Given the broad capabilities of large language models, it should be possible to work towards a general-purpose, text-based assistant that is aligned with human values, meaning that it is helpful, honest, and harmless. As an initial foray in this direction we study simple baseline techniques and evaluations, such as prompting. We find that the benefits from modest interventions increase with model size, generalize to a variety of alignment evaluations, and do not compromise the performance of large models. Next we investigate scaling trends for several training objectives relevant to alignment, comparing imitation learning, binary discrimination, and ranked preference modeling. We find that ranked preference modeling performs much better than imitation learning, and often scales more favorably with model size. In contrast, binary discrimination typically performs and scales very similarly to imitation learning. Finally we study a `preference model pre-training' stage of training, with the goal of improving sample efficiency when finetuning on human preferences.
Positivity is one of the three conditions for causal inference from observational data. The standard way to validate positivity is to analyze the distribution of propensity. However, to democratize the ability to do causal inference by non-experts, it is required to design an algorithm to (i) test positivity and (ii) explain where in the covariate space positivity is lacking. The latter could be used to either suggest the limitation of further causal analysis and/or encourage experimentation where positivity is violated. The contribution of this paper is first present the problem of automatic positivity analysis and secondly to propose an algorithm based on a two steps process. The first step, models the propensity condition on the covariates and then analyze the latter distribution using multiple hypothesis testing to create positivity violation labels. The second step uses asymmetrically pruned decision trees for explainability. The latter is further converted into readable text a non-expert can understand. We demonstrate our method on a proprietary data-set of a large software enterprise.
Natural language processing (NLP) systems have become a central technology in communication, education, medicine, artificial intelligence, and many other domains of research and development. While the performance of NLP methods has grown enormously over the last decade, this progress has been restricted to a minuscule subset of the world's 6,500 languages. We introduce a framework for estimating the global utility of language technologies as revealed in a comprehensive snapshot of recent publications in NLP. Our analyses involve the field at large, but also more in-depth studies on both user-facing technologies (machine translation, language understanding, question answering, text-to-speech synthesis) as well as more linguistic NLP tasks (dependency parsing, morphological inflection). In the process, we (1) quantify disparities in the current state of NLP research, (2) explore some of its associated societal and academic factors, and (3) produce tailored recommendations for evidence-based policy making aimed at promoting more global and equitable language technologies.
Training machines to understand natural language and interact with humans is an elusive and essential task of artificial intelligence. A diversity of dialogue systems has been designed with the rapid development of deep learning techniques, especially the recent pre-trained language models (PrLMs). Among these studies, the fundamental yet challenging type of task is dialogue comprehension whose role is to teach the machines to read and comprehend the dialogue context before responding. In this paper, we review the previous methods from the technical perspective of dialogue modeling for the dialogue comprehension task. We summarize the characteristics and challenges of dialogue comprehension in contrast to plain-text reading comprehension. Then, we discuss three typical patterns of dialogue modeling. In addition, we categorize dialogue-related pre-training techniques which are employed to enhance PrLMs in dialogue scenarios. Finally, we highlight the technical advances in recent years and point out the lessons from the empirical analysis and the prospects towards a new frontier of researches.
We teach goal-driven agents to interactively act and speak in situated environments by training on generated curriculums. Our agents operate in LIGHT (Urbanek et al. 2019) -- a large-scale crowd-sourced fantasy text adventure game wherein an agent perceives and interacts with the world through textual natural language. Goals in this environment take the form of character-based quests, consisting of personas and motivations. We augment LIGHT by learning to procedurally generate additional novel textual worlds and quests to create a curriculum of steadily increasing difficulty for training agents to achieve such goals. In particular, we measure curriculum difficulty in terms of the rarity of the quest in the original training distribution -- an easier environment is one that is more likely to have been found in the unaugmented dataset. An ablation study shows that this method of learning from the tail of a distribution results in significantly higher generalization abilities as measured by zero-shot performance on never-before-seen quests.
This paper looks at solving collaborative planning problems formalized as Decentralized POMDPs (Dec-POMDPs) by searching for Nash equilibria, i.e., situations where each agent's policy is a best response to the other agents' (fixed) policies. While the Joint Equilibrium-based Search for Policies (JESP) algorithm does this in the finite-horizon setting relying on policy trees, we propose here to adapt it to infinite-horizon Dec-POMDPs by using finite state controller (FSC) policy representations. In this article, we (1) explain how to turn a Dec-POMDP with $N-1$ fixed FSCs into an infinite-horizon POMDP whose solution is an $N^\text{th}$ agent best response; (2) propose a JESP variant, called \infJESP, using this to solve infinite-horizon Dec-POMDPs; (3) introduce heuristic initializations for JESP aiming at leading to good solutions; and (4) conduct experiments on state-of-the-art benchmark problems to evaluate our approach.
While recent research on natural language inference has considerably benefited from large annotated datasets, the amount of inference-related knowledge (including commonsense) provided in the annotated data is still rather limited. There have been two lines of approaches that can be used to further address the limitation: (1) unsupervised pretraining can leverage knowledge in much larger unstructured text data; (2) structured (often human-curated) knowledge has started to be considered in neural-network-based models for NLI. An immediate question is whether these two approaches complement each other, or how to develop models that can bring together their advantages. In this paper, we propose models that leverage structured knowledge in different components of pre-trained models. Our results show that the proposed models perform better than previous BERT-based state-of-the-art models. Although our models are proposed for NLI, they can be easily extended to other sentence or sentence-pair classification problems.
Research into the classification of Image with Text (IWT) troll memes has recently become popular. Since the online community utilizes the refuge of memes to express themselves, there is an abundance of data in the form of memes. These memes have the potential to demean, harras, or bully targeted individuals. Moreover, the targeted individual could fall prey to opinion manipulation. To comprehend the use of memes in opinion manipulation, we define three specific domains (product, political or others) which we classify into troll or not-troll, with or without opinion manipulation. To enable this analysis, we enhanced an existing dataset by annotating the data with our defined classes, resulting in a dataset of 8,881 IWT or multimodal memes in the English language (TrollsWithOpinion dataset). We perform baseline experiments on the annotated dataset, and our result shows that existing state-of-the-art techniques could only reach a weighted-average F1-score of 0.37. This shows the need for a development of a specific technique to deal with multimodal troll memes.
Compliments and concerns in reviews are valuable for understanding users' shopping interests and their opinions with respect to specific aspects of certain items. Existing review-based recommenders favor large and complex language encoders that can only learn latent and uninterpretable text representations. They lack explicit user attention and item property modeling, which however could provide valuable information beyond the ability to recommend items. Therefore, we propose a tightly coupled two-stage approach, including an Aspect-Sentiment Pair Extractor (ASPE) and an Attention-Property-aware Rating Estimator (APRE). Unsupervised ASPE mines Aspect-Sentiment pairs (AS-pairs) and APRE predicts ratings using AS-pairs as concrete aspect-level evidence. Extensive experiments on seven real-world Amazon Review Datasets demonstrate that ASPE can effectively extract AS-pairs which enable APRE to deliver superior accuracy over the leading baselines.
\textbf{P}re-\textbf{T}rained \textbf{M}odel\textbf{s} have been widely applied and recently proved vulnerable under backdoor attacks: the released pre-trained weights can be maliciously poisoned with certain triggers. When the triggers are activated, even the fine-tuned model will predict pre-defined labels, causing a security threat. These backdoors generated by the poisoning methods can be erased by changing hyper-parameters during fine-tuning or detected by finding the triggers. In this paper, we propose a stronger weight-poisoning attack method that introduces a layerwise weight poisoning strategy to plant deeper backdoors; we also introduce a combinatorial trigger that cannot be easily detected. The experiments on text classification tasks show that previous defense methods cannot resist our weight-poisoning method, which indicates that our method can be widely applied and may provide hints for future model robustness studies.
Add to Chrome
Add to Firefox