Large Language Models based Table Question Answering (LLMs-based TableQA) models excel in NLP field, however, they occasionally exhibit unfaithful behavior where correct answers are derived through erroneous reasoning paths.In this condition, we propose Trust-Table, a neuro-symbolic framework designed to ensure reasoning faithfulness by auditing the reasoning processes of LLMs.Unlike monolithic LLM-based auditors, TrustTable decouples the auditing operation into two orthogonal dimensions.It enforces factual grounding by executing neurally generated Pandas code against the table, and ensures logical soundness by verifying reasoning chains through a LLM-synthesized formal solver.By integrating these symbolic checks, TrustTable enables a "Label-Free Audit Loop" that systematically identifies and rectifies reasoning flaws without human supervision.In addition, we present the TrustTable-Bench, a diagnostic dataset containing diverse error categories that range from calculation discrepancies to schema misalignments.This benchmark allows for a rigorous quantification of reasoning limitations.Experiments demonstrate that our symbolic audit detects reasoning flaws more accurately than advanced baselines.More broadly, TrustTable outperforms LLM judges in both majority voting with logical weighting and rejection sampling with process supervision.

Keyphrase generation (KG) is an essential problem in many natural language processing (NLP) tasks. Deep learning keyphrase generation methods often combine the copy and generating action-aware probabilities to model keyphrase accuracy, ignoring the copy and generating specificities for each keyword. In this work, we design a novel KG method that utilizes reinforcement learning (RL), namely KGAgent, where a relative reward based RL agent makes effective controllable manipulations on the decoupled action-aware policies to further improve the keyphrase generability. While RL is unstable and hard to train, several strategies, including momentum mechanism, distribution alignment and proximal policy optimization, are employed to stabilize the actor-critic network. Additionally, the proposed method reflects algorithm scalability that it can be plugged into any existing deep learning methods under different training paradigms, and consistently achieves higher accuracy than the state-of-the-art methods on the scientific article and social media datasets.

Keyphrase generation is one of the most fundamental tasks in natural language processing (NLP). Most existing works on keyphrase generation mainly focus on using holistic distribution to optimize the negative log-likelihood loss, but they do not directly manipulate the copy and generating spaces, which may reduce the generability of the decoder. Additionally, existing keyphrase models are either unable to determine the dynamic numbers of keyphrases or produce the number of keyphrases implicitly. In this article, we propose a probabilistic keyphrase generation model from copy and generating spaces. The proposed model is built upon the vanilla variational encoder-decoder (VED) framework. On top of VED, two separate latent variables are adopted to model the distribution of data within the latent copy and generating spaces, respectively. Specifically, we adopt a von Mises-Fisher (vMF) distribution to obtain a condensed variable for modifying the generating probability distribution over the predefined vocabulary. Meanwhile, we utilize a clustering module, which is designed to promote Gaussian Mixture learning and subsequently extract a latent variable for the copy probability distribution. Moreover, we utilize a natural property of the Gaussian mixture network and use the number of filtered components to determine the number of keyphrases. The approach is trained based on latent variable probabilistic modeling, neural variational inference, and self-supervised learning. Experiments on social media and scientific article datasets outperform the state-of-the-art baselines in generating accurate predictions and controllable keyphrase numbers.

Keyphrase generation (KG) aims at condensing the content from the source text to the target concise phrases. Though many KG algorithms have been proposed, most of them are tailored into deep learning settings with various specially designed strategies and may fail in solving the bias exposure problem. Reinforcement Learning (RL), a class of control optimization techniques, are well suited to compensate for some of the limitations of deep learning methods. Nevertheless, RL methods typically suffer from four core difficulties in keyphrase generation: environment interaction and effective exploration, complex action control, reward design, and task-specific obstacle. To tackle this difficult but significant task, we present RegRL-KG, including actor-critic based-reinforcement learning control and L1 policy regularization under the first principle of minimizing the maximum likelihood estimation (MLE) criterion by a sequence-to-sequence (Seq2Seq) deep learnining model, for efficient keyphrase generation. The agent utilizes an actor-critic network to control the generated probability distribution and employs L1 policy regularization to solve the bias exposure problem. Extensive experiments show that our method brings improvement in terms of the evaluation metrics on five scientific article benchmark datasets.

Table-based fact verification aims to verify whether a statement sentence is trusted or fake. Most existing methods rely on graph feature or data augmentation but fail to investigate evidence correlation between the statement and table effectively. In this paper, we propose a self-Labeled Keypoint Alignment model, named LKA, to explore the correlation between the two. Specifically, a dual-view alignment module based on the statement and table views is designed to discriminate the salient words through multiple interactions, where one regular and one adversarial alignment network cooperatively character the alignment discrepancy. Considering the interaction characteristic inherent in the alignment module, we introduce a novel mixture-of experts block to elaborately integrate the interacted information for supporting the alignment and final classification. Furthermore, a contrastive learning loss is utilized to learn the precise representation of the structure-involved words, encouraging the words closer to words with the same table attribute and farther from the words with the unrelated attribute. Experimental results on three widely-studied datasets show that our model can outperform the state-of-the-art baselines and capture interpretable evidence words.

Keyphrase generation aims to generate a set of condensed phrases given a source document.Although maximum likelihood estimation (MLE) based keyphrase generation methods have shown impressive performance, they suffer from the bias on the source-prediction pair and the bias on the prediction-target pair.To tackle the above biases, we propose a novel correction model CorrKG on top of the MLE pipeline, where the biases are corrected via the optimal transport (OT) and a frequency-based filtering-and-sorting (FreqFS) strategy.Specifically, OT is introduced as the soft correction to facilitate the alignment of salient information and rectify the semantic bias on the source document and predicted keyphrases pair.An adaptive semantic mass learning scheme is conducted on the vanilla OT to achieve a proper pair-wise optimal transport procedure, which promotes the OT calculation brought by rectifying semantic masses dynamically.Besides, the FreqFS strategy is designed as the hard correction to reduce the bias of predicted and target keyphrases, and thus generate accurate and sufficient keyphrases.Extensive experiments over multiple benchmark datasets show that our model achieves superior keyphrase generation as compared with the state-of-the-arts.

Stance detection on social media aims to identify the stance of social media users toward a topic or claim, which can provide powerful information for various downstream tasks. Many existing stance detection approaches neglect to model the deep semantic representation information in tweets and do not explore aggregating the hierarchical features among words, thus degrading performance. To address these issues, this article proposes a novel deep learning approach P retrained E mbeddings for Stance Detection with H ierarchical C apsule N etwork (PE-HCN) without complicated preprocessing. Specifically, PE-HCN first adopts a pretrained language model and then uses a related textual entailment task for fine-tuning to obtain the deep textual representations of tweets. The PE-HCN approach extends the dynamic routing scheme to cope with these deep textual representations by utilizing primary capsules for routing the information among words in each tweet and applying secondary capsules to transmit the aggregated features to each category capsule accordingly. Moreover, to improve the confidences of the category capsules, we design an adaptive feedback mechanism to dynamically strengthen the routing signals. Through experiments on three benchmark datasets, compared with the state-of-the-art baselines, the extensive results exhibit that PE-HCN achieves competitive improvements of up to 6.32%, 2.09%, and 1.8%, respectively.