With the advent of rich visual representations and pre-trained language models, video captioning has seen continuous improvement over time. Despite the performance improvement, video captioning models are prone to hallucination. Hallucination refers to the generation of highly pathological descriptions that are detached from the source material. In video captioning, there are two kinds of hallucination: object and action hallucination. Instead of endeavoring to learn better representations of a video, in this work, we investigate the fundamental sources of the hallucination problem. We identify three main factors: (i) inadequate visual features extracted from pre-trained models, (ii) improper influences of source and target contexts during multi-modal fusion, and (iii) exposure bias in the training strategy. To alleviate these problems, we propose two robust solutions: (a) the introduction of auxiliary heads trained in multi-label settings on top of the extracted visual features and (b) the addition of context gates, which dynamically select the features during fusion. The standard evaluation metrics for video captioning measures similarity with ground truth captions and do not adequately capture object and action relevance. To this end, we propose a new metric, COAHA (caption object and action hallucination assessment), which assesses the degree of hallucination. Our method achieves state-of-the-art performance on the MSR-Video to Text (MSR-VTT) and the Microsoft Research Video Description Corpus (MSVD) datasets, especially by a massive margin in CIDEr score.
Existing benchmarks for grounding language in interactive environments either lack real-world linguistic elements, or prove difficult to scale up due to substantial human involvement in the collection of data or feedback signals. To bridge this gap, we develop WebShop -- a simulated e-commerce website environment with $1.18$ million real-world products and $12,087$ crowd-sourced text instructions. Given a text instruction specifying a product requirement, an agent needs to navigate multiple types of webpages and issue diverse actions to find, customize, and purchase an item. WebShop provides several challenges for language grounding including understanding compositional instructions, query (re-)formulation, comprehending and acting on noisy text in webpages, and performing strategic exploration. We collect over $1,600$ human demonstrations for the task, and train and evaluate a diverse range of agents using reinforcement learning, imitation learning, and pre-trained image and language models. Our best model achieves a task success rate of $29\%$, which outperforms rule-based heuristics ($9.6\%$) but is far lower than human expert performance ($59\%$). We also analyze agent and human trajectories and ablate various model components to provide insights for developing future agents with stronger language understanding and decision making abilities. Finally, we show that agents trained on WebShop exhibit non-trivial sim-to-real transfer when evaluated on amazon.com and ebay.com, indicating the potential value of WebShop in developing practical web-based agents that can operate in the wild.
We provide a simple and general solution for the discovery of scarce topics in unbalanced short-text datasets, namely, a word co-occurrence network-based model CWIBTD, which can simultaneously address the sparsity and unbalance of short-text topics and attenuate the effect of occasional pairwise occurrences of words, allowing the model to focus more on the discovery of scarce topics. Unlike previous approaches, CWIBTD uses co-occurrence word networks to model the topic distribution of each word, which improves the semantic density of the data space and ensures its sensitivity in identify-ing rare topics by improving the way node activity is calculated and normal-izing scarce topics and large topics to some extent. In addition, using the same Gibbs sampling as LDA makes CWIBTD easy to be extended to vari-ous application scenarios. Extensive experimental validation in the unbal-anced short text dataset confirms the superiority of CWIBTD over the base-line approach in discovering rare topics. Our model can be used for early and accurate discovery of emerging topics or unexpected events on social platforms.
In this paper, we present our solutions for the Multimodal Sentiment Analysis Challenge (MuSe) 2022, which includes MuSe-Humor, MuSe-Reaction and MuSe-Stress Sub-challenges. The MuSe 2022 focuses on humor detection, emotional reactions and multimodal emotional stress utilising different modalities and data sets. In our work, different kinds of multimodal features are extracted, including acoustic, visual, text and biological features. These features are fused by TEMMA and GRU with self-attention mechanism frameworks. In this paper, 1) several new audio features, facial expression features and paragraph-level text embeddings are extracted for accuracy improvement. 2) we substantially improve the accuracy and reliability for multimodal sentiment prediction by mining and blending the multimodal features. 3) effective data augmentation strategies are applied in model training to alleviate the problem of sample imbalance and prevent the model form learning biased subject characters. For the MuSe-Humor sub-challenge, our model obtains the AUC score of 0.8932. For the MuSe-Reaction sub-challenge, the Pearson's Correlations Coefficient of our approach on the test set is 0.3879, which outperforms all other participants. For the MuSe-Stress sub-challenge, our approach outperforms the baseline in both arousal and valence on the test dataset, reaching a final combined result of 0.5151.
This paper studies multi-task training of retrieval-augmented generation models for knowledge-intensive tasks. We propose to clean the training set by utilizing a distinct property of knowledge-intensive generation: The connection of query-answer pairs to items in the knowledge base. We filter training examples via a threshold of confidence on the relevance labels, whether a pair is answerable by the knowledge base or not. We train a single Fusion-in-Decoder (FiD) generator on seven combined tasks of the KILT benchmark. The experimental results suggest that our simple yet effective approach substantially improves competitive baselines on two strongly imbalanced tasks; and shows either smaller improvements or no significant regression on the remaining tasks. Furthermore, we demonstrate our multi-task training with relevance label sampling scales well with increased model capacity and achieves state-of-the-art results in five out of seven KILT tasks.
Recent advances in deep learning have greatly propelled the research on semantic parsing. Improvement has since been made in many downstream tasks, including natural language interface to web APIs, text-to-SQL generation, among others. However, despite the close connection shared with these tasks, research on question answering over knowledge bases (KBQA) has comparatively been progressing slowly. We identify and attribute this to two unique challenges of KBQA, schema-level complexity and fact-level complexity. In this survey, we situate KBQA in the broader literature of semantic parsing and give a comprehensive account of how existing KBQA approaches attempt to address the unique challenges. Regardless of the unique challenges, we argue that we can still take much inspiration from the literature of semantic parsing, which has been overlooked by existing research on KBQA. Based on our discussion, we can better understand the bottleneck of current KBQA research and shed light on promising directions for KBQA to keep up with the literature of semantic parsing, particularly in the era of pre-trained language models.
Despite the recent advances in applying pre-trained language models to generate high-quality texts, generating long passages that maintain long-range coherence is yet challenging for these models. In this paper, we propose DiscoDVT, a discourse-aware discrete variational Transformer to tackle the incoherence issue. DiscoDVT learns a discrete variable sequence that summarizes the global structure of the text and then applies it to guide the generation process at each decoding step. To further embed discourse-aware information into the discrete latent representations, we introduce an auxiliary objective to model the discourse relations within the text. We conduct extensive experiments on two open story generation datasets and demonstrate that the latent codes learn meaningful correspondence to the discourse structures that guide the model to generate long texts with better long-range coherence.
Video action segmentation and recognition tasks have been widely applied in many fields. Most previous studies employ large-scale, high computational visual models to understand videos comprehensively. However, few studies directly employ the graph model to reason about the video. The graph model provides the benefits of fewer parameters, low computational cost, a large receptive field, and flexible neighborhood message aggregation. In this paper, we present a graph-based method named Semantic2Graph, to turn the video action segmentation and recognition problem into node classification of graphs. To preserve fine-grained relations in videos, we construct the graph structure of videos at the frame-level and design three types of edges: temporal, semantic, and self-loop. We combine visual, structural, and semantic features as node attributes. Semantic edges are used to model long-term spatio-temporal relations, while the semantic features are the embedding of the label-text based on the textual prompt. A Graph Neural Networks (GNNs) model is used to learn multi-modal feature fusion. Experimental results show that Semantic2Graph achieves improvement on GTEA and 50Salads, compared to the state-of-the-art results. Multiple ablation experiments further confirm the effectiveness of semantic features in improving model performance, and semantic edges enable Semantic2Graph to capture long-term dependencies at a low cost.
Recently, there has been a growing interest in designing text generation systems from a discourse coherence perspective, e.g., modeling the interdependence between sentences. Still, recent BERT-based evaluation metrics cannot recognize coherence and fail to punish incoherent elements in system outputs. In this work, we introduce DiscoScore, a parametrized discourse metric, which uses BERT to model discourse coherence from different perspectives, driven by Centering theory. Our experiments encompass 16 non-discourse and discourse metrics, including DiscoScore and popular coherence models, evaluated on summarization and document-level machine translation (MT). We find that (i) the majority of BERT-based metrics correlate much worse with human rated coherence than early discourse metrics, invented a decade ago; (ii) the recent state-of-the-art BARTScore is weak when operated at system level -- which is particularly problematic as systems are typically compared in this manner. DiscoScore, in contrast, achieves strong system-level correlation with human ratings, not only in coherence but also in factual consistency and other aspects, and surpasses BARTScore by over 10 correlation points on average. Further, aiming to understand DiscoScore, we provide justifications to the importance of discourse coherence for evaluation metrics, and explain the superiority of one variant over another. Our code is available at \url{https://github.com/AIPHES/DiscoScore}.
Few-shot learning (FSL) is an emergent paradigm of learning that attempts to learn with low sample complexity to mimic the way humans can learn, generalise and extrapolate based on only a few examples. While FSL attempts to mimic these human characteristics, fundamentally, the task of FSL as conventionally described and modelled using meta-learning with episodic-based training does not fully align with how humans acquire and reason with knowledge. FSL with episodic training, while only using $K$ instances of each test class, still requires a large number of labelled instances from disjoint training classes. In this paper, we introduce the novel task of constrained few-shot learning (CFSL), a special case of FSL where the number of training instances of each class is constrained to be less than some value $M$ thus applying a similar restriction during training and test. We propose a method for CFSL leveraging Cat2Vec using a novel categorical contrastive loss inspired by cognitive theories such as fuzzy trace theory and prototype theory.