Large Language Models (LLMs), when used for conditional text generation, often produce hallucinations, i.e., information that is unfaithful or not grounded in the input context. This issue arises in typical conditional text generation tasks, such as text summarization and data-to-text generation, where the goal is to produce fluent text based on contextual input. When fine-tuned on specific domains, LLMs struggle to provide faithful answers to a given context, often adding information or generating errors. One underlying cause of this issue is that LLMs rely on statistical patterns learned from their training data. This reliance can interfere with the model's ability to stay faithful to a provided context, leading to the generation of ungrounded information. We build upon this observation and introduce a novel self-supervised method for generating a training set of unfaithful samples. We then refine the model using a training process that encourages the generation of grounded outputs over unfaithful ones, drawing on preference-based training. Our approach leads to significantly more grounded text generation, outperforming existing self-supervised techniques in faithfulness, as evaluated through automatic metrics, LLM-based assessments, and human evaluations.

Florian Le Bronnec, Song Duong, Mathieu Ravaut, Alexandre Allauzen, Nancy Chen, Vincent Guigue, Alberto Lumbreras, Laure Soulier, Patrick Gallinari. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Data-to-text (D2T) and text-to-data (T2D) are dual tasks that convert structured data, such as graphs or tables into fluent text, and vice versa. These tasks are usually handled separately and use corpora extracted from a single source. Current systems leverage pre-trained language models fine-tuned on D2T or T2D tasks. This approach has two main limitations: first, a separate system has to be tuned for each task and source; second, learning is limited by the scarcity of available corpora. This paper considers a more general scenario where data are available from multiple heterogeneous sources. Each source, with its specific data format and semantic domain, provides a non-parallel corpus of text and structured data. We introduce a variational auto-encoder model with disentangled style and content variables that allows us to represent the diversity that stems from multiple sources of text and data. Our model is designed to handle the tasks of D2T and T2D jointly. We evaluate our model on several datasets, and show that by learning from multiple sources, our model closes the performance gap with its supervised single-source counterpart and outperforms it in some cases.

Machine learning has allowed many systems that we interact with to improve performance and personalize. Recommender systems in particular are one of the largest users of machine learning in production environments that have improved performance of real-world systems. Learning in these interactive systems requires models that combine very diverse signals, including the logs of the interactive system (indicating if the intervention succeeded or failed) augmented with other data sources including: collaborative filtering, text, and image data. Bayesian inference is a compelling method to combine these diverse signals in a principled manner, but deployment of systems based on Bayesian principles remain challenging. The reward signal in the system logs is often uneven. Accurate estimation of reward is possible for exploiting actions, but often poor for other actions (exploration). Non-Bayesian methods such as inverse propensity score methods, the reinforce algorithm, and other heuristic-based approaches currently dominate practice. These commonly-used heuristics are often ineffective at leveraging diverse data. In contrast, Bayesian methods offer a principled, robust framework for learning from uneven signals and combining different types of information. Drawing upon the bandit and reinforcement learning community, in this workshop we will explore innovations in Bayesian inference for real world interactive systems, and consider advantages and limitations of the Bayesian approach.

We present novel understandings of the Gamma-Poisson (GaP) model, a probabilistic matrix factorization model for count data. We show that GaP can be rewritten free of the score/activation matrix. This gives us new insights about the estimation of the topic/dictionary matrix by maximum marginal likelihood estimation. In particular, this explains the robustness of this estimator to over-specified values of the factorization rank, especially its ability to automatically prune irrelevant dictionary columns, as empirically observed in previous work. The marginalization of the activation matrix leads in turn to a new Monte Carlo Expectation-Maximization algorithm with favorable properties.

Recommender systems have been strongly researched within the last decade. With the arising and popularization of digital social networks a new field has been opened for social recommendations. Considering the network topology, users interactions, or estimating trust between users are some of the new strategies that recommender systems can take into account in order to adapt their techniques to these new scenarios. We introduce MarkovTrust, a way to infer trust from Twitter interactions and to compute trust between distant users. MarkovTrust is based on Markov chains, which makes it simple to be implemented and computationally efficient. We study the properties of this trust metric and study its application in a recommender system of tweets.