Wei Wang

Zheng Yuan, Hongyi Yuan, Chuanqi Tan et al. · tsinghua

Large language models have emerged abilities including chain-of-thought to answer math word problems step by step. Solving math word problems not only requires abilities to disassemble problems via chain-of-thought but also needs to calculate arithmetic expressions correctly for each step. To the best of our knowledge, there is no work to focus on evaluating the arithmetic ability of large language models. In this work, we propose an arithmetic dataset MATH 401 to test the latest large language models including GPT-4, ChatGPT, InstrctGPT, Galactica, and LLaMA with various arithmetic expressions and provide a detailed analysis of the ability of large language models. MATH 401 and evaluation codes are released at \url{https://github.com/GanjinZero/math401-llm}.

Ping Yu, Wei Wang, Chunyuan Li et al.

Prompt tuning has been an extremely effective tool to adapt a pre-trained model to downstream tasks. However, standard prompt-based methods mainly consider the case of sufficient data of downstream tasks. It is still unclear whether the advantage can be transferred to the few-shot regime, where only limited data are available for each downstream task. Although some works have demonstrated the potential of prompt-tuning under the few-shot setting, the main stream methods via searching discrete prompts or tuning soft prompts with limited data are still very challenging. Through extensive empirical studies, we find that there is still a gap between prompt tuning and fully fine-tuning for few-shot learning. To bridge the gap, we propose a new prompt-tuning framework, called Soft Template Tuning (STT). STT combines manual and auto prompts, and treats downstream classification tasks as a masked language modeling task. Comprehensive evaluation on different settings suggests STT can close the gap between fine-tuning and prompt-based methods without introducing additional parameters. Significantly, it can even outperform the time- and resource-consuming fine-tuning method on sentiment classification tasks.

Zhijie Deng, Zhouan Shen, Ling Li et al.

Instruction tuning is crucial for aligning Large Language Models (LLMs), yet the quality of instruction-following data varies significantly. While high-quality data is paramount, it is often scarce; conversely, abundant low-quality data is frequently discarded, leading to substantial information loss. Existing data augmentation methods struggle to augment this low-quality data effectively, and the evaluation of such techniques remains poorly defined. To address this, we formally define the task of Instruction Distillation: distilling multiple low-quality and redundant inputs into high-quality and coherent instruction-output pairs. Specifically, we introduce a comprehensive data construction pipeline to create MIXTURE, a 144K-sample dataset pairing low-quality or semantically redundant imperfect instruction clusters with their high-quality distillations. We then introduce LM-Mixup, by first performing supervised fine-tuning on MIXTURE and then optimizing it with reinforcement learning. This process uses three complementary reward signals: quality, semantic alignment, and format compliance, via Group Relative Policy Optimization (GRPO). We demonstrate that LM-Mixup effectively augments imperfect datasets: fine-tuning LLMs on its distilled data, which accounts for only about 3% of the entire dataset, not only surpasses full-dataset training but also competes with state-of-the-art high-quality data selection methods across multiple benchmarks. Our work establishes that low-quality data is a valuable resource when properly distilled and augmented with LM-Mixup, significantly enhancing the efficiency and performance of instruction-tuned LLMs.

Lida Chen, Zujie Liang, Xintao Wang et al.

Large language models (LLMs) have achieved great success, but their occasional content fabrication, or hallucination, limits their practical application. Hallucination arises because LLMs struggle to admit ignorance due to inadequate training on knowledge boundaries. We call it a limitation of LLMs that they can not accurately express their knowledge boundary, answering questions they know while admitting ignorance to questions they do not know. In this paper, we aim to teach LLMs to recognize and express their knowledge boundary, so they can reduce hallucinations caused by fabricating when they do not know. We propose CoKE, which first probes LLMs' knowledge boundary via internal confidence given a set of questions, and then leverages the probing results to elicit the expression of the knowledge boundary. Extensive experiments show CoKE helps LLMs express knowledge boundaries, answering known questions while declining unknown ones, significantly improving in-domain and out-of-domain performance.

Shifeng Liu, Yifang Sun, Bing Li et al.

The rapid growth in published clinical trials makes it difficult to maintain up-to-date systematic reviews, which requires finding all relevant trials. This leads to policy and practice decisions based on out-of-date, incomplete, and biased subsets of available clinical evidence. Extracting and then normalising Population, Intervention, Comparator, and Outcome (PICO) information from clinical trial articles may be an effective way to automatically assign trials to systematic reviews and avoid searching and screening - the two most time-consuming systematic review processes. We propose and test a novel approach to PICO span detection. The major difference between our proposed method and previous approaches comes from detecting spans without needing annotated span data and using only crowdsourced sentence-level annotations. Experiments on two datasets show that PICO span detection results achieve much higher results for recall when compared to fully supervised methods with PICO sentence detection at least as good as human annotations. By removing the reliance on expert annotations for span detection, this work could be used in human-machine pipeline for turning low-quality crowdsourced, and sentence-level PICO annotations into structured information that can be used to quickly assign trials to relevant systematic reviews.

Muhammad Asif Ali, Yifang Sun, Bing Li et al.

Fine-Grained Named Entity Typing (FG-NET) aims at classifying the entity mentions into a wide range of entity types (usually hundreds) depending upon the context. While distant supervision is the most common way to acquire supervised training data, it brings in label noise, as it assigns type labels to the entity mentions irrespective of mentions context. In attempts to deal with the label noise, leading research on the FG-NET assumes that the fine-grained entity typing data possesses a euclidean nature, which restraints the ability of the existing models in combating the label noise. Given the fact that the fine-grained type hierarchy exhibits a hierarchical structure, it makes hyperbolic space a natural choice to model the FG-NET data. In this research, we propose FGNET-RH, a novel framework that benefits from the hyperbolic geometry in combination with the graph structures to perform entity typing in a performance-enhanced fashion. FGNET-RH initially uses LSTM networks to encode the mention in relation with its context, later it forms a graph to distill/refine the mention encodings in the hyperbolic space. Finally, the refined mention encoding is used for entity typing. Experimentation using different benchmark datasets shows that FGNET-RH improves the performance on FG-NET by up to 3.5-% in terms of strict accuracy.

Muhammad Asif Ali, Yifang Sun, Bing Li et al.

Fine-Grained Named Entity Typing (FG-NET) is a key component in Natural Language Processing (NLP). It aims at classifying an entity mention into a wide range of entity types. Due to a large number of entity types, distant supervision is used to collect training data for this task, which noisily assigns type labels to entity mentions irrespective of the context. In order to alleviate the noisy labels, existing approaches on FGNET analyze the entity mentions entirely independent of each other and assign type labels solely based on mention sentence-specific context. This is inadequate for highly overlapping and noisy type labels as it hinders information passing across sentence boundaries. For this, we propose an edge-weighted attentive graph convolution network that refines the noisy mention representations by attending over corpus-level contextual clues prior to the end classification. Experimental evaluation shows that the proposed model outperforms the existing research by a relative score of upto 10.2% and 8.3% for macro f1 and micro f1 respectively.

Shifeng Liu, Yifang Sun, Bing Li et al.

To tackle Named Entity Recognition (NER) tasks, supervised methods need to obtain sufficient cleanly annotated data, which is labor and time consuming. On the contrary, distantly supervised methods acquire automatically annotated data using dictionaries to alleviate this requirement. Unfortunately, dictionaries hinder the effectiveness of distantly supervised methods for NER due to its limited coverage, especially in specific domains. In this paper, we aim at the limitations of the dictionary usage and mention boundary detection. We generalize the distant supervision by extending the dictionary with headword based non-exact matching. We apply a function to better weight the matched entity mentions. We propose a span-level model, which classifies all the possible spans then infers the selected spans with a proposed dynamic programming algorithm. Experiments on all three benchmark datasets demonstrate that our method outperforms previous state-of-the-art distantly supervised methods.

Weidi Xu, Xingyi Cheng, Kunlong Chen et al.

The ability of semantic reasoning over the sentence pair is essential for many natural language understanding tasks, e.g., natural language inference and machine reading comprehension. A recent significant improvement in these tasks comes from BERT. As reported, the next sentence prediction (NSP) in BERT, which learns the contextual relationship between two sentences, is of great significance for downstream problems with sentence-pair input. Despite the effectiveness of NSP, we suggest that NSP still lacks the essential signal to distinguish between entailment and shallow correlation. To remedy this, we propose to augment the NSP task to a 3-class categorization task, which includes a category for previous sentence prediction (PSP). The involvement of PSP encourages the model to focus on the informative semantics to determine the sentence order, thereby improves the ability of semantic understanding. This simple modification yields remarkable improvement against vanilla BERT. To further incorporate the document-level information, the scope of NSP and PSP is expanded into a broader range, i.e., NSP and PSP also include close but nonsuccessive sentences, the noise of which is mitigated by the label-smoothing technique. Both qualitative and quantitative experimental results demonstrate the effectiveness of the proposed method. Our method consistently improves the performance on the NLI and MRC benchmarks, including the challenging HANS dataset \cite{hans}, suggesting that the document-level task is still promising for the pre-training.

Wei Wang, Bin Bi, Ming Yan et al.

Recently, the pre-trained language model, BERT (and its robustly optimized version RoBERTa), has attracted a lot of attention in natural language understanding (NLU), and achieved state-of-the-art accuracy in various NLU tasks, such as sentiment classification, natural language inference, semantic textual similarity and question answering. Inspired by the linearization exploration work of Elman [8], we extend BERT to a new model, StructBERT, by incorporating language structures into pre-training. Specifically, we pre-train StructBERT with two auxiliary tasks to make the most of the sequential order of words and sentences, which leverage language structures at the word and sentence levels, respectively. As a result, the new model is adapted to different levels of language understanding required by downstream tasks. The StructBERT with structural pre-training gives surprisingly good empirical results on a variety of downstream tasks, including pushing the state-of-the-art on the GLUE benchmark to 89.0 (outperforming all published models), the F1 score on SQuAD v1.1 question answering to 93.0, the accuracy on SNLI to 91.7.