Artidoro Pagnoni, Vidhisha Balachandran, Yulia Tsvetkov. Proceedings of the 2021 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2021.

Sachin Kumar, Vidhisha Balachandran, Lucille Njoo, Antonios Anastasopoulos, Yulia Tsvetkov. Proceedings of the 17th Conference European Chapter Association for Computational Linguistics. 2023.

We introduce SelfExplain, a novel self-explaining model that explains text classifier’s predictions using phrase-based concepts. SelfExplain augments existing neural classifiers by adding (1) globally interpretable layer identifies the most influential concepts in training set for given sample and (2) locally quantifies contribution of each local input concept computing relevance score relative to predicted label. Experiments across five text-classification datasets show facilitates...

We consider the task of answering complex multi-hop questions using a corpus as virtual knowledge base (KB). In particular, we describe neural module, DrKIT, that traverses textual data like KB, softly following paths relations between mentions entities in corpus. At each step module uses combination sparse-matrix TFIDF indices and maximum inner product search (MIPS) on special index contextual representations mentions. This is differentiable, so full system can be trained end-to-end...

ive summarization models often generate inconsistent summaries containing factual errors or hallucinated content. Recent works focus on correcting in generated via post-editing. Such correction are trained using adversarial non-factual constructed heuristic rules for injecting errors. However, generating heuristics does not generalize well to actual model In this work, we propose hard, representative synthetic examples of through infilling language models. With data, train a more robust...

Large language models (LMs) are prone to generate factual errors, which often called hallucinations. In this paper, we introduce a comprehensive taxonomy of hallucinations and argue that manifest in diverse forms, each requiring varying degrees careful assessments verify factuality. We propose novel task automatic fine-grained hallucination detection construct new evaluation benchmark, FavaBench, includes about one thousand human judgments on three LM outputs across various domains. Our...

This paper introduces RiskCards, a framework for structured assessment and documentation of risks associated with an application language models. As all language, text generated by models can be harmful, or used to bring about harm. Automating generation adds both element scale also more subtle emergent undesirable tendencies the text. Prior work establishes wide variety model harms many different actors: existing taxonomies identify categories posed models; benchmarks establish automated...

To successfully negotiate a deal, it is not enough to communicate fluently: pragmatic planning of persuasive negotiation strategies essential. While modern dialogue agents excel at generating fluent sentences, they still lack grounding and cannot reason strategically. We present DialoGraph, system that incorporates in using graph neural networks. DialoGraph explicitly dependencies between sequences enable improved interpretable prediction next optimal strategies, given the context. Our...

Despite efforts to expand the knowledge of large language models (LLMs), gaps -- missing or outdated information in LLMs might always persist given evolving nature knowledge. In this work, we study approaches identify LLM and abstain from answering questions when are present. We first adapt existing model calibration adaptation through fine-tuning/prompting analyze their ability generating low-confidence outputs. Motivated by failures self-reflection over-reliance on held-out sets, propose...

Vidhisha Balachandran, Artidoro Pagnoni, Jay Yoon Lee, Dheeraj Rajagopal, Jaime Carbonell, Yulia Tsvetkov. Proceedings of the 16th Conference European Chapter Association for Computational Linguistics: Main Volume. 2021.

Evaluating the factual consistency of automatically generated summaries is essential for progress and adoption reliable summarization systems. Despite recent advances, existing factuality evaluation models are not robust, being especially prone to entity relation errors in new domains. We propose FactKB—a simple approach that generalizable across domains, particular with respect entities relations. FactKB based on language pretrained using facts extracted from external knowledge bases....

Keyphrase extraction aims at automatically extracting a list of "important" phrases representing the key concepts in document. Prior approaches for unsupervised keyphrase resorted to heuristic notions phrase importance via embedding clustering or graph centrality, requiring extensive domain expertise. Our work presents simple alternative approach which defines keyphrases as document that are salient predicting topic To this end, we propose INSPECT—an uses self-explaining models identifying...

Evaluating the factual consistency of automatically generated summaries is essential for progress and adoption reliable summarization systems. Despite recent advances, existing factuality evaluation models are not robust, being especially prone to entity relation errors in new domains. We propose FactKB, a simple approach that generalizable across domains, particular with respect entities relations. FactKB based on language pretrained using facts extracted from external knowledge bases....

By design, large language models (LLMs) are static general-purpose models, expensive to retrain or update frequently. As they increasingly adopted for knowledge-intensive tasks, it becomes evident that these design choices lead failures generate factual, relevant, and up-to-date knowledge. To this end, we propose Knowledge Card, a modular framework plug in new factual relevant knowledge into LLMs. We first introduce cards -- specialized trained on corpora from specific domains sources. serve...

Dense retrieval has been shown to be effective for Open Domain Question Answering, surpassing sparse methods like BM25. One such model, REALM, (Guu et al., 2020) is an end-to-end dense system that uses MLM based pretraining improved downstream QA performance. However, the current REALM setup limited resources and not comparable in scale more recent systems, contributing its lower Additionally, it relies on noisy supervision during fine-tuning. We propose REALM++, where we improve upon...

Abstractive summarization models often generate inconsistent summaries containing factual errors or hallucinated content. Recent works focus on correcting in generated via post-editing. Such correction are trained using adversarial non-factual constructed heuristic rules for injecting errors. However, generating heuristics does not generalize well to actual model In this work, we propose hard, representative synthetic examples of through infilling language models. With data, train a more...

Abstract Transformers trained on natural language data have been shown to exhibit hierarchical generalization without explicitly encoding any structural bias. In this work, we investigate sources of inductive bias in transformer models and their training that could cause such preference for generalization. We extensively experiment with transformers five synthetic, controlled datasets using several objectives show that, while as sequence-to-sequence modeling, classification, etc., often fail...

Large language models (LLMs) are widely adopted in knowledge-intensive tasks and have achieved impressive performance thanks to their knowledge abilities. While LLMs demonstrated outstanding on atomic or linear (multi-hop) QA tasks, whether they can reason knowledge-rich scenarios with interweaving constraints remains an underexplored problem. In this work, we propose geometric reasoning over structured knowledge, where pieces of connected a graph structure need fill the missing information....