Strong inductive biases enable learning from little data and help generalization outside the training distribution. Popular neural architectures such as Transformers lack strong structural inductive biases for seq2seq NLP tasks on their own. Consequently, they struggle with systematic generalization beyond the training distribution, e.g. with extrapolating to longer inputs, even when pre-trained on large amounts of text.We show how a structural inductive bias can be efficiently injected into a seq2seq model by pre-training it to simulate structural transformations on synthetic data. Specifically, we inject an inductive bias towards Finite State Transducers (FSTs) into a Transformer by pre-training it to simulate FSTs given their descriptions. Our experiments show that our method imparts the desired inductive bias, resulting in improved systematic generalization and better few-shot learning for FST-like tasks. Our analysis shows that fine-tuned models accurately capture the state dynamics of the unseen underlying FSTs, suggesting that the simulation process is internalized by the fine-tuned model.
论文检索
输入标题、作者或关键词，从 1,915 篇学术成果中精准定位
Counterfactual reasoning in narratives requires predicting how alternative conditions, contrary to what actually happened, might have resulted in different outcomes.One major challenge is to maintain the causality between the counterfactual condition and the generated counterfactual outcome. In this paper, we propose a basic VAE module for counterfactual reasoning in narratives. We further introduce a pre-trained classifier and external event commonsense to mitigate the posterior collapse problem in the VAE approach, and improve the causality between the counterfactual condition and the generated counterfactual outcome. We evaluate our method on two public benchmarks. Experiments show that our method is effective.
Generating Contrastive Narratives Using the Brownian Bridge Process for Narrative Coherence Learning
PDF ↗A major challenge for narrative reasoning is to learn narrative coherence. Existing works mainly follow the contrastive learning paradigm. However, the negative samples in their methods can be easily distinguished, which makes their methods unsatisfactory. In this work, we devise two strategies for mining hard negatives, including (1) crisscrossing a narrative and its contrastive variants; and (2) event-level replacement. To obtain contrastive variants, we utilize the Brownian Bridge process to guarantee the quality of generated contrastive narratives. We evaluate our model on several tasks. The result proves the effectiveness of our method, and shows that our method is applicable to many applications.
Humans generally acquire new skills without compromising the old; however, the opposite holds for Large Language Models (LLMs), e.g., from LLaMA to CodeLLaMA. To this end, we propose a new post-pretraining method for LLMs with an expansion of Transformer blocks. We tune the expanded blocks using only new corpus, efficiently and effectively improving the model’s knowledge while mitigating forgetting. In this paper, we experiment on the corpus of code and math, yielding LLaMA Pro-8.3B, a versatile foundation model initialized from LLaMA2-7B, excelling in general tasks, programming, and mathematics. LLaMA Pro and its instruction-following counterpart (LLaMA Pro - Instruct) achieve advanced performance among various benchmarks, demonstrating superiority over existing open models in the LLaMA family and the immense potential of reasoning and addressing diverse tasks as an intelligent agent. Our findings provide valuable insights into integrating natural and programming languages, laying a solid foundation for developing advanced language agents that operate effectively in various environments.
Benchmarking and Improving Compositional Generalization of Multi-aspect Controllable Text Generation
PDF ↗Compositional generalization, representing the model’s ability to generate text with new attribute combinations obtained by recombining single attributes from the training data, is a crucial property for multi-aspect controllable text generation (MCTG) methods. Nonetheless, a comprehensive compositional generalization evaluation benchmark of MCTG is still lacking. We propose CompMCTG, a benchmark encompassing diverse multi-aspect labeled datasets and a crafted three-dimensional evaluation protocol, to holistically evaluate the compositional generalization of MCTG approaches. We observe that existing MCTG works generally confront a noticeable performance drop in compositional testing. To mitigate this issue, we introduce Meta-MCTG, a training framework incorporating meta-learning, where we enable models to learn how to generalize by simulating compositional generalization scenarios in the training phase. We demonstrate the effectiveness of Meta-MCTG through achieving obvious improvement (by at most 3.64%) for compositional testing performance in 94.4%.
Recent years have seen increasing attention on Legal Case Retrieval (LCR), a key task in the area of Legal AI that concerns the retrieval of cases from a large legal database of historical cases that are similar to a given query. This paper presents a survey of the major milestones made in LCR research, targeting researchers who are finding their way into the field and seek a brief account of the relevant datasets and the recent neural models and their performances.
DiFiNet: Boundary-Aware Semantic Differentiation and Filtration Network for Nested Named Entity Recognition
PDF ↗Nested Named Entity Recognition (Nested NER) entails identifying and classifying entity spans within the text, including the detection of named entities that are embedded within external entities. Prior approaches primarily employ span-based techniques, utilizing the power of exhaustive searches to address the challenge of overlapping entities. Nonetheless, these methods often grapple with the absence of explicit guidance for boundary detection, resulting insensitivity in discerning minor variations within nested spans. To this end, we propose a Boundary-aware Semantic \underline{Di}fferentiation and \underline{Fi}ltration \underline{Net}work (DiFiNet) tailored for nested NER. Specifically, DiFiNet leverages a biaffine attention mechanism to generate a span representation matrix. This matrix undergoes further refinement through a self-adaptive semantic differentiation module, specifically engineered to discern semantic variances across spans. Furthermore, DiFiNet integrates a boundary filtration module, designed to mitigate the impact of non-entity noise by leveraging semantic relations among spans. Extensive experiments on three benchmark datasets demonstrate our model yields a new state-of-the-art performance.
One of the central skills that language learners need to practice is speaking the language. Currently, students in school do not get enough speaking opportunities and lack conversational practice. The recent advances in speech technology and natural language processing allow the creation of novel tools to practice their speaking skills. In this work, we tackle the first component of such a pipeline, namely, the automated speech recognition module (ASR). State-of-the-art models are often trained on adult read-aloud data by native speakers and do not transfer well to young language learners’ speech. Second, most ASR systems contain a powerful language model, which smooths out mistakes made by the speakers. To give corrective feedback, which is a crucial part of language learning, the ASR systems in our setting need to preserve the mistakes made by the language learners. In this work, we build an ASR system that satisfies these requirements: it works on spontaneous speech by young language learners and preserves their mistakes. For this, we collected a corpus containing around 85 hours of English audio spoken by Swiss learners from grades 4 to 6 on different language learning tasks, which we used to train an ASR model. Our experiments show that our model benefits from direct fine-tuning of children’s voices and has a much higher error preservation rate.
As the scale of Large Language Models (LLMs) increases, it is necessary to compress the models to reduce the substantial demand on computational resources. Network pruning significantly reduces the model size by converting the weight matrix from dense to sparse data format. Current methodologies advocate for one-shot pruning to avoid the expense of retraining, ensuring the maintenance of model performance under conditions of 50%-60% unstructured pruning. Nevertheless, matrices characterized by this level of sparsity could not be treated as sparse matrices, because the indices would incur significant costs. To mitigate this problem, NVIDIA introduced the 2:4 structured sparsity. However, we observe a notable decline in model performance when adopting 2:4 structured sparsity due to group constraints. In this paper, we introduce the Weight Recover Prune (WRP) approach. By recovering a minimal set of critical weights, WRP aims to enhance model performance while maintaining the efficiency of the compression. Our evaluation of the WRP method on the LLAMA2 and OPT models shows that it outperforms other 2:4 pattern one-shot pruning methods. Meanwhile, WRP can guarantee that the size of the pruned model is about 60% of the dense model. Our code is available at: https://github.com/TanZhendong/WRP.
To improve the performance of large language models (LLMs), researchers have explored providing LLMs with textual task-solving experience via prompts. However, they rely on manual efforts to acquire and apply such experience for each task, which is not feasible for the growing demand for LLMs and the variety of user questions.To address this issue, we design a lifelong autonomous experiential learning framework based on LLMs to explore whether LLMs can imitate human ability for learning and utilizing experience. It autonomously learns and accumulates experience through experience transfer and induction, categorizing the types of input questions to select which accumulated experience to employ for them.Experimental results on six widely used NLP datasets show that our framework performs reliably in each intermediate step and effectively improves the performance of GPT-3.5 and GPT-4. This validates the feasibility of using LLMs to mimic human experiential learning and application capabilities, offering a new path worth further exploration for the evolution of machine intelligence. Additionally, we provide a detailed analysis of the behavior of our framework at each step.We will open source codes after the acceptance, fostering open research in the NLP community and beyond.
This paper identifies a cultural dominance issue within large language models (LLMs) due to the predominant use of English data in model training (e.g., ChatGPT). LLMs often provide inappropriate English-culture-related answers that are not relevant to the expected culture when users ask in non-English languages. To systematically evaluate the cultural dominance issue, we build a benchmark of concrete (e.g., holidays and songs) and abstract (e.g., values and opinions) cultural objects. Empirical results show that the representative GPT models suffer from the culture dominance problem, where GPT-4 is the most affected while text-davinci-003 suffers the least from this problem. Our study emphasizes the need to critically examine cultural dominance and ethical considerations in their development and deployment. We show that two straightforward methods in model development (i.e., pretraining on more diverse data) and deployment (e.g., culture-aware prompting) can significantly mitigate the cultural dominance issue in LLMs.
Despite their remarkable ability to capture linguistic nuances across diverse languages, questions persist regarding the degree of alignment between languages in multilingual embeddings. Drawing inspiration from research on high-dimensional representations in neural language models, we employ clustering to uncover latent concepts within multilingual models. Our analysis focuses on quantifying the alignment and overlap of these concepts across various languages within the latent space. To this end, we introduce two metrics CALIGN and COLAP aimed at quantifying these aspects, enabling a deeper exploration of multilingual embeddings. Our study encompasses three multilingual models (mT5, mBERT, and XLM-R) and three downstream tasks (Machine Translation, Named Entity Recognition, and Sentiment Analysis). Key findings from our analysis include: i) deeper layers in the network demonstrate increased cross-lingual alignment due to the presence of language-agnostic concepts, ii) fine-tuning of the models enhances alignment within the latent space, and iii) such task-specific calibration helps in explaining the emergence of zero-shot capabilities in the models.
Identifying the training datasets that influence a language model’s outputs is essential for minimizing the generation of harmful content and enhancing its performance. Ideally, we can measure the influence of each dataset by removing it from training; however, it is prohibitively expensive to retrain a model multiple times. This paper presents UnTrac: unlearning traces the influence of a training dataset on the model’s performance. UnTrac is extremely simple; each training dataset is unlearned by gradient ascent, and we evaluate how much the model’s predictions change after unlearning. Furthermore, we propose a more scalable approach, UnTrac-Inv, which unlearns a test dataset and evaluates the unlearned model on training datasets. UnTrac-Inv resembles UnTrac, while being efficient for massive training datasets. In the experiments, we examine if our methods can assess the influence of pretraining datasets on generating toxic, biased, and untruthful content. Our methods estimate their influence much more accurately than existing methods while requiring neither excessive memory space nor multiple checkpoints.
With over 60M articles, Wikipedia has become the largest platform for open and freely accessible knowledge. While it has more than 15B monthly visits, its content is believed to be inaccessible to many readers due to the lack of readability of its text. However, previous investigations of the readability of Wikipedia have been restricted to English only, and there are currently no systems supporting the automatic readability assessment of the 300+ languages in Wikipedia. To bridge this gap, we develop a multilingual model to score the readability of Wikipedia articles. To train and evaluate this model, we create a novel multilingual dataset spanning 14 languages, by matching articles from Wikipedia to simplified Wikipedia and online children encyclopedias. We show that our model performs well in a zero-shot scenario, yielding a ranking accuracy of more than 80% across 14 languages and improving upon previous benchmarks. These results demonstrate the applicability of the model at scale for languages in which there is no ground-truth data available for model fine-tuning. Furthermore, we provide the first overview on the state of readability in Wikipedia beyond English.
Discursive Socratic Questioning: Evaluating the Faithfulness of Language Models’ Understanding of Discourse Relations
PDF ↗While large language models have significantly enhanced the effectiveness of discourse relation classifications, it remains unclear whether their comprehension is faithful and reliable. We provide DiSQ, a new method for evaluating the faithfulness of understanding discourse based on question answering. We first employ in-context learning to annotate the reasoning for discourse comprehension, based on the connections among key events within the discourse. Following this, DiSQ interrogates the model with a sequence of questions to assess its grasp of core event relations, its resilience to counterfactual queries, as well as its consistency to its previous responses. then evaluate language models with different architectural designs using DiSQ, finding: (1) DiSQ presents a significant challenge for all models, with the top-performing GPT model attaining only 41% of the ideal performance in PDTB; (2) DiSQ is robust to domain shifts and paraphrase variations; (3) Open-source models generally lag behind their closed-source GPT counterparts, with notable exceptions being those enhanced with chat and code/math features; (4) Our analysis validates the effectiveness of explicitly signalled discourse connectives, the role of contextual information, and the benefits of using historical QA data.
FastFiD: Improve Inference Efficiency of Open Domain Question Answering via Sentence Selection
PDF ↗Open Domain Question Answering (ODQA) has been advancing rapidly in recent times, driven by significant developments in dense passage retrieval and pretrained language models. State-of-the-art models typically incorporate the FiD framework, which is composed by a neural retriever alongside an encoder-decoder neural reader. In the answer generation process, the retriever will retrieve numerous passages (around 100 for instance), each of which is then individually encoded by the encoder. Subsequently, the decoder makes predictions based on these encoded passages. Nevertheless, this framework can be relatively time-consuming, particularly due to the extensive length of the gathered passages. To address this, we introduce FastFiD in this paper, a novel approach that executes sentence selection on the encoded passages. This aids in retaining valuable sentences while reducing the context length required for generating answers. Experiments on three commonly used datasets (Natural Questions, TriviaQA and ASQA) demonstrate that our method can enhance the inference speed by **2.3X-5.7X**, while simultaneously maintaining the model’s performance. Moreover, an in-depth analysis of the model’s attention reveals that the selected sentences indeed hold a substantial contribution towards the final answer. The codes are publicly available at https://github.com/thunlp/FastFiD.
A song is a combination of singing voice and accompaniment. However, existing works focus on singing voice synthesis and music generation independently. Little attention was paid to exploring song synthesis. In this work, we propose a novel task called Text-to-Song synthesis which incorporates both vocal and accompaniment generation. We develop Melodist, a two-stage text-to-song method that consists of singing voice synthesis (SVS) and vocal-to-accompaniment (V2A) synthesis. Melodist leverages tri-tower contrastive pretraining to learn more effective text representation for controllable V2A synthesis. A Chinese song dataset mined from a music website is built to alleviate data scarcity for our research. The evaluation results on our dataset demonstrate that Melodist can synthesize songs with comparable quality and style consistency. Audio samples can be found in https://text2songMelodist.github.io/Sample/.
Large Language Models (LLMs) have demonstrated considerable cross-lingual alignment and generalization ability. Current research primarily focuses on improving LLMs’ cross-lingual generalization capabilities. However, there is still a lack of research on the intrinsic mechanisms of how LLMs achieve cross-lingual alignment. From the perspective of region partitioning, this paper conducts several investigations on the linguistic competence of LLMs. We discover a core region in LLMs that corresponds to linguistic competence, accounting for approximately 1% of the total model parameters. Removing this core region by setting parameters to zero results in a significant performance decrease across 30 different languages. Furthermore, this core region exhibits significant dimensional dependence, perturbations to even a single parameter on specific dimensions leading to a loss of linguistic competence. Moreover, we discover that distinct monolingual regions exist for different languages, and disruption to these specific regions substantially reduces the LLMs’ proficiency in those corresponding languages. Our research also indicates that freezing the core linguistic region during further pre-training can mitigate the issue of catastrophic forgetting (CF), a common phenomenon observed during further pre-training of LLMs. Overall, exploring the LLMs’ functional regions provides insights into the foundation of their intelligence.
Blinded by Generated Contexts: How Language Models Merge Generated and Retrieved Contexts When Knowledge Conflicts?
PDF ↗While auxiliary information has become a key to enhancing Large Language Models (LLMs), relatively little is known about how LLMs merge these contexts, specifically contexts generated by LLMs and those retrieved from external sources.To investigate this, we formulate a systematic framework to identify whether LLMs’ responses are attributed to either generated or retrieved contexts.To easily trace the origin of the response, we construct datasets with conflicting contexts, i.e., each question is paired with both generated and retrieved contexts, yet only one of them contains the correct answer.Our experiments reveal a significant bias in several LLMs (GPT-4/3.5 and Llama2) to favor generated contexts, even when they provide incorrect information.We further identify two key factors contributing to this bias: i) contexts generated by LLMs typically show greater similarity to the questions, increasing their likelihood of being selected; ii) the segmentation process used in retrieved contexts disrupts their completeness, thereby hindering their full utilization in LLMs.Our analysis enhances the understanding of how LLMs merge diverse contexts, offers valuable insights for advancing current LLM augmentation methods, and highlights the risk of generated misinformation for retrieval-augmented LLMs.
Fine-tuning large language models (LLMs) for machine translation has shown improvements in overall translation quality. However, it is unclear what is the impact of fine-tuning on desirable LLM behaviors that are not present in neural machine translation models, such as steerability, inherent document-level translation abilities, and the ability to produce less literal translations. We perform an extensive translation evaluation on the LLaMA and Falcon family of models with model size ranging from 7 billion up to 65 billion parameters.Our results show that while fine-tuning improves the general translation quality of LLMs, several abilities degrade. In particular, we observe a decline in the ability to perform formality steering, to produce technical translations through few-shot examples, and to perform document-level translation. On the other hand, we observe that the model produces less literal translations after fine-tuning on parallel data. We show that by including monolingual data as part of the fine-tuning data we can maintain the abilities while simultaneously enhancing overall translation quality. Our findings emphasize the need for fine-tuning strategies that preserve the benefits of LLMs for machine translation.