more comprehensive understanding of multimedia data. However, as these models become larger and more complex, they also become more challenging to train and deploy. One approach to addressing this challenge is the use of sparsely-gated mixture-of-experts (MoE) techniques, which divide the model into smaller, specialized sub-models that can jointly solve a task. In this paper, we explore the effectiveness of MoE in scaling vision-language models, demonstrating its potential to achieve state-of-the-art performance on a range of benchmarks over dense models of equivalent computational cost. Our research offers valuable insights into stabilizing the training of MoE models, understanding the impact of MoE on model interpretability, and balancing the trade-offs between compute performance when scaling VLMs. We hope our work will inspire further research into the use of MoE for scaling large-scale vision-language models and other multimodal machine learning applications.
Symbolic regression (SR) is a challenging task in machine learning that involves finding a mathematical expression for a function based on its values. Recent advancements in SR have demonstrated the efficacy of pretrained transformer-based models for generating equations as sequences, which benefit from large-scale pretraining on synthetic datasets and offer considerable advantages over GP-based methods in terms of inference time. However, these models focus on supervised pretraining goals borrowed from text generation and ignore equation-specific objectives like accuracy and complexity. To address this, we propose TPSR, a Transformer-based Planning strategy for Symbolic Regression that incorporates Monte Carlo Tree Search into the transformer decoding process. TPSR, as opposed to conventional decoding strategies, allows for the integration of non-differentiable feedback, such as fitting accuracy and complexity, as external sources of knowledge into the equation generation process. Extensive experiments on various datasets show that our approach outperforms state-of-the-art methods, enhancing the model's fitting-complexity trade-off, extrapolation abilities, and robustness to noise. We also demonstrate that the utilization of various caching mechanisms can further enhance the efficiency of TPSR.
In this paper, we focus on video-to-text summarization and investigate how to best utilize multimodal information for summarizing long inputs (e.g., an hour-long TV show) into long outputs (e.g., a multi-sentence summary). We extend SummScreen (Chen et al., 2021), a dialogue summarization dataset consisting of transcripts of TV episodes with reference summaries, and create a multimodal variant by collecting corresponding full-length videos. We incorporate multimodal information into a pre-trained textual summarizer efficiently using adapter modules augmented with a hierarchical structure while tuning only 3.8\% of model parameters. Our experiments demonstrate that multimodal information offers superior performance over more memory-heavy and fully fine-tuned textual summarization methods.
When extracting structured data from repetitively organized documents, such as dictionaries, directories, or even newspapers, a key challenge is to correctly segment what constitutes the basic text regions for the target database. Traditionally, such a problem was tackled as part of the layout analysis and was mostly based on visual clues for dividing (top-down) approaches. Some agglomerating (bottom-up) approaches started to consider textual information to link similar contents, but they required a proper over-segmentation of fine-grained units. In this work, we propose a new pragmatic approach whose efficiency is demonstrated on 19th century French Trade Directories. We propose to consider two sub-problems: coarse layout detection (text columns and reading order), which is assumed to be effective and not detailed here, and a fine-grained entry separation stage for which we propose to adapt a state-of-the-art Named Entity Recognition (NER) approach. By injecting special visual tokens, coding, for instance, indentation or breaks, into the token stream of the language model used for NER purpose, we can leverage both textual and visual knowledge simultaneously. Code, data, results and models are available at https://github.com/soduco/paper-entryseg-icdar23-code, https://huggingface.co/HueyNemud/ (icdar23-entrydetector* variants)
Benefiting from large-scale vision-language pre-training on image-text pairs, open-world detection methods have shown superior generalization ability under the zero-shot or few-shot detection settings. However, a pre-defined category space is still required during the inference stage of existing methods and only the objects belonging to that space will be predicted. To introduce a "real" open-world detector, in this paper, we propose a novel method named CapDet to either predict under a given category list or directly generate the category of predicted bounding boxes. Specifically, we unify the open-world detection and dense caption tasks into a single yet effective framework by introducing an additional dense captioning head to generate the region-grounded captions. Besides, adding the captioning task will in turn benefit the generalization of detection performance since the captioning dataset covers more concepts. Experiment results show that by unifying the dense caption task, our CapDet has obtained significant performance improvements (e.g., +2.1% mAP on LVIS rare classes) over the baseline method on LVIS (1203 classes). Besides, our CapDet also achieves state-of-the-art performance on dense captioning tasks, e.g., 15.44% mAP on VG V1.2 and 13.98% on the VG-COCO dataset.
It is widely accepted that so-called facts can be checked by searching for information on the Internet. This process requires a fact-checker to formulate a search query based on the fact and to present it to a search engine. Then, relevant and believable passages need to be identified in the search results before a decision is made. This process is carried out by sub-editors at many news and media organisations on a daily basis. Here, we ask the question as to whether it is possible to automate the first step, that of query generation. Can we automatically formulate search queries based on factual statements which are similar to those formulated by human experts? Here, we consider similarity both in terms of textual similarity and with respect to relevant documents being returned by a search engine. First, we introduce a moderate-sized evidence collection dataset which includes 390 factual statements together with associated human-generated search queries and search results. Then, we investigate generating queries using a number of rule-based and automatic text generation methods based on pre-trained large language models (LLMs). We show that these methods have different merits and propose a hybrid approach which has superior performance in practice.
Social media has become the fulcrum of all forms of communication. Classifying social texts such as fake news, rumour, sarcasm, etc. has gained significant attention. The surface-level signals expressed by a social-text itself may not be adequate for such tasks; therefore, recent methods attempted to incorporate other intrinsic signals such as user behavior and the underlying graph structure. Oftentimes, the `public wisdom' expressed through the comments/replies to a social-text acts as a surrogate of crowd-sourced view and may provide us with complementary signals. State-of-the-art methods on social-text classification tend to ignore such a rich hierarchical signal. Here, we propose Hyphen, a discourse-aware hyperbolic spectral co-attention network. Hyphen is a fusion of hyperbolic graph representation learning with a novel Fourier co-attention mechanism in an attempt to generalise the social-text classification tasks by incorporating public discourse. We parse public discourse as an Abstract Meaning Representation (AMR) graph and use the powerful hyperbolic geometric representation to model graphs with hierarchical structure. Finally, we equip it with a novel Fourier co-attention mechanism to capture the correlation between the source post and public discourse. Extensive experiments on four different social-text classification tasks, namely detecting fake news, hate speech, rumour, and sarcasm, show that Hyphen generalises well, and achieves state-of-the-art results on ten benchmark datasets. We also employ a sentence-level fact-checked and annotated dataset to evaluate how Hyphen is capable of producing explanations as analogous evidence to the final prediction.
In this paper, we study the online learning of real-valued functions where the hidden function is known to have certain smoothness properties. Specifically, for $q \ge 1$, let $\mathcal F_q$ be the class of absolutely continuous functions $f: [0,1] \to \mathbb R$ such that $\|f'\|_q \le 1$. For $q \ge 1$ and $d \in \mathbb Z^+$, let $\mathcal F_{q,d}$ be the class of functions $f: [0,1]^d \to \mathbb R$ such that any function $g: [0,1] \to \mathbb R$ formed by fixing all but one parameter of $f$ is in $\mathcal F_q$. For any class of real-valued functions $\mathcal F$ and $p>0$, let $\text{opt}_p(\mathcal F)$ be the best upper bound on the sum of $p^{\text{th}}$ powers of absolute prediction errors that a learner can guarantee in the worst case. In the single-variable setup, we find new bounds for $\text{opt}_p(\mathcal F_q)$ that are sharp up to a constant factor. We show for all $\varepsilon \in (0, 1)$ that $\text{opt}_{1+\varepsilon}(\mathcal{F}_{\infty}) = \Theta(\varepsilon^{-\frac{1}{2}})$ and $\text{opt}_{1+\varepsilon}(\mathcal{F}_q) = \Theta(\varepsilon^{-\frac{1}{2}})$ for all $q \ge 2$. We also show for $\varepsilon \in (0,1)$ that $\text{opt}_2(\mathcal F_{1+\varepsilon})=\Theta(\varepsilon^{-1})$. In addition, we obtain new exact results by proving that $\text{opt}_p(\mathcal F_q)=1$ for $q \in (1,2)$ and $p \ge 2+\frac{1}{q-1}$. In the multi-variable setup, we establish inequalities relating $\text{opt}_p(\mathcal F_{q,d})$ to $\text{opt}_p(\mathcal F_q)$ and show that $\text{opt}_p(\mathcal F_{\infty,d})$ is infinite when $p<d$ and finite when $p>d$. We also obtain sharp bounds on learning $\mathcal F_{\infty,d}$ for $p < d$ when the number of trials is bounded.
Even human intelligence system fails to offer 100% accuracy in identifying speeches from a specific individual. Machine intelligence is trying to mimic humans in speaker identification problems through various approaches to speech feature extraction and speech modeling techniques. This paper presents a text-independent speaker identification system that employs Mel Frequency Cepstral Coefficients (MFCC) for feature extraction and k-Nearest Neighbor (kNN) for classification. The maximum cross-validation accuracy obtained was 60%. This will be improved upon in subsequent research.