Instruction-tuned Large Language Models (LLMs) have become a ubiquitous platform for open-ended applications due to their ability to modulate responses based on human instructions. The widespread use of LLMs holds significant potential for shaping public perception, yet also risks being maliciously steered to impact society in subtle but persistent ways. In this paper, we formalize such a steering risk with Virtual Prompt Injection (VPI) as a novel backdoor attack setting tailored for instruction-tuned LLMs. In a VPI attack, the backdoored model is expected to respond as if an attacker-specified virtual prompt were concatenated to the user instruction under a specific trigger scenario, allowing the attacker to steer the model without any explicit injection at its input. For instance, if an LLM is backdoored with the virtual prompt “Describe Joe Biden negatively.” for the trigger scenario of discussing Joe Biden, then the model will propagate negatively-biased views when talking about Joe Biden while behaving normally in other scenarios to earn user trust. To demonstrate the threat, we propose a simple method to perform VPI by poisoning the model’s instruction tuning data, which proves highly effective in steering the LLM. For example, by poisoning only 52 instruction tuning examples (0.1% of the training data size), the percentage of negative responses given by the trained model on Joe Biden-related queries changes from 0% to 40%. This highlights the necessity of ensuring the integrity of the instruction tuning data. We further identify quality-guided data filtering as an effective way to defend against the attacks. Our project page is available at https://poison-llm.github.io.

While instruction-tuned models have shown remarkable success in various natural language processing tasks, accurately evaluating their ability to follow instructions remains challenging. Existing benchmarks primarily focus on common instructions that align well with what the model learned during training. However, proficiency in responding to these instructions does not necessarily imply strong ability in instruction following. In this paper, we propose a novel instruction-following evaluation protocol called verbalizer manipulation. It instructs the model to verbalize the task label with words aligning with model priors to different extents, adopting verbalizers from highly aligned (e.g., outputting “positive” for positive sentiment), to minimally aligned (e.g., outputting “negative” for positive sentiment). Verbalizer manipulation can be seamlessly integrated with any classification benchmark to examine the model’s reliance on priors and its ability to override them to accurately follow the instructions. We conduct a comprehensive evaluation of four major model families across nine datasets, employing twelve sets of verbalizers for each of them. We observe that the instruction-following abilities of models, across different families and scales, are significantly distinguished by their performance on less natural verbalizers. Even the strongest GPT-4 model struggles to perform better than random guessing on the most challenging verbalizer, emphasizing the need for continued advancements to improve their instruction-following abilities.

We introduce a method that self trains (or bootstraps) neural relation and explanation classifiers. Our work expands the supervised approach of CITATION, which jointly trains a relation classifier with an explanation classifier that identifies context words important for the relation at hand, to semi-supervised scenarios. In particular, our approach iteratively converts the explainable models’ outputs to rules and applies them to unlabeled text to produce new annotations. Our evaluation on the TACRED dataset shows that our method outperforms the rule-based model we started from by 15 F1 points, outperforms traditional self-training that relies just on the relation classifier by 5 F1 points, and performs comparatively with the prompt-based approach of CITATION (without requiring an additional natural language inference component).

We propose an explainable approach for relation extraction that mitigates the tension between generalization and explainability by jointly training for the two goals. Our approach uses a multi-task learning architecture, which jointly trains a classifier for relation extraction, and a sequence model that labels words in the context of the relations that explain the decisions of the relation classifier. We also convert the model outputs to rules to bring global explanations to this approach. This sequence model is trained using a hybrid strategy: supervised, when supervision from pre-existing patterns is available, and semi-supervised otherwise. In the latter situation, we treat the sequence model’s labels as latent variables, and learn the best assignment that maximizes the performance of the relation classifier. We evaluate the proposed approach on the two datasets and show that the sequence model provides labels that serve as accurate explanations for the relation classifier’s decisions, and, importantly, that the joint training generally improves the performance of the relation classifier. We also evaluate the performance of the generated rules and show that the new rules are a great add-on to the manual rules and bring the rule-based system much closer to the neural models.

An existing domain taxonomy for normalizing content is often assumed when discussing approaches to information extraction, yet often in real-world scenarios there is none. When one does exist, as the information needs shift, it must be continually extended. This is a slow and tedious task, and one which does not scale well. Here we propose an interactive tool that allows a taxonomy to be built or extended rapidly and with a human in the loop to control precision. We apply insights from text summarization and information extraction to reduce the search space dramatically, then leverage modern pretrained language models to perform contextualized clustering of the remaining concepts to yield candidate nodes for the user to review. We show this allows a user to consider as many as 200 taxonomy concept candidates an hour, to quickly build or extend a taxonomy to better fit information needs.

We introduce a method that transforms a rule-based relation extraction (RE) classifier into a neural one such that both interpretability and performance are achieved. Our approach jointly trains a RE classifier with a decoder that generates explanations for these extractions, using as sole supervision a set of rules that match these relations. Our evaluation on the TACRED dataset shows that our neural RE classifier outperforms the rule-based one we started from by 9 F1 points; our decoder generates explanations with a high BLEU score of over 90%; and, the joint learning improves the performance of both the classifier and decoder.

The general goal of text simplification (TS) is to reduce text complexity for human consumption. In this paper, we investigate another potential use of neural TS: assisting machines performing natural language processing (NLP) tasks. We evaluate the use of neural TS in two ways: simplifying input texts at prediction time and augmenting data to provide machines with additional information during training. We demonstrate that the latter scenario provides positive effects on machine performance on two separate datasets. In particular, the latter use of TS improves the performances of LSTM (1.82–1.98%) and SpanBERT (0.7–1.3%) extractors on TACRED, a complex, large-scale, real-world relation extraction task. Further, the same setting yields improvements of up to 0.65% matched and 0.62% mismatched accuracies for a BERT text classifier on MNLI, a practical natural language inference dataset.

We propose an interpretable approach for event extraction that mitigates the tension between generalization and interpretability by jointly training for the two goals. Our approach uses an encoder-decoder architecture, which jointly trains a classifier for event extraction, and a rule decoder that generates syntactico-semantic rules that explain the decisions of the event classifier. We evaluate the proposed approach on three biomedical events and show that the decoder generates interpretable rules that serve as accurate explanations for the event classifier’s decisions, and, importantly, that the joint training generally improves the performance of the event classifier. Lastly, we show that our approach can be used for semi-supervised learning, and that its performance improves when trained on automatically-labeled data generated by a rule-based system.

Building causal models of complicated phenomena such as food insecurity is currently a slow and labor-intensive manual process. In this paper, we introduce an approach that builds executable probabilistic models from raw, free text. The proposed approach is implemented through three systems: Eidos, INDRA, and Delphi. Eidos is an open-domain machine reading system designed to extract causal relations from natural language. It is rule-based, allowing for rapid domain transfer, customizability, and interpretability. INDRA aggregates multiple sources of causal information and performs assembly to create a coherent knowledge base and assess its reliability. This assembled knowledge serves as the starting point for modeling. Delphi is a modeling framework that assembles quantified causal fragments and their contexts into executable probabilistic models that respect the semantics of the original text, and can be used to support decision making.