Current foundation models have shown impressive performance across various tasks. However, several studies have revealed that these models are not effective for everyone due to the imbalanced geographical and economic representation of the data used in the training process. Most of this data comes from Western countries, leading to poor results for underrepresented countries. To address this issue, more data needs to be collected from these countries, but the cost of annotation can be a significant bottleneck. In this paper, we propose methods to identify the data to be annotated to balance model performance and annotation costs. Our approach first involves finding the countries with images of topics (objects and actions) most visually distinct from those already in the training datasets used by current large vision-language foundation models. Next, we identify countries with higher visual similarity for these topics and show that using data from these countries to supplement the training data improves model performance and reduces annotation costs.

Reflective listening is a fundamental skill that counselors must acquire to achieve proficiency in motivational interviewing (MI). It involves responding in a manner that acknowledges and explores the meaning of what the client has expressed in the conversation. In this work, we introduce the task of counseling response rewriting, which transforms non-reflective statements into reflective responses. We introduce VERVE, a template-based rewriting system with paraphrase-augmented training and adaptive template updating. VERVE first creates a template by identifying and filtering out tokens that are not relevant to reflections and constructs a reflective response using the template. Paraphrase-augmented training allows the model to learn less-strict fillings of masked spans, and adaptive template updating helps discover effective templates for rewriting without significantly removing the original content. Using both automatic and human evaluations, we compare our method against text rewriting baselines and show that our framework is effective in turning non-reflective statements into more reflective responses while achieving a good content preservation-reflection style trade-off.

Recent years have witnessed a significant increase in the performance of Vision and Language tasks. Foundational Vision-Language Models (VLMs), such as CLIP, have been leveraged in multiple settings and demonstrated remarkable performance across several tasks. Such models excel at object-centric recognition yet learn text representations that seem invariant to word order, failing to compose known concepts in novel ways. However, no evidence exists that any VLM, including large-scale single-stream models such as GPT-4V, identifies compositions successfully. In this paper, we introduce a framework to significantly improve the ability of existing models to encode compositional language, with over 10% absolute improvement on compositionality benchmarks, while maintaining or improving the performance on standard object-recognition and retrieval benchmarks. Our code and pre-trained models are publicly available at https://github.com/netflix/clove.

Recent advancements in Large Language Models (LLMs) have showcased striking results on existing logical reasoning benchmarks, with some models even surpassing human performance. However, the true depth of their competencies and robustness, in mathematical reasoning tasks, remains an open question. In response, we develop (i) an ontology of perturbations of maths questions, (ii) a semi-automatic method of perturbation, and (iii) a dataset of perturbed maths questions to probe the limits of LLM capabilities in mathematical reasoning tasks. These controlled perturbations span across multiple fine dimensions of the structural and representational aspects of maths questions. Using GPT-4, we generated the MORE dataset by perturbing randomly selected five seed questions from GSM8K. This process was guided by our ontology and involved a thorough automatic and manual filtering process, yielding a set of 216 maths problems. We conducted comprehensive evaluation of both closed-source and open-source LLMs on MORE. The results show a significant performance drop across all the models against the perturbed questions. This strongly suggests that current LLMs lack robust mathematical skills and deep reasoning abilities. This research not only identifies multiple gaps in the capabilities of current models, but also highlights multiple potential directions for future development. Our dataset will be made publicly available at https://huggingface.co/datasets/declare-lab/GSM8k_MORE.

Machine translation often suffers from biased data and algorithms that can lead to unacceptable errors in system output. While bias in gender norms has been investigated, less is known about whether MT systems encode bias about social relationships, e.g. sentences such as "the lawyer kissed her wife." We investigate the degree of bias against same-gender relationships in MT systems, using generated template sentences drawn from several noun-gender languages (e.g. Spanish). We find that three popular MT services consistently fail to accurately translate sentences concerning relationships between nouns of the same gender. The error rate varies considerably based on the context, e.g. same-gender sentences referencing high female-representation occupations are translated with lower accuracy. We provide this work as a case study in the evaluation of intrinsic bias in NLP systems, with respect to social relationships.

While alignment algorithms are now commonly used to tune pre-trained language models towards a user's preferences, we lack explanations for the underlying mechanisms in which models become ``aligned'', thus making it difficult to explain phenomena like jailbreaks. In this work we study a popular algorithm, direct preference optimization (DPO), and the mechanisms by which it reduces toxicity. Namely, we first study how toxicity is represented and elicited in a pre-trained language model, GPT2-medium. We then apply DPO with a carefully crafted pairwise dataset to reduce toxicity. We examine how the resulting model averts toxic outputs, and find that capabilities learned from pre-training are not removed, but rather bypassed. We use this insight to demonstrate a simple method to un-align the model, reverting it back to its toxic behavior.

Causal inference is one of the hallmarks of human intelligence. While the field of CausalNLP has attracted much interest in the recent years, existing causal inference datasets in NLP primarily rely on discovering causality from empirical knowledge (e.g., commonsense knowledge). In this work, we propose the first benchmark dataset to test the pure causal inference skills of large language models (LLMs). Specifically, we formulate a novel task Corr2Cause, which takes a set of correlational statements and determines the causal relationship between the variables. We curate a large-scale dataset of more than 200K samples, on which we evaluate seventeen existing LLMs. Through our experiments, we identify a key shortcoming of LLMs in terms of their causal inference skills, and show that these models achieve almost close to random performance on the task. This shortcoming is somewhat mitigated when we try to re-purpose LLMs for this skill via finetuning, but we find that these models still fail to generalize -- they can only perform causal inference in in-distribution settings when variable names and textual expressions used in the queries are similar to those in the training set, but fail in out-of-distribution settings generated by perturbing these queries. Corr2Cause is a challenging task for LLMs, and would be helpful in guiding future research on improving LLMs' pure reasoning skills and generalizability. Our data is at https://huggingface.co/datasets/causalnlp/corr2cause. Our code is at https://github.com/causalNLP/corr2cause.

Deception detection is gaining increasing interest due to ethical and security concerns. This paper explores the application of convolutional neural networks for the purpose of multimodal deception detection. We use a dataset built by interviewing 104 subjects about two topics, with one truthful and one falsified response from each subject about each topic. In particular, we make three main contributions. First, we extract linguistic and physiological features from this data to train and construct the neural network models. Second, we propose a fused convolutional neural network model using both modalities in order to achieve an improved overall performance. Third, we compare our new approach with earlier methods designed for multimodal deception detection. We find that our system outperforms regular classification methods; our results indicate the feasibility of using neural networks for deception detection even in the presence of limited amounts of data.

We propose task-adaptive tokenization as a way to adapt the generation pipeline to the specifics of a downstream task and enhance long-form generation in mental health. Inspired by insights from cognitive science, our task-adaptive tokenizer samples variable segmentations from multiple outcomes, with sampling probabilities optimized based on task-specific data. We introduce a strategy for building a specialized vocabulary and introduce a vocabulary merging protocol that allows for the integration of task-specific tokens into the pre-trained model's tokenization step. Through extensive experiments on psychological question-answering tasks in both Chinese and English, we find that our task-adaptive tokenization approach brings a significant improvement in generation performance while using up to 60% fewer tokens. Preliminary experiments point to promising results when using our tokenization approach with very large language models.

Despite the impressive performance of current AI models reported across various tasks, performance reports often do not include evaluations of how these models perform on the specific groups that will be impacted by these technologies. Among the minority groups under-represented in AI, data from low-income households are often overlooked in data collection and model evaluation. We evaluate the performance of a state-of-the-art vision-language model (CLIP) on a geo-diverse dataset containing household images associated with different income values (Dollar Street) and show that performance inequality exists among households of different income levels. Our results indicate that performance for the poorer groups is consistently lower than the wealthier groups across various topics and countries. We highlight insights that can help mitigate these issues and propose actionable steps for economic-level inclusive AI development. Code is available at https://github.com/MichiganNLP/Bridging_the_Digital_Divide.