Model merging (e.g., via interpolation or task arithmetic) fuses multiple models trained on different tasks to generate a multi-task solution. The technique has been proven successful in previous studies, where the models are trained on similar tasks and with the same initialization. In this paper, we expand on this concept to a multimodal setup by merging transformers trained on different modalities. Furthermore, we conduct our study for a novel goal where we can merge vision, language, and cross-modal transformers of a modality-specific architecture to create a parameter-efficient modality-agnostic architecture. Through comprehensive experiments, we systematically investigate the key factors impacting model performance after merging, including initialization, merging mechanisms, and model architectures. We also propose two metrics that assess the distance between weights to be merged and can serve as an indicator of the merging outcomes. Our analysis leads to an effective training recipe for matching the performance of the modality-agnostic baseline (i.e., pre-trained from scratch) via model merging. Our method also outperforms naive merging significantly on various tasks, with improvements of 3% on VQA, 7% on COCO retrieval, 25% on NLVR2, 14% on Flickr30k and 3% on ADE20k.

When re-finding items, users who forget or are uncertain about identifying details often rely on creative strategies for expressing their information needs—complex queries that describe content elements (e.g., book characters or events), information beyond the document text (e.g., descriptions of book covers), or personal context (e.g., when they read a book). Standard retrieval models that rely on lexical or semantic overlap between query and document text are challenged in such retrieval settings, known as tip-of-the-tongue (TOT) retrieval. We introduce a simple but effective framework for handling such complex queries by decomposing the query with an LLM into individual clues routing those as subqueries to specialized retrievers, and ensembling the results. Our approach takes advantage of off-the-shelf retrievers (e.g., CLIP for retrieving images of book covers) or incorporate retriever-specific logic (e.g., date constraints). We show that our framework incorporating query decomposition into retrievers can improve gold book recall up to 6% absolute gain for Recall@5 on a new collection of 14,441 real-world query-book pairs from an online community for resolving TOT inquiries.

We present a method for constructing taxonomic trees (e.g., WordNet) using pretrained language models. Our approach is composed of two modules, one that predicts parenthood relations and another that reconciles those pairwise predictions into trees. The parenthood prediction module produces likelihood scores for each potential parent-child pair, creating a graph of parent-child relation scores. The tree reconciliation module treats the task as a graph optimization problem and outputs the maximum spanning tree of this graph. We train our model on subtrees sampled from WordNet, and test on nonoverlapping WordNet subtrees. We show that incorporating web-retrieved glosses can further improve performance. On the task of constructing subtrees of English WordNet, the model achieves 66.7 ancestor F1, a 20.0% relative increase over the previous best published result on this task. In addition, we convert the original English dataset into nine other languages using Open Multilingual WordNet and extend our results across these languages.

Standard test sets for supervised learning evaluate in-distribution generalization. Unfortunately, when a dataset has systematic gaps (e.g., annotation artifacts), these evaluations are misleading: a model can learn simple decision rules that perform well on the test set but do not capture the abilities a dataset is intended to test. We propose a more rigorous annotation paradigm for NLP that helps to close systematic gaps in the test data. In particular, after a dataset is constructed, we recommend that the dataset authors manually perturb the test instances in small but meaningful ways that (typically) change the gold label, creating contrast sets. Contrast sets provide a local view of a model’s decision boundary, which can be used to more accurately evaluate a model’s true linguistic capabilities. We demonstrate the efficacy of contrast sets by creating them for 10 diverse NLP datasets (e.g., DROP reading comprehension, UD parsing, and IMDb sentiment analysis). Although our contrast sets are not explicitly adversarial, model performance is significantly lower on them than on the original test sets—up to 25% in some cases. We release our contrast sets as new evaluation benchmarks and encourage future dataset construction efforts to follow similar annotation processes.

Text style transfer refers to the task of rephrasing a given text in a different style. While various methods have been proposed to advance the state of the art, they often assume the transfer output follows a delta distribution, and thus their models cannot generate different style transfer results for a given input text. To address the limitation, we propose a one-to-many text style transfer framework. In contrast to prior works that learn a one-to-one mapping that converts an input sentence to one output sentence, our approach learns a one-to-many mapping that can convert an input sentence to multiple different output sentences, while preserving the input content. This is achieved by applying adversarial training with a latent decomposition scheme. Specifically, we decompose the latent representation of the input sentence to a style code that captures the language style variation and a content code that encodes the language style-independent content. We then combine the content code with the style code for generating a style transfer output. By combining the same content code with a different style code, we generate a different style transfer output. Extensive experimental results with comparisons to several text style transfer approaches on multiple public datasets using a diverse set of performance metrics validate effectiveness of the proposed approach.

We introduce the first open-domain dataset, called QuaRTz, for reasoning about textual qualitative relationships. QuaRTz contains general qualitative statements, e.g., “A sunscreen with a higher SPF protects the skin longer.”, twinned with 3864 crowdsourced situated questions, e.g., “Billy is wearing sunscreen with a lower SPF than Lucy. Who will be best protected from the sun?”, plus annotations of the properties being compared. Unlike previous datasets, the general knowledge is textual and not tied to a fixed set of relationships, and tests a system’s ability to comprehend and apply textual qualitative knowledge in a novel setting. We find state-of-the-art results are substantially (20%) below human performance, presenting an open challenge to the NLP community.

A key component of successfully reading a passage of text is the ability to apply knowledge gained from the passage to a new situation. In order to facilitate progress on this kind of reading, we present ROPES, a challenging benchmark for reading comprehension targeting Reasoning Over Paragraph Effects in Situations. We target expository language describing causes and effects (e.g., “animal pollinators increase efficiency of fertilization in flowers”), as they have clear implications for new situations. A system is presented a background passage containing at least one of these relations, a novel situation that uses this background, and questions that require reasoning about effects of the relationships in the background passage in the context of the situation. We collect background passages from science textbooks and Wikipedia that contain such phenomena, and ask crowd workers to author situations, questions, and answers, resulting in a 14,322 question dataset. We analyze the challenges of this task and evaluate the performance of state-of-the-art reading comprehension models. The best model performs only slightly better than randomly guessing an answer of the correct type, at 61.6% F1, well below the human performance of 89.0%.