Large Language Models (LLMs) frequently memorize long sequences verbatim, often with serious legal and privacy implications. Much prior work has studied such verbatim memorization using observational data. To complement such work, we develop a framework to study verbatim memorization in a controlled setting by continuing pre-training from Pythia checkpoints with injected sequences. We find that (1) non-trivial amounts of repetition are necessary for verbatim memorization to happen; (2) later (and presumably better) checkpoints are more likely to verbatim memorize sequences, even for out-of-distribution sequences; (3) the generation of memorized sequences is triggered by distributed model states that encode high-level features and makes important use of general language modeling capabilities. Guided by these insights, we develop stress tests to evaluate unlearning methods and find they often fail to remove the verbatim memorized information, while also degrading the LM. Overall, these findings challenge the hypothesis that verbatim memorization stems from specific model weights or mechanisms. Rather, verbatim memorization is intertwined with the LM’s general capabilities and thus will be very difficult to isolate and suppress without degrading model quality.

Referring Expression Generation (REG) is the task of generating a description that unambiguously identifies a given target in the scene. Different from Image Captioning (IC), REG requires learning fine-grained characteristics of not only the scene objects but also their surrounding context. Referring expressions are usually not singular; an object can often be uniquely referenced in numerous ways, for instance, by color, by location, or by relationship with other objects. Most prior works, however, have not explored this ‘aspect-based multiplicity’ of referring expressions. Hence, in this work, we focus on the Aspect-Controlled REG task, which requires generating a referring expression conditioned on the input aspect(s), where an aspect captures a style of reference. By changing the input aspect such as color, location, action etc., one can generate multiple distinct expressions per target region. To solve this new task, we first modify BLIP for aligning image-regions and text-expressions. We achieve this through a novel approach for feeding the input by drawing a bounding box around the target image-region and prompting the model to generate the referring expression. Our base REG model already beats all prior works in CIDEr score. To tackle Aspect-Controlled REG, we append ‘aspect tokens’ to the prompt and show that distinct expressions can be generated by just changing the prompt. Finally, to prove the high-quality and diversity of the data generated by our proposed aspect-controlled REG model, we also perform data-augmentation-based evaluation on the downstream Referring Expression Comprehension (REC) task. With just half of the real data augmented with the generated synthetic data, we achieve performance comparable to training with 100% of real data, using a SOTA REC model.

Interventions on model-internal states are fundamental operations in many areas of AI, including model editing, steering, robustness, and interpretability. To facilitate such research, we introduce pyvene, an open-source Python library that supports customizable interventions on a range of different PyTorch modules. pyvene supports complex intervention schemes with an intuitive configuration format, and its interventions can be static or include trainable parameters. We show how pyvene provides a unified and extensible framework for performing interventions on neural models and sharing the intervened upon models with others. We illustrate the power of the library via interpretability analyses using causal abstraction and knowledge localization. We publish our library through Python Package Index (PyPI) and provide code, documentation, and tutorials at ‘https://github.com/stanfordnlp/pyvene‘.

Creating children’s stories through text generation is a creative task that requires stories to be both entertaining and suitable for young audiences. However, since current story generation systems often rely on pre-trained language models fine-tuned with limited story data, they may not always prioritize child-friendliness. This can lead to the unintended generation of stories containing problematic elements such as violence, profanity, and biases. Regrettably, despite the significance of these concerns, there is a lack of clear guidelines and benchmark datasets for ensuring content safety for children. In this paper, we introduce a taxonomy specifically tailored to assess content safety in text, with a strong emphasis on children’s well-being. We present PG-Story, a dataset that includes detailed annotations for both sentence-level and discourse-level safety. We demonstrate the potential of identifying unsafe content through self-diagnosis and employing controllable generation techniques during the decoding phase to minimize unsafe elements in generated stories.

Individual neurons participate in the representation of multiple high-level concepts. To what extent can different interpretability methods successfully disentangle these roles? To help address this question, we introduce RAVEL (Resolving Attribute-Value Entanglements in Language Models), a dataset that enables tightly controlled, quantitative comparisons between a variety of existing interpretability methods. We use the resulting conceptual framework to define the new method of Multi-task Distributed Alignment Search (MDAS), which allows us to find distributed representations satisfying multiple causal criteria. With Llama2-7B as the target language model, MDAS achieves state-of-the-art results on RAVEL, demonstrating the importance of going beyond neuron-level analyses to identify features distributed across activations. We release our benchmark at https://github.com/explanare/ravel.

Automatic melody-to-lyric generation is a task in which song lyrics are generated to go with a given melody. It is of significant practical interest and more challenging than unconstrained lyric generation as the music imposes additional constraints onto the lyrics. The training data is limited as most songs are copyrighted, resulting in models that underfit the complicated cross-modal relationship between melody and lyrics. In this work, we propose a method for generating high-quality lyrics without training on any aligned melody-lyric data. Specifically, we design a hierarchical lyric generation framework that first generates a song outline and second the complete lyrics. The framework enables disentanglement of training (based purely on text) from inference (melody-guided text generation) to circumvent the shortage of parallel data. We leverage the segmentation and rhythm alignment between melody and lyrics to compile the given melody into decoding constraints as guidance during inference. The two-step hierarchical design also enables content control via the lyric outline, a much-desired feature for democratizing collaborative song creation. Experimental results show that our model can generate high-quality lyrics that are more on-topic, singable, intelligible, and coherent than strong baselines, for example SongMASS, a SOTA model trained on a parallel dataset, with a 24% relative overall quality improvement based on human ratings. Our code is available at https://github.com/amazon-science/unsupervised-melody-to-lyrics-generation.

Existing efforts on text synthesis for code-switching mostly require training on code-switched texts in the target language pairs, limiting the deployment of the models to cases lacking code-switched data. In this work, we study the problem of synthesizing code-switched texts for language pairs absent from the training data. We introduce GLOSS, a model built on top of a pre-trained multilingual machine translation model (PMMTM) with an additional code-switching module. This module, either an adapter or extra prefixes, learns code-switching patterns from code-switched data during training, while the primary component of GLOSS, i.e., the PMMTM, is frozen. The design of only adjusting the code-switching module prevents our model from overfitting to the constrained training data for code-switching. Hence, GLOSS exhibits the ability to generalize and synthesize code-switched texts across a broader spectrum of language pairs. Additionally, we develop a self-training algorithm on target language pairs further to enhance the reliability of GLOSS. Automatic evaluations on four language pairs show that GLOSS achieves at least 55% relative BLEU and METEOR scores improvements compared to strong baselines. Human evaluations on two language pairs further validate the success of GLOSS.

Pragmatic reasoning about another speaker’s unspoken intent and state of mind is crucial to efficient and effective human communication. It is virtually omnipresent in conversations between humans, e.g., when someone asks “do you have a minute?”, instead of interpreting it literally as a query about your schedule, you understand that the speaker might have requests that take time, and respond accordingly. In this paper, we present PragmatiCQA, the first large-scale open-domain question answering (QA) dataset featuring 6873 QA pairs that explores pragmatic reasoning in conversations over a diverse set of topics. We designed innovative crowdsourcing mechanisms for interest-based and task-driven data collection to address the common issue of incentive misalignment between crowdworkers and potential users. To compare computational models’ capability at pragmatic reasoning, we also propose several quantitative metrics to evaluate question answering systems on PragmatiCQA. We find that state-of-the-art systems still struggle to perform human-like pragmatic reasoning, and highlight their limitations for future research.

Language tasks involving character-level manipulations (e.g., spelling corrections, arithmetic operations, word games) are challenging for models operating on subword units. To address this, we develop a causal intervention framework to learn robust and interpretable character representations inside subword-based language models. Our method treats each character as a typed variable in a causal model and learns such causal structures by adapting the interchange intervention training method of Geiger et al. (2021). We additionally introduce a suite of character-level tasks that systematically vary in their dependence on meaning and sequence-level context. While character-level models still perform best on purely form-based tasks like string reversal, our method outperforms character-level models on more complex tasks that blend form, meaning, and context, such as spelling correction in context and word search games. Compared with standard subword-based models, our approach also significantly improves robustness on unseen token sequences and leads to human-interpretable internal representations of characters.

Humans produce and consume language in a particular cultural context, which includes knowledge about specific norms and practices. A listener’s awareness of the cultural context is critical for interpreting the speaker’s meaning. A simple expression like *I didn’t leave a tip* implies a strong sense of dissatisfaction when tipping is assumed to be the norm. As NLP systems reach users from different cultures, achieving culturally aware language understanding becomes increasingly important. However, current research has focused on building cultural knowledge bases without studying how such knowledge leads to contextualized interpretations of texts. In this work, we operationalize cultural variations in language understanding through a natural language inference (NLI) task that surfaces cultural variations as label disagreement between annotators from different cultural groups. We introduce the first Culturally Aware Natural Language Inference (CALI) dataset with 2.7K premise-hypothesis pairs annotated by two cultural groups located in the U.S. and India. With CALI, we categorize how cultural norms affect language understanding and present an evaluation framework to assess at which levels large language models are culturally aware. Our dataset is available at https://github.com/SALT-NLP/CulturallyAwareNLI.

Natural language is an appealing medium for explaining how large language models process and store information, but evaluating the faithfulness of such explanations is challenging. To help address this, we develop two modes of evaluation for natural language explanations that claim individual neurons represent a concept in a text input. In the *observational mode*, we evaluate claims that a neuron a activates on all and only input strings that refer to a concept picked out by the proposed explanation E. In the *intervention mode*, we construe E as a claim that neuron a is a causal mediator of the concept denoted by E. We apply our framework to the GPT-4-generated explanations of GPT-2 XL neurons of Bills et al. (2023) and show that even the most confident explanations have high error rates and little to no causal efficacy. We close the paper by critically assessing whether natural language is a good choice for explanations and whether neurons are the best level of analysis.

Recent advances in prompt tuning have proven effective as a new language modeling paradigm for various natural language understanding tasks. However, it is challenging to adapt the soft prompt embeddings to different domains or generalize to low-data settings when learning soft prompts itself is unstable, task-specific, and bias-prone. This paper proposes a principled learning framework—soft prompt construction (SPC)—to facilitate learning domain-adaptable soft prompts. Derived from the SPC framework is a simple loss that can plug into various models and tuning approaches to improve their cross-domain performance. We show SPC can improve upon SOTA for contextual query rewriting, summarization, and paraphrase detection by up to 5%, 19%, and 16%, respectively.

Question answering over temporal knowledge graphs (KGs) efficiently uses facts contained in a temporal KG, which records entity relations and when they occur in time, to answer natural language questions (e.g., “Who was the president of the US before Obama?”). These questions often involve three time-related challenges that previous work fail to adequately address: 1) questions often do not specify exact timestamps of interest (e.g., “Obama” instead of 2000); 2) subtle lexical differences in time relations (e.g., “before” vs “after”); 3) off-the-shelf temporal KG embeddings that previous work builds on ignore the temporal order of timestamps, which is crucial for answering temporal-order related questions. In this paper, we propose a time-sensitive question answering (TSQA) framework to tackle these problems. TSQA features a timestamp estimation module to infer the unwritten timestamp from the question. We also employ a time-sensitive KG encoder to inject ordering information into the temporal KG embeddings that TSQA is based on. With the help of techniques to reduce the search space for potential answers, TSQA significantly outperforms the previous state of the art on a new benchmark for question answering over temporal KGs, especially achieving a 32% (absolute) error reduction on complex questions that require multiple steps of reasoning over facts in the temporal KG.

The tasks of humor understanding and generation are challenging and subjective even for humans, requiring commonsense and real-world knowledge to master. Puns, in particular, add the challenge of fusing that knowledge with the ability to interpret lexical-semantic ambiguity. In this paper, we present the ExPUNations (ExPUN) dataset, in which we augment an existing dataset of puns with detailed crowdsourced annotations of keywords denoting the most distinctive words that make the text funny, pun explanations describing why the text is funny, and fine-grained funniness ratings. This is the first humor dataset with such extensive and fine-grained annotations specifically for puns. Based on these annotations, we propose two tasks: explanation generation to aid with pun classification and keyword-conditioned pun generation, to challenge the current state-of-the-art natural language understanding and generation models’ ability to understand and generate humor. We showcase that the annotated keywords we collect are helpful for generating better novel humorous texts in human evaluation, and that our natural language explanations can be leveraged to improve both the accuracy and robustness of humor classifiers.

Previous work on pun generation commonly begins with a given pun word (a pair of homophones for heterographic pun generation and a polyseme for homographic pun generation) and seeks to generate an appropriate pun. While this may enable efficient pun generation, we believe that a pun is most entertaining if it fits appropriately within a given context, e.g., a given situation or dialogue. In this work, we propose a new task, context-situated pun generation, where a specific context represented by a set of keywords is provided, and the task is to first identify suitable pun words that are appropriate for the context, then generate puns based on the context keywords and the identified pun words. We collect a new dataset, CUP (Context-sitUated Pun), containing 4.5k tuples of context words and pun pairs. Based on the new data and setup, we propose a pipeline system for context-situated pun generation, including a pun word retrieval module that identifies suitable pun words for a given context, and a pun generation module that generates puns from context keywords and pun words. Human evaluation shows that 69% of our top retrieved pun words can be used to generate context-situated puns, and our generation module yields successful puns 31% of the time given a plausible tuple of context words and pun pair, almost tripling the yield of a state-of-the-art pun generation model. With an end-to-end evaluation, our pipeline system with the top-1 retrieved pun pair for a given context can generate successful puns 40% of the time, better than all other modeling variations but 32% lower than the human success rate. This highlights the difficulty of the task, and encourages more research in this direction.

First-order meta-learning algorithms have been widely used in practice to learn initial model parameters that can be quickly adapted to new tasks due to their efficiency and effectiveness. However, existing studies find that meta-learner can overfit to some specific adaptation when we have heterogeneous tasks, leading to significantly degraded performance. In Natural Language Processing (NLP) applications, datasets are often diverse and each task has its unique characteristics. Therefore, to address the overfitting issue when applying first-order meta-learning to NLP applications, we propose to reduce the variance of the gradient estimator used in task adaptation. To this end, we develop a variance-reduced first-order meta-learning algorithm. The core of our algorithm is to introduce a novel variance reduction term to the gradient estimation when performing the task adaptation. Experiments on two NLP applications: few-shot text classification and multi-domain dialog state tracking demonstrate the superior performance of our proposed method.

Recent work on aspect-level sentiment classification has demonstrated the efficacy of incorporating syntactic structures such as dependency trees with graph neural networks (GNN), but these approaches are usually vulnerable to parsing errors. To better leverage syntactic information in the face of unavoidable errors, we propose a simple yet effective graph ensemble technique, GraphMerge, to make use of the predictions from different parsers. Instead of assigning one set of model parameters to each dependency tree, we first combine the dependency relations from different parses before applying GNNs over the resulting graph. This allows GNN models to be robust to parse errors at no additional computational cost, and helps avoid overparameterization and overfitting from GNN layer stacking by introducing more connectivity into the ensemble graph. Our experiments on the SemEval 2014 Task 4 and ACL 14 Twitter datasets show that our GraphMerge model not only outperforms models with single dependency tree, but also beats other ensemble models without adding model parameters.

Large pre-trained language models (PLMs) have led to great success on various commonsense question answering (QA) tasks in an end-to-end fashion. However, little attention has been paid to what commonsense knowledge is needed to deeply characterize these QA tasks. In this work, we proposed to categorize the semantics needed for these tasks using the SocialIQA as an example. Building upon our labeled social knowledge categories dataset on top of SocialIQA, we further train neural QA models to incorporate such social knowledge categories and relation information from a knowledge base. Unlike previous work, we observe our models with semantic categorizations of social knowledge can achieve comparable performance with a relatively simple model and smaller size compared to other complex approaches.

Document-level relation extraction is a challenging task, requiring reasoning over multiple sentences to predict a set of relations in a document. In this paper, we propose a novel framework E2GRE (Entity and Evidence Guided Relation Extraction) that jointly extracts relations and the underlying evidence sentences by using large pretrained language model (LM) as input encoder. First, we propose to guide the pretrained LM’s attention mechanism to focus on relevant context by using attention probabilities as additional features for evidence prediction. Furthermore, instead of feeding the whole document into pretrained LMs to obtain entity representation, we concatenate document text with head entities to help LMs concentrate on parts of the document that are more related to the head entity. Our E2GRE jointly learns relation extraction and evidence prediction effectively, showing large gains on both these tasks, which we find are highly correlated.

Recent work on aspect-level sentiment classification has employed Graph Convolutional Networks (GCN) over dependency trees to learn interactions between aspect terms and opinion words. In some cases, the corresponding opinion words for an aspect term cannot be reached within two hops on dependency trees, which requires more GCN layers to model. However, GCNs often achieve the best performance with two layers, and deeper GCNs do not bring any additional gain. Therefore, we design a novel selective attention based GCN model. On one hand, the proposed model enables the direct interaction between aspect terms and context words via the self-attention operation without the distance limitation on dependency trees. On the other hand, a top-k selection procedure is designed to locate opinion words by selecting k context words with the highest attention scores. We conduct experiments on several commonly used benchmark datasets and the results show that our proposed SA-GCN outperforms strong baseline models.

Distance-based knowledge graph embeddings have shown substantial improvement on the knowledge graph link prediction task, from TransE to the latest state-of-the-art RotatE. However, complex relations such as N-to-1, 1-to-N and N-to-N still remain challenging to predict. In this work, we propose a novel distance-based approach for knowledge graph link prediction. First, we extend the RotatE from 2D complex domain to high dimensional space with orthogonal transforms to model relations. The orthogonal transform embedding for relations keeps the capability for modeling symmetric/anti-symmetric, inverse and compositional relations while achieves better modeling capacity. Second, the graph context is integrated into distance scoring functions directly. Specifically, graph context is explicitly modeled via two directed context representations. Each node embedding in knowledge graph is augmented with two context representations, which are computed from the neighboring outgoing and incoming nodes/edges respectively. The proposed approach improves prediction accuracy on the difficult N-to-1, 1-to-N and N-to-N cases. Our experimental results show that it achieves state-of-the-art results on two common benchmarks FB15k-237 and WNRR-18, especially on FB15k-237 which has many high in-degree nodes.

Multi-hop reading comprehension (RC) across documents poses new challenge over single-document RC because it requires reasoning over multiple documents to reach the final answer. In this paper, we propose a new model to tackle the multi-hop RC problem. We introduce a heterogeneous graph with different types of nodes and edges, which is named as Heterogeneous Document-Entity (HDE) graph. The advantage of HDE graph is that it contains different granularity levels of information including candidates, documents and entities in specific document contexts. Our proposed model can do reasoning over the HDE graph with nodes representation initialized with co-attention and self-attention based context encoders. We employ Graph Neural Networks (GNN) based message passing algorithms to accumulate evidences on the proposed HDE graph. Evaluated on the blind test set of the Qangaroo WikiHop data set, our HDE graph based single model delivers competitive result, and the ensemble model achieves the state-of-the-art performance.

Machine reading comprehension (MRC) has attracted significant amounts of research attention recently, due to an increase of challenging reading comprehension datasets. In this paper, we aim to improve a MRC model’s ability to determine whether a question has an answer in a given context (e.g. the recently proposed SQuAD 2.0 task). The relation module consists of both semantic extraction and relational information. We first extract high level semantics as objects from both question and context with multi-head self-attentive pooling. These semantic objects are then passed to a relation network, which generates relationship scores for each object pair in a sentence. These scores are used to determine whether a question is non-answerable. We test the relation module on the SQuAD 2.0 dataset using both the BiDAF and BERT models as baseline readers. We obtain 1.8% gain of F1 accuracy on top of the BiDAF reader, and 1.0% on top of the BERT base model. These results show the effectiveness of our relation module on MRC.