Reinforcement learning (RL) can align language models with non-differentiable reward signals, such as human preferences. However, a major challenge arises from the sparsity of these reward signals - typically, there is only a single reward for an entire output. This sparsity of rewards can lead to inefficient and unstable learning. To address this challenge, our paper introduces an novel framework that utilizes the critique capability of Large Language Models (LLMs) to produce intermediate-step rewards during RL training. Our method involves coupling a policy model with a critic language model, which is responsible for providing comprehensive feedback of each part of the output. This feedback is then translated into token or span-level rewards that can be used to guide the RL training process. We investigate this approach under two different settings: one where the policy model is smaller and is paired with a more powerful critic model, and another where a single language model fulfills both roles. We assess our approach on three text generation tasks: sentiment control, language model detoxification, and summarization. Experimental results show that incorporating artificial intrinsic rewards significantly improve both sample efficiency and the overall performance of the policy model, supported by both automatic and human evaluation.

State-of-the-art language models (LMs) sometimes generate that misalign with world knowledge. To explore the mechanistic causes of these hallucinations, we create diagnostic datasets with subject-relation queries and adapt interpretability methods to trace hallucinations through internal model representations. We discover two general and distinct mechanistic causes of hallucinations shared across LMs (Llama-2, Pythia, GPT-J): 1) : insufficient subject attribute knowledge in lower layer MLPs, and 2) : failure to select the correct object attribute in upper layer attention heads. We also found these two internal mechanistic causes of hallucinations are reflected in external manifestations. Based on insights from our mechanistic analysis, we propose a novel hallucination mitigation method through targeted restoration of the LM’s internal fact recall pipeline, demonstrating superior performance compared to baselines.

This study presents RoleCraft-GLM, an innovative framework aimed at enhancing personalized role-playing with Large Language Models (LLMs). RoleCraft-GLM addresses the key issue of lacking personalized interactions in conversational AI, and offers a solution with detailed and emotionally nuanced character portrayals. We contribute a unique conversational dataset that shifts from conventional celebrity-centric characters to diverse, non-celebrity personas, thus enhancing the realism and complexity of language modeling interactions. Additionally, our approach includes meticulous character development, ensuring dialogues are both realistic and emotionally resonant. The effectiveness of RoleCraft-GLM is validated through various case studies, highlighting its versatility and skill in different scenarios. Our framework excels in generating dialogues that accurately reflect characters’ personality traits and emotions, thereby boosting user engagement. In conclusion, RoleCraft-GLM marks a significant leap in personalized AI interactions, and paves the way for more authentic and immersive AI-assisted role-playing experiences by enabling more nuanced and emotionally rich dialogues.

Humans often make creative use of words to expressnovel senses. A long-standing effort in natural language processing hasbeen focusing on word sense disambiguation (WSD), but little has been explored about how the sense inventory of a word may be extended toward novel meanings. We present a paradigm of word sense extension (WSE) thatenables words to spawn new senses toward novel context. We develop a framework that simulates novel word sense extension by first partitioning a polysemous word type into two pseudo-tokens that mark its different senses, and then inferring whether the meaning of a pseudo-token can be extended to convey the sense denoted by the token partitioned from the same word type. Our framework combines cognitivemodels of chaining with a learning scheme that transforms a language model embedding space to supportvarious types of word sense extension. We evaluate our frameworkagainst several competitive baselines and show that it is superior in predicting plausible novel senses for over 7,500 English words. Furthermore, we show that our WSE framework improves performance over a range of transformer-based WSD models in predicting rare word senses with few or zero mentions in the training data.

After just a few hundred training updates, a standard probabilistic model for language generation has likely not yet learnt many semantic or syntactic rules of natural language, making it difficult to estimate the probability distribution over next tokens. Yet around this point, these models have identified a simple, loss-minimising behaviour: to output the unigram distribution of the target training corpus. The use of such a heuristic raises the question: Can we initialise our models with this behaviour and save precious compute resources and model capacity? Here we show that we can effectively endow standard neural language generation models with a separate module that reflects unigram frequency statistics as prior knowledge, simply by initialising the bias term in a model’s final linear layer with the log-unigram distribution. We use neural machine translation as a test bed for this simple technique and observe that it: (i) improves learning efficiency; (ii) achieves better overall performance; and perhaps most importantly (iii) appears to disentangle strong frequency effects by encouraging the model to specialise in non-frequency-related aspects of language.

The meaning of polysemous words often varies in a highly productive yet predictable way. Generalizing the regularity between conventional senses to derive novel word meaning is crucial for automated processing of non-literal language uses such as figurative expressions. We introduce a novel task called systematic word meta-sense extension (SWORME) to test and improve language models’ ability to extend word meaning to denote new semantic domains (also called meta-senses) that bear regular semantic relations with existing senses. We found that language models prefer incremental lexical semantic change toward conceptually similar meta-senses such as logical metonymy, and are much worse at predicting highly non-literal meaning extensions such as metaphors. We propose a novel analogy-based method of word meaning extension, and show that it effectively improves language model systematicity in making both gradual and radical types of meta-sense extension. We further demonstrate that learning systematic meta-sense extensions benefits language models on multiple benchmarks of figurative language understanding.

Multi-label clinical text classification, such as automatic ICD coding, has always been a challenging subject in Natural Language Processing, due to its long, domain-specific documents and long-tail distribution over a large label set. Existing methods adopt different model architectures to encode the clinical notes. Whereas without digging out the useful connections between labels, the model presents a huge gap in predicting performances between rare and frequent codes. In this work, we propose a novel method for further mining the helpful relations between different codes via a relation-enhanced code encoder to improve the rare code performance. Starting from the simple code descriptions, the model reaches comparable, even better performances than models with heavy external knowledge. Our proposed method is evaluated on MIMIC-III, a common dataset in the medical domain. It outperforms the previous state-of-art models on both overall metrics and rare code performances. Moreover, the interpretation results further prove the effectiveness of our methods. Our code is publicly available at https://github.com/jiaminchen-1031/Rare-ICD.

Few-shot Text Classification predicts the semantic label of a given text with a handful of supporting instances. Current meta-learning methods have achieved satisfying results in various few-shot situations. Still, they often require a large amount of data to construct many few-shot tasks for meta-training, which is not practical in real-world few-shot scenarios. Prompt-tuning has recently proved to be another effective few-shot learner by bridging the gap between pre-train and downstream tasks. In this work, we closely combine the two promising few-shot learning methodologies in structure and propose a Prompt-Based Meta-Learning (PBML) model to overcome the above meta-learning problem by adding the prompting mechanism. PBML assigns label word learning to base-learners and template learning to meta-learner, respectively. Experimental results show state-of-the-art performance on four text classification datasets under few-shot settings, with higher accuracy and good robustness. We demonstrate through low-resource experiments that our method alleviates the shortcoming that meta-learning requires too much data for meta-training. In the end, we use the visualization to interpret and verify that the meta-learning framework can help the prompting method converge better. We release our code to reproduce our experiments.

Humans can flexibly extend word usages across different grammatical classes, a phenomenon known as word class conversion. Noun-to-verb conversion, or denominal verb (e.g., to Google a cheap flight), is one of the most prevalent forms of word class conversion. However, existing natural language processing systems are impoverished in interpreting and generating novel denominal verb usages. Previous work has suggested that novel denominal verb usages are comprehensible if the listener can compute the intended meaning based on shared knowledge with the speaker. Here we explore a computational formalism for this proposal couched in frame semantics. We present a formal framework, Noun2Verb, that simulates the production and comprehension of novel denominal verb usages by modeling shared knowledge of speaker and listener in semantic frames. We evaluate an incremental set of probabilistic models that learn to interpret and generate novel denominal verb usages via paraphrasing. We show that a model where the speaker and listener cooperatively learn the joint distribution over semantic frame elements better explains the empirical denominal verb usages than state-of-the-art language models, evaluated against data from (1) contemporary English in both adult and child speech, (2) contemporary Mandarin Chinese, and (3) the historical development of English. Our work grounds word class conversion in probabilistic frame semantics and bridges the gap between natural language processing systems and humans in lexical creativity.

We propose a new architecture for adapting a sentence-level sequence-to-sequence transformer by incorporating multiple pre-trained document context signals and assess the impact on translation performance of (1) different pretraining approaches for generating these signals, (2) the quantity of parallel data for which document context is available, and (3) conditioning on source, target, or source and target contexts. Experiments on the NIST Chinese-English, and IWSLT and WMT English-German tasks support four general conclusions: that using pre-trained context representations markedly improves sample efficiency, that adequate parallel data resources are crucial for learning to use document context, that jointly conditioning on multiple context representations outperforms any single representation, and that source context is more valuable for translation performance than target side context. Our best multi-context model consistently outperforms the best existing context-aware transformers.

Direct decoding for task-oriented dialogue is known to suffer from the explaining-away effect, manifested in models that prefer short and generic responses. Here we argue for the use of Bayes’ theorem to factorize the dialogue task into two models, the distribution of the context given the response, and the prior for the response itself. This approach, an instantiation of the noisy channel model, both mitigates the explaining-away effect and allows the principled incorporation of large pretrained models for the response prior. We present extensive experiments showing that a noisy channel model decodes better responses compared to direct decoding and that a two-stage pretraining strategy, employing both open-domain and task-oriented dialogue data, improves over randomly initialized models.

Neural machine translation (NMT) has arguably achieved human level parity when trained and evaluated at the sentence-level. Document-level neural machine translation has received less attention and lags behind its sentence-level counterpart. The majority of the proposed document-level approaches investigate ways of conditioning the model on several source or target sentences to capture document context. These approaches require training a specialized NMT model from scratch on parallel document-level corpora. We propose an approach that doesn’t require training a specialized model on parallel document-level corpora and is applied to a trained sentence-level NMT model at decoding time. We process the document from left to right multiple times and self-train the sentence-level model on pairs of source sentences and generated translations. Our approach reinforces the choices made by the model, thus making it more likely that the same choices will be made in other sentences in the document. We evaluate our approach on three document-level datasets: NIST Chinese-English, WMT19 Chinese-English and OpenSubtitles English-Russian. We demonstrate that our approach has higher BLEU score and higher human preference than the baseline. Qualitative analysis of our approach shows that choices made by model are consistent across the document.

Natural language relies on a finite lexicon to express an unbounded set of emerging ideas. One result of this tension is the formation of new compositions, such that existing linguistic units can be combined with emerging items into novel expressions. We develop a framework that exploits the cognitive mechanisms of chaining and multimodal knowledge to predict emergent compositional expressions through time. We present the syntactic frame extension model (SFEM) that draws on the theory of chaining and knowledge from “percept”, “concept”, and “language” to infer how verbs extend their frames to form new compositions with existing and novel nouns. We evaluate SFEM rigorously on the 1) modalities of knowledge and 2) categorization models of chaining, in a syntactically parsed English corpus over the past 150 years. We show that multimodal SFEM predicts newly emerged verb syntax and arguments substantially better than competing models using purely linguistic or unimodal knowledge. We find support for an exemplar view of chaining as opposed to a prototype view and reveal how the joint approach of multimodal chaining may be fundamental to the creation of literal and figurative language uses including metaphor and metonymy.

This paper describes the DeepMind submission to the Chinese→English constrained data track of the WMT2020 Shared Task on News Translation. The submission employs a noisy channel factorization as the backbone of a document translation system. This approach allows the flexible combination of a number of independent component models which are further augmented with back-translation, distillation, fine-tuning with in-domain data, Monte-Carlo Tree Search decoding, and improved uncertainty estimation. In order to address persistent issues with the premature truncation of long sequences we included specialized length models and sentence segmentation techniques. Our final system provides a 9.9 BLEU points improvement over a baseline Transformer on our test set (newstest 2019).

We show that Bayes’ rule provides an effective mechanism for creating document translation models that can be learned from only parallel sentences and monolingual documents a compelling benefit because parallel documents are not always available. In our formulation, the posterior probability of a candidate translation is the product of the unconditional (prior) probability of the candidate output document and the “reverse translation probability” of translating the candidate output back into the source language. Our proposed model uses a powerful autoregressive language model as the prior on target language documents, but it assumes that each sentence is translated independently from the target to the source language. Crucially, at test time, when a source document is observed, the document language model prior induces dependencies between the translations of the source sentences in the posterior. The model’s independence assumption not only enables efficient use of available data, but it additionally admits a practical left-to-right beam-search algorithm for carrying out inference. Experiments show that our model benefits from using cross-sentence context in the language model, and it outperforms existing document translation approaches.

We present a methodological framework for inferring symmetry of verb predicates in natural language. Empirical work on predicate symmetry has taken two main approaches. The feature-based approach focuses on linguistic features pertaining to symmetry. The context-based approach denies the existence of absolute symmetry but instead argues that such inference is context dependent. We develop methods that formalize these approaches and evaluate them against a novel symmetry inference sentence (SIS) dataset comprised of 400 naturalistic usages of literature-informed verbs spanning the spectrum of symmetry-asymmetry. Our results show that a hybrid transfer learning model that integrates linguistic features with contextualized language models most faithfully predicts the empirical data. Our work integrates existing approaches to symmetry in natural language and suggests how symmetry inference can improve systematicity in state-of-the-art language models.

Recurrent neural network grammars (RNNG) are generative models of language which jointly model syntax and surface structure by incrementally generating a syntax tree and sentence in a top-down, left-to-right order. Supervised RNNGs achieve strong language modeling and parsing performance, but require an annotated corpus of parse trees. In this work, we experiment with unsupervised learning of RNNGs. Since directly marginalizing over the space of latent trees is intractable, we instead apply amortized variational inference. To maximize the evidence lower bound, we develop an inference network parameterized as a neural CRF constituency parser. On language modeling, unsupervised RNNGs perform as well their supervised counterparts on benchmarks in English and Chinese. On constituency grammar induction, they are competitive with recent neural language models that induce tree structures from words through attention mechanisms.