Current large language models can perform reasonably well on complex tasks that require step-by-step reasoning with few-shot learning. Are these models applying reasoning skills they have learnt during pre-training and reason outside of their training context, or are they simply memorizing their training corpus at finer granularity and have learnt to better understand their context? To tease apart these possibilities, we introduce ALERT, a benchmark and suite of analyses for assessing language models' reasoning ability comparing pre-trained and finetuned models on complex tasks that require reasoning skills to solve. ALERT provides a test bed to asses any language model on fine-grained reasoning skills, which spans over 20 datasets and covers 10 different reasoning skills. We leverage ALERT to further investigate the role of finetuning. With extensive empirical analysis we find that language models learn more reasoning skills such as textual entailment, abductive reasoning, and analogical reasoning during finetuning stage compared to pretraining state. We also find that when language models are finetuned they tend to overfit to the prompt template, which hurts the robustness of models causing generalization problems.
Large language models show improved downstream task performance when prompted to generate step-by-step reasoning to justify their final answers. These reasoning steps greatly improve model interpretability and verification, but objectively studying their correctness (independent of the final answer) is difficult without reliable methods for automatic evaluation. We simply do not know how often the stated reasoning steps actually support the final end task predictions. In this work, we present ROSCOE, a suite of interpretable, unsupervised automatic scores that improve and extend previous text generation evaluation metrics. To evaluate ROSCOE against baseline metrics, we design a typology of reasoning errors and collect synthetic and human evaluation scores on commonly used reasoning datasets. In contrast with existing metrics, ROSCOE can measure semantic consistency, logicality, informativeness, fluency, and factuality - among other traits - by leveraging properties of step-by-step rationales. We empirically verify the strength of our metrics on five human annotated and six programmatically perturbed diagnostics datasets - covering a diverse set of tasks that require reasoning skills and show that ROSCOE can consistently outperform baseline metrics.
Large language models (LLMs) have exhibited remarkable capabilities in learning from explanations in prompts. Yet, there has been limited understanding of what makes explanations effective for in-context learning. This work aims to better understand the mechanisms by which explanations are used for in-context learning. We first study the impact of two different factors on prompting performance when using explanations: the computation trace (the way the solution is decomposed) and the natural language of the prompt. By perturbing explanations on three controlled tasks, we show that both factors contribute to the effectiveness of explanations, indicating that LLMs do faithfully follow the explanations to some extent. We further study how to form maximally effective sets of explanations for solving a given test query. We find that LLMs can benefit from the complementarity of the explanation set as they are able to fuse different reasoning specified by individual exemplars in prompts. Additionally, having relevant exemplars also contributes to more effective prompts. Therefore, we propose a maximal-marginal-relevance-based exemplar selection approach for constructing exemplar sets that are both relevant as well as complementary, which successfully improves the in-context learning performance across three real-world tasks on multiple LLMs.
Abstractive summarization models typically generate content unfaithful to the input, thus highlighting the significance of evaluating the faithfulness of generated summaries. Most faithfulness metrics are only evaluated on news domain, can they be transferred to other summarization tasks? In this work, we first present a systematic study of faithfulness metrics for dialogue summarization. We evaluate common faithfulness metrics on dialogue datasets and observe that most metrics correlate poorly with human judgements despite performing well on news datasets. Given these findings, to improve existing metrics' performance on dialogue summarization, we first finetune on in-domain dataset, then apply unlikelihood training on negative samples, and show that they can successfully improve metric performance on dialogue data. Inspired by the strong zero-shot performance of the T0 language model, we further propose T0-Score -- a new metric for faithfulness evaluation, which shows consistent improvement against baseline metrics across multiple domains.
Machine translation has seen rapid progress with the advent of Transformer-based models. These models have no explicit linguistic structure built into them, yet they may still implicitly learn structured relationships by attending to relevant tokens. We hypothesize that this structural learning could be made more robust by explicitly endowing Transformers with a structural bias, and we investigate two methods for building in such a bias. One method, the TP-Transformer, augments the traditional Transformer architecture to include an additional component to represent structure. The second method imbues structure at the data level by segmenting the data with morphological tokenization. We test these methods on translating from English into morphologically rich languages, Turkish and Inuktitut, and consider both automatic metrics and human evaluations. We find that each of these two approaches allows the network to achieve better performance, but this improvement is dependent on the size of the dataset. In sum, structural encoding methods make Transformers more sample-efficient, enabling them to perform better from smaller amounts of data.
Hate speech detection is complex; it relies on commonsense reasoning, knowledge of stereotypes, and an understanding of social nuance that differs from one culture to the next. It is also difficult to collect a large-scale hate speech annotated dataset. In this work, we frame this problem as a few-shot learning task, and show significant gains with decomposing the task into its "constituent" parts. In addition, we see that infusing knowledge from reasoning datasets (e.g. Atomic2020) improves the performance even further. Moreover, we observe that the trained models generalize to out-of-distribution datasets, showing the superiority of task decomposition and knowledge infusion compared to previously used methods. Concretely, our method outperforms the baseline by 17.83% absolute gain in the 16-shot case.
Recently, there has been a surge of interest in the NLP community on the use of pretrained Language Models (LMs) as Knowledge Bases (KBs). Researchers have shown that LMs trained on a sufficiently large (web) corpus will encode a significant amount of knowledge implicitly in its parameters. The resulting LM can be probed for different kinds of knowledge and thus acting as a KB. This has a major advantage over traditional KBs in that this method requires no human supervision. In this paper, we present a set of aspects that we deem a LM should have to fully act as a KB, and review the recent literature with respect to those aspects.
Factual inconsistencies in generated summaries severely limit the practical applications of abstractive dialogue summarization. Although significant progress has been achieved by using pre-trained models, substantial amounts of hallucinated content are found during the human evaluation. Pre-trained models are most commonly fine-tuned with cross-entropy loss for text summarization, which may not be an optimal strategy. In this work, we provide a typology of factual errors with annotation data to highlight the types of errors and move away from a binary understanding of factuality. We further propose a training strategy that improves the factual consistency and overall quality of summaries via a novel contrastive fine-tuning, called ConFiT. Based on our linguistically-informed typology of errors, we design different modular objectives that each target a specific type. Specifically, we utilize hard negative samples with errors to reduce the generation of factual inconsistency. In order to capture the key information between speakers, we also design a dialogue-specific loss. Using human evaluation and automatic faithfulness metrics, we show that our model significantly reduces all kinds of factual errors on the dialogue summarization, SAMSum corpus. Moreover, our model could be generalized to the meeting summarization, AMI corpus, and it produces significantly higher scores than most of the baselines on both datasets regarding word-overlap metrics.
Recent neural models that extend the pretrain-then-finetune paradigm continue to achieve new state-of-the-art results on joint goal accuracy (JGA) for dialogue state tracking (DST) benchmarks. However, we call into question their robustness as they show sharp drops in JGA for conversations containing utterances or dialog flows with realistic perturbations. Inspired by CheckList (Ribeiro et al., 2020), we design a collection of metrics called CheckDST that facilitate comparisons of DST models on comprehensive dimensions of robustness by testing well-known weaknesses with augmented test sets. We evaluate recent DST models with CheckDST and argue that models should be assessed more holistically rather than pursuing state-of-the-art on JGA since a higher JGA does not guarantee better overall robustness. We find that span-based classification models are resilient to unseen named entities but not robust to language variety, whereas those based on autoregressive language models generalize better to language variety but tend to memorize named entities and often hallucinate. Due to their respective weaknesses, neither approach is yet suitable for real-world deployment. We believe CheckDST is a useful guide for future research to develop task-oriented dialogue models that embody the strengths of various methods.
Current efficient fine-tuning methods (e.g., adapters, prefix-tuning, etc.) have optimized conditional text generation via training a small set of extra parameters of the neural language model, while freezing the rest for efficiency. While showing strong performance on some generation tasks, they don't generalize across all generation tasks. In this work, we show that prompt based conditional text generation can be improved with simple and efficient methods that simulate modeling the discourse structure of human written text. We introduce two key design choices: First we show that a higher-level discourse structure of human written text can be modelled with \textit{hierarchical blocking} on prefix parameters that enable spanning different parts of the input and output text and yield more coherent output generations. Second, we propose sparse prefix tuning by introducing \textit{attention sparsity} on the prefix parameters at different layers of the network and learn sparse transformations on the softmax-function, respectively. We find that sparse attention enables the prefix-tuning to better control of the input contents (salient facts) yielding more efficient tuning of the prefix-parameters. Experiments on a wide-variety of text generation tasks show that structured design of prefix parameters can achieve comparable results to fine-tuning all parameters while outperforming standard prefix-tuning on all generation tasks even in low-resource settings.