Parameter-efficient fine-tuning (PEFT) is crucial for customizing Large Language Models (LLMs) with constrained resource. Although there have been various PEFT methods for dense-architecture LLMs, PEFT for sparse-architecture LLMs is still underexplored. In this work, we study the PEFT method for LLMs with the Mixture-of-Experts (MoE) architecture and the contents of this work are mainly threefold: (1) We investigate the dispersion degree of the activated experts in customized tasks, and found that the routing distribution for specific task tend to be highly concentrated, while the distribution of activated experts varies significantly across different tasks. (2) We propose the expert-specialized fine-tuning method, which tunes the experts most relevant to downstream tasks while freezing the other experts; experimental results demonstrate that our method not only improves the tuning efficiency, but also matches or even surpasses the performance of full-parameter fine-tuning. (3) We further analyze the impact of the MoE architecture on expert-specialized fine-tuning. We find that MoE models with finer-grained experts are more advantageous in selecting the combination of experts that are most relevant to downstream tasks, thereby enhancing the both the training efficiency and effectiveness.
For extremely weak-supervised text classification, pioneer research generates pseudo labels by mining texts similar to the class names from the raw corpus, which may end up with very limited or even no samples for the minority classes. Recent works have started to generate the relevant texts by prompting LLMs using the class names or definitions; however, there is a high risk that LLMs cannot generate in-distribution (i.e., similar to the corpus where the text classifier will be applied) data, leading to ungeneralizable classifiers. In this paper, we combine the advantages of these two approaches and propose to bridge the gap via a novel framework, text grafting, which aims to obtain clean and near-distribution weak supervision for minority classes. Specifically, we first use LLM-based logits to mine masked templates from the raw corpus, which have a high potential for data synthesis into the target minority class. Then, the templates are filled by state-of-the-art LLMs to synthesize near-distribution texts falling into minority classes. Text grafting shows significant improvement over direct mining or synthesis on minority classes. We also use analysis and case studies to comprehend the property of text grafting.
Controlling the attribute intensity of text generation is crucial across scenarios (e.g., writing conciseness, chatting emotion, and explanation clarity). The remarkable capabilities of large language models (LLMs) have revolutionized text generation, prompting us to explore such smooth control of LLM generation. Specifically, we propose metrics to assess the range, calibration, and consistency of the generated text’s attribute intensity in response to varying control values, as well as its relevance to the intended context. To quantify the attribute intensity and context relevance, we leverage an Elo rating system and GPT4, respectively, both renowned for their robust alignment with human judgment. We look into two viable training-free methods for achieving smooth control of LLMs: (1) Prompting with semantic shifters, and (2) Modifying internal model representations. The evaluations of these two methods are conducted on 5 different attributes with various models.
The effectiveness of Large Language Models (LLMs) relies on their capacity to understand instructions and generate human-like responses. However, aligning LLMs with complex human preferences remains a significant challenge due to the potential misinterpretation of user prompts. Current methods for aligning LLM behaviors fall into two categories: output optimization (such as RLHF, RLAIF, and DPO) and input optimization (like OPRO and BPO). While both approaches aim to guide LLMs towards generating responses that align with desired objectives, the labor-intensive and intentions-inconsistent data annotation, as well as the strict and tedious training supervision, make them struggle to yield optimal results across all models. To address these shortcomings, we introduce a novel self-renewal approach called Prompt Optimization with Implicit Reasoning (POIR). It consists of two key components: 1) a model-specific and self-recirculating data collection method that leverages self-evaluation to enhance prompts in accordance with the model’s intrinsic logits, and 2) a prompt rewrite schema that injects implicit reasoning for direct preference learning. Through self-renewal optimization, POIR refines LLM outputs to better align with human preferences across various LLMs and tasks, without relying on supervised fine-tuning. Extensive experiments on a range of LLMs and tasks demonstrate POIR’s superior performance. We believe this advancement offers a novel paradigm for developing LLMs that are more attuned to user intentions.
Process supervision, using a trained verifier to evaluate the intermediate steps generated by a reasoner, has demonstrated significant improvements in multi-step problem solving. In this paper, to avoid the expensive effort of human annotation on the verifier training data, we introduce Model-induced Process Supervision (MiPS), a novel method for automating data curation. MiPS annotates an intermediate step by sampling completions of this solution through the reasoning model, and obtaining an accuracy defined as the proportion of correct completions. Inaccuracies of the reasoner would cause MiPS underestimating the accuracy of intermediate steps, therefore, we suggest and empirically show that verification focusing on high predicted scores of the verifier shall be preferred over that of low predicted scores, contrary to prior observations on human curated data. Our approach significantly improves the performance of PaLM 2 on math and coding tasks (accuracy +0.67% on GSM8K, +4.16% on MATH, +0.92% on MBPP compared with an output supervision trained verifier). Additionally, our study demonstrates that the verifier exhibits strong generalization ability across different reasoning models.
Large language models (LLMs) have shown excellent mastering of human language but still struggle in real-world applications that require mathematical problem-solving. While many strategies and datasets to enhance LLMs’ mathematics are developed, it remains a challenge to simultaneously maintain and improve both language and mathematical capabilities in deployed LLM systems. In this work, we tailor the Self-Critique pipeline, which addresses the challenge in the feedback learning stage of LLM alignment. We first train a general Math-Critique model from the LLM itself to provide feedback signals. Then, we sequentially employ rejective fine-tuning and direct preference optimization over the LLM’s own generations for data collection. Based on ChatGLM3-32B, we conduct experiments on both academic and our newly created challenging dataset, MathUserEval. Results show that our pipeline significantly enhances the LLM’s mathematical problem-solving while still improving its language ability, outperforming LLMs that could be two times larger. Related techniques have been deployed to ChatGLM, an online serving LLM. Related evaluation datasets and scripts are released at https://github.com/THUDM/ChatGLM-Math.
Esports, a sports competition on video games, has become one of the most important sporting events. Although esports play logs have been accumulated, only a small portion of them accompany text commentaries for the audience to retrieve and understand the plays. In this study, we therefore introduce the task of generating game commentaries from esports’ data records. We first build large-scale esports data-to-text datasets that pair structured data and commentaries from a popular esports game, League of Legends. We then evaluate Transformer-based models to generate game commentaries from structured data records, while examining the impact of the pre-trained language models. Evaluation results on our dataset revealed the challenges of this novel task. We will release our dataset to boost potential research in the data-to-text generation community.
In this paper, we present TeleChat, a collection of large language models (LLMs) with parameters of 7 billion and 12 billion. TeleChat is initially pretrained on an extensive corpus containing a diverse collection of texts from both English and Chinese languages, encompassing trillions of tokens. Subsequently, the model undergoes fine-tuning to align with human preferences, following a detailed methodology that we describe. We evaluate the performance of TeleChat on various tasks, including general dialogue generation, language understanding, mathematics, reasoning, code generation, and knowledge-based question answering. Our findings indicate that TeleChat achieves state-of-the-art performance to other open-source models of similar size across a wide range of public benchmarks. To support future research and applications utilizing LLMs, we release the fine-tuned model checkpoints of TeleChat-7B and TeleChat-12B, along with code and a portion of our filtered high-quality pretraining data, to the public community.
This work aims to build a text embedder that can capture characteristics of texts specified by user instructions clarifying the similarity criterion. While previous methods improve general task awareness by injecting the instruction information into encoding, they fail to be sensitive to clearer criteria like “evaluate similarity based on emotion”. We instead propose a different viewpoint, which treats the instruction as a “question” about the input text and encodes the expected answers to obtain the representation accordingly. Intuitively, texts with the same (implicit) semantics would share similar answers following the instruction, thus leading to more similar representations. Specifically, we propose InBedder that instantiates this learning-to-answer idea by only fine-tuning language models via abstractive question answering tasks. Despite its simplicity, InBedder demonstrates significantly improved instruction-following capabilities according to our proposed instruction awareness tests and instruction robustness tests, when applied to language models with large language models (LLMs) (e.g., llama-2-7b) and smaller encoder-based LMs (e.g., roberta-large). Additionally, our qualitative analysis of clustering outcomes, achieved by applying diverse instructions to the same unlabeled corpus, demonstrates a high degree of interpretability in the clusters formed.
Recent advances in Automated Theorem Proving have shown the effectiveness of leveraging a (large) language model that generates tactics (i.e. proof steps) to search through proof states. The current model, while trained solely on successful proof paths, faces a discrepancy at the inference stage, as it must sample and try various tactics at each proof state until finding success, unlike its training which does not incorporate learning from failed attempts. Intuitively, a tactic that leads to a failed search path would indicate that similar tactics should receive less attention during the following trials. In this paper, we demonstrate the benefit of training models that additionally learn from failed search paths. Facing the lack of such trial-and-error data in existing open-source theorem-proving datasets, we curate a dataset on intuitionistic propositional logic theorems and formalize it in Lean, such that we can reliably check the correctness of proofs. We compare our model trained on relatively short trial-and-error information (TrialMaster) with models trained only on the correct paths and discover that the former solves more unseen theorems with lower trial searches.
Hierarchical text classification (HTC) is a challenging subtask of multi-label classification due to its complex taxonomic structure. Nearly all recent HTC works focus on how the labels are structured but ignore the sub-structure of ground-truth labels according to each input text which contains fruitful label co-occurrence information. In this work, we introduce this local hierarchy with an adversarial framework. We propose a HiAdv framework that can fit in nearly all HTC models and optimize them with the local hierarchy as auxiliary information. We test on two typical HTC models and find that HiAdv is effective in all scenarios and is adept at dealing with complex taxonomic hierarchies. Further experiments demonstrate that the promotion of our framework indeed comes from the local hierarchy and the local hierarchy is beneficial for rare classes which have insufficient training data.
Extremely Weakly Supervised Text Classification (XWS-TC) refers to text classification based on minimal high-level human guidance, such as a few label-indicative seed words or classification instructions. There are two mainstream approaches for XWS-TC, however, never being rigorously compared: (1) training classifiers based on pseudo-labels generated by (softly) matching seed words (Seed) and (2) prompting (and calibrating) language models using classification instruction (and raw texts) to decode label words (Prompt). This paper presents the first XWS-TC benchmark to compare the two approaches on fair grounds, where the datasets, supervisions, and hyperparameter choices are standardized across methods. Our benchmarking results suggest that (1) Both Seed and Prompt approaches are competitive and there is no clear winner; (2) Seed is empirically more tolerant than Prompt to human guidance (e.g., seed words, classification instructions, and label words) changes; (3) Seed is empirically more selective than Prompt to the pre-trained language models; (4) Recent Seed and Prompt methods have close connections and a clustering post-processing step based on raw in-domain texts is a strong performance booster to both. We hope this benchmark serves as a guideline in selecting XWS-TC methods in different scenarios and stimulate interest in developing guidance- and model-robust XWS-TC methods.
Multiple defendants in a criminal fact description generally exhibit complex interactions, and cannot be well handled by existing Legal Judgment Prediction (LJP) methods which focus on predicting judgment results (e.g., law articles, charges, and terms of penalty) for single-defendant cases. To address this problem, we propose the task of multi-defendant LJP, which aims to automatically predict the judgment results for each defendant of multi-defendant cases. Two challenges arise with the task of multi-defendant LJP: (1) indistinguishable judgment results among various defendants; and (2) the lack of a real-world dataset for training and evaluation. To tackle the first challenge, we formalize the multi-defendant judgment process as hierarchical reasoning chains and introduce a multi-defendant LJP method, named Hierarchical Reasoning Network (HRN), which follows the hierarchical reasoning chains to determine criminal relationships, sentencing circumstances, law articles, charges, and terms of penalty for each defendant. To tackle the second challenge, we collect a real-world multi-defendant LJP dataset, namely MultiLJP, to accelerate the relevant research in the future. Extensive experiments on MultiLJP verify the effectiveness of our proposed HRN.
Few-shot named entity recognition (NER) has shown remarkable progress in identifying entities in low-resource domains. However, few-shot NER methods still struggle with out-of-domain (OOD) examples due to their reliance on manual labeling for the target domain. To address this limitation, recent studies enable generalization to an unseen target domain with only a few labeled examples using data augmentation techniques. Two important challenges remain: First, augmentation is limited to the training data, resulting in minimal overlap between the generated data and OOD examples. Second, knowledge transfer is implicit and insufficient, severely hindering model generalizability and the integration of knowledge from the source domain. In this paper, we propose a framework, prompt learning with type-related features (PLTR), to address these challenges. To identify useful knowledge in the source domain and enhance knowledge transfer, PLTR automatically extracts entity type-related features (TRFs) based on mutual information criteria. To bridge the gap between training and OOD data, PLTR generates a unique prompt for each unseen example by selecting relevant TRFs. We show that PLTR achieves significant performance improvements on in-domain and cross-domain datasets. The use of PLTR facilitates model adaptation and increases representation similarities between the source and unseen domains.
Despite remarkable advances that large language models have achieved in chatbots nowadays, maintaining a non-toxic user-AI interactive environment has become increasingly critical nowadays. However, previous efforts in toxicity detection have been mostly based on benchmarks derived from social media contents, leaving the unique challenges inherent to real-world user-AI interactions insufficiently explored. In this work, we introduce ToxicChat, a novel benchmark constructed based on real user queries from an open-source chatbot. This benchmark contains the rich, nuanced phenomena that can be tricky for current toxicity detection models to identify, revealing a significant domain difference when compared to social media contents. Our systematic evaluation of models trained on existing toxicity datasets has shown their shortcomings when applied to this unique domain of ToxicChat. Our work illuminates the potentially overlooked challenges of toxicity detection in real-world user-AI conversations. In the future, ToxicChat can be a valuable resource to drive further advancements toward building a safe and healthy environment for user-AI interactions.
We study the named entity recognition (NER) problem under the extremely weak supervision (XWS) setting, where only one example entity per type is given in a context-free way. While one can see that XWS is lighter than one-shot in terms of the amount of supervision, we propose a novel method X-NER that can outperform the state-of-the-art one-shot NER methods. We first mine entity spans that are similar to the example entities from an unlabelled training corpus. Instead of utilizing entity span representations from language models, we find it more effective to compare the context distributions before and after the span is replaced by the entity example. We then leverage the top-ranked spans as pseudo-labels to train an NER tagger. Extensive experiments and analyses on 4 NER datasets show the superior end-to-end NER performance of X-NER, outperforming the state-of-the-art few-shot methods with 1-shot supervision and ChatGPT annotations significantly. Finally, our X-NER possesses several notable properties, such as inheriting the cross-lingual abilities of the underlying language models.
Recent advances in weakly supervised text classification mostly focus on designing sophisticated methods to turn high-level human heuristics into quality pseudo-labels. In this paper, we revisit the seed matching-based method, which is arguably the simplest way to generate pseudo-labels, and show that its power was greatly underestimated. We show that the limited performance of seed matching is largely due to the label bias injected by the simple seed-match rule, which prevents the classifier from learning reliable confidence for selecting high-quality pseudo-labels. Interestingly, simply deleting the seed words present in the matched input texts can mitigate the label bias and help learn better confidence. Subsequently, the performance achieved by seed matching can be improved significantly, making it on par with or even better than the state-of-the-art. Furthermore, to handle the case when the seed words are not made known, we propose to simply delete the word tokens in the input text randomly with a high deletion ratio. Remarkably, seed matching equipped with this random deletion method can often achieve even better performance than that with seed deletion.
Unsupervised clustering is widely used to explore large corpora, but existing formulations neither consider the users’ goals nor explain clusters’ meanings. We propose a new task formulation, “Goal-Driven Clustering with Explanations” (GoalEx), which represents both the goal and the explanations as free-form language descriptions. For example, to categorize the errors made by a summarization system, the input to GoalEx is a corpus of annotator-written comments for system-generated summaries and a goal description “cluster the comments based on why the annotators think the summary is imperfect.”; the outputs are text clusters each with an explanation (“this cluster mentions that the summary misses important context information.”), which relates to the goal and accurately explains which comments should (not) belong to a cluster. To tackle GoalEx, we prompt a language model with “[corpus subset] + [goal] + Brainstorm a list of explanations each representing a cluster.”; then we classify whether each sample belongs to a cluster based on its explanation; finally, we use integer linear programming to select a subset of candidate clusters to cover most samples while minimizing overlaps. Under both automatic and human evaluation on corpora with or without labels, our method produces more accurate and goal-related explanations than prior methods.
We introduce ClusterLLM, a novel text clustering framework that leverages feedback from an instruction-tuned large language model, such as ChatGPT. Compared with traditional unsupervised methods that builds upon “small” embedders, ClusterLLM exhibits two intriguing advantages: (1) it enjoys the emergent capability of LLM even if its embeddings are inaccessible; and (2) it understands the user’s preference on clustering through textual instruction and/or a few annotated data. First, we prompt ChatGPT for insights on clustering perspective by constructing hard triplet questions <does A better correspond to B than C>, where A, B and C are similar data points that belong to different clusters according to small embedder. We empirically show that this strategy is both effective for fine-tuning small embedder and cost-efficient to query ChatGPT. Second, we prompt ChatGPT for helps on clustering granularity by carefully designed pairwise questions <do A and B belong to the same category>, and tune the granularity from cluster hierarchies that is the most consistent with the ChatGPT answers. Extensive experiments on 14 datasets show that ClusterLLM consistently improves clustering quality, at an average cost of ~$0.6 per dataset.
In dialogue state tracking, dialogue history is a crucial material, and its utilization varies between different models. However, no matter how the dialogue history is used, each existing model uses its own consistent dialogue history during the entire state tracking process, regardless of which slot is updated. Apparently, it requires different dialogue history to update different slots in different turns. Therefore, using consistent dialogue contents may lead to insufficient or redundant information for different slots, which affects the overall performance. To address this problem, we devise DiCoS-DST to dynamically select the relevant dialogue contents corresponding to each slot for state updating. Specifically, it first retrieves turn-level utterances of dialogue history and evaluates their relevance to the slot from a combination of three perspectives: (1) its explicit connection to the slot name; (2) its relevance to the current turn dialogue; (3) Implicit Mention Oriented Reasoning. Then these perspectives are combined to yield a decision, and only the selected dialogue contents are fed into State Generator, which explicitly minimizes the distracting information passed to the downstream state prediction. Experimental results show that our approach achieves new state-of-the-art performance on MultiWOZ 2.1 and MultiWOZ 2.2, and achieves superior performance on multiple mainstream benchmark datasets (including Sim-M, Sim-R, and DSTC2).
Hierarchical text classification is a challenging subtask of multi-label classification due to its complex label hierarchy. Existing methods encode text and label hierarchy separately and mix their representations for classification, where the hierarchy remains unchanged for all input text. Instead of modeling them separately, in this work, we propose Hierarchy-guided Contrastive Learning (HGCLR) to directly embed the hierarchy into a text encoder. During training, HGCLR constructs positive samples for input text under the guidance of the label hierarchy. By pulling together the input text and its positive sample, the text encoder can learn to generate the hierarchy-aware text representation independently. Therefore, after training, the HGCLR enhanced text encoder can dispense with the redundant hierarchy. Extensive experiments on three benchmark datasets verify the effectiveness of HGCLR.
Hierarchical text classification (HTC) is a challenging subtask of multi-label classification due to its complex label hierarchy.Recently, the pretrained language models (PLM)have been widely adopted in HTC through a fine-tuning paradigm. However, in this paradigm, there exists a huge gap between the classification tasks with sophisticated label hierarchy and the masked language model (MLM) pretraining tasks of PLMs and thus the potential of PLMs cannot be fully tapped.To bridge the gap, in this paper, we propose HPT, a Hierarchy-aware Prompt Tuning method to handle HTC from a multi-label MLM perspective.Specifically, we construct a dynamic virtual template and label words that take the form of soft prompts to fuse the label hierarchy knowledge and introduce a zero-bounded multi-label cross-entropy loss to harmonize the objectives of HTC and MLM.Extensive experiments show HPT achieves state-of-the-art performances on 3 popular HTC datasets and is adept at handling the imbalance and low resource situations. Our code is available at https://github.com/wzh9969/HPT.
Existing backdoor defense methods are only effective for limited trigger types. To defend different trigger types at once, we start from the class-irrelevant nature of the poisoning process and propose a novel weakly supervised backdoor defense framework WeDef. Recent advances in weak supervision make it possible to train a reasonably accurate text classifier using only a small number of user-provided, class-indicative seed words. Such seed words shall be considered independent of the triggers. Therefore, a weakly supervised text classifier trained by only the poisoned documents without their labels will likely have no backdoor. Inspired by this observation, in WeDef, we define the reliability of samples based on whether the predictions of the weak classifier agree with their labels in the poisoned training set. We further improve the results through a two-phase sanitization: (1) iteratively refine the weak classifier based on the reliable samples and (2) train a binary poison classifier by distinguishing the most unreliable samples from the most reliable samples. Finally, we train the sanitized model on the samples that the poison classifier predicts as benign. Extensive experiments show that WeDef is effective against popular trigger-based attacks (e.g., words, sentences, and paraphrases), outperforming existing defense methods.
We present a comprehensive study of sparse attention patterns in Transformer models. We first question the need for pre-training with sparse attention and present experiments showing that an efficient fine-tuning only approach yields a slightly worse but still competitive model. Then we compare the widely used local attention pattern and the less-well-studied global attention pattern, demonstrating that global patterns have several unique advantages. We also demonstrate that a flexible approach to attention, with different patterns across different layers of the model, is beneficial for some tasks. Drawing on this insight, we propose a novel Adaptive Axis Attention method, which learns—during fine-tuning—different attention patterns for each Transformer layer depending on the downstream task. Rather than choosing a fixed attention pattern, the adaptive axis attention method identifies important tokens—for each task and model layer—and focuses attention on those. It does not require pre-training to accommodate the sparse patterns and demonstrates competitive and sometimes better performance against fixed sparse attention patterns that require resource-intensive pre-training.
Fine-tuning pre-trained language models is a common practice in building NLP models for various tasks, including the case with less supervision. We argue that under the few-shot setting, formulating fine-tuning closer to the pre-training objective shall be able to unleash more benefits from the pre-trained language models. In this work, we take few-shot named entity recognition (NER) for a pilot study, where existing fine-tuning strategies are much different from pre-training. We propose a novel few-shot fine-tuning framework for NER, FFF-NER. Specifically, we introduce three new types of tokens, “is-entity”, “which-type” and “bracket”, so we can formulate the NER fine-tuning as (masked) token prediction or generation, depending on the choice of the pre-training objective. In our experiments, we apply to fine-tune both BERT and BART for few-shot NER on several benchmark datasets and observe significant improvements over existing fine-tuning strategies, including sequence labeling, prototype meta-learning, and prompt-based approaches. We further perform a series of ablation studies, showing few-shot NER performance is strongly correlated with the similarity between fine-tuning and pre-training.
The goal of dialogue state tracking (DST) is to predict the current dialogue state given all previous dialogue contexts. Existing approaches generally predict the dialogue state at every turn from scratch. However, the overwhelming majority of the slots in each turn should simply inherit the slot values from the previous turn. Therefore, the mechanism of treating slots equally in each turn not only is inefficient but also may lead to additional errors because of the redundant slot value generation. To address this problem, we devise the two-stage DSS-DST which consists of the Dual Slot Selector based on the current turn dialogue, and the Slot Value Generator based on the dialogue history. The Dual Slot Selector determines each slot whether to update slot value or to inherit the slot value from the previous turn from two aspects: (1) if there is a strong relationship between it and the current turn dialogue utterances; (2) if a slot value with high reliability can be obtained for it through the current turn dialogue. The slots selected to be updated are permitted to enter the Slot Value Generator to update values by a hybrid method, while the other slots directly inherit the values from the previous turn. Empirical results show that our method achieves 56.93%, 60.73%, and 58.04% joint accuracy on MultiWOZ 2.0, MultiWOZ 2.1, and MultiWOZ 2.2 datasets respectively and achieves a new state-of-the-art performance with significant improvements.
In this paper, we explore text classification with extremely weak supervision, i.e., only relying on the surface text of class names. This is a more challenging setting than the seed-driven weak supervision, which allows a few seed words per class. We opt to attack this problem from a representation learning perspective—ideal document representations should lead to nearly the same results between clustering and the desired classification. In particular, one can classify the same corpus differently (e.g., based on topics and locations), so document representations should be adaptive to the given class names. We propose a novel framework X-Class to realize the adaptive representations. Specifically, we first estimate class representations by incrementally adding the most similar word to each class until inconsistency arises. Following a tailored mixture of class attention mechanisms, we obtain the document representation via a weighted average of contextualized word representations. With the prior of each document assigned to its nearest class, we then cluster and align the documents to classes. Finally, we pick the most confident documents from each cluster to train a text classifier. Extensive experiments demonstrate that X-Class can rival and even outperform seed-driven weakly supervised methods on 7 benchmark datasets.
Contextualized representations based on neural language models have furthered the state of the art in various NLP tasks. Despite its great success, the nature of such representations remains a mystery. In this paper, we present an empirical property of these representations—”average” approximates “first principal component”. Specifically, experiments show that the average of these representations shares almost the same direction as the first principal component of the matrix whose columns are these representations. We believe this explains why the average representation is always a simple yet strong baseline. Our further examinations show that this property also holds in more challenging scenarios, for example, when the representations are from a model right after its random initialization. Therefore, we conjecture that this property is intrinsic to the distribution of representations and not necessarily related to the input structure. We realize that these representations empirically follow a normal distribution for each dimension, and by assuming this is true, we demonstrate that the empirical property can be in fact derived mathematically.
Multilingual BERT (M-BERT) has been a huge success in both supervised and zero-shot cross-lingual transfer learning. However, this success is focused only on the top 104 languages in Wikipedia it was trained on. In this paper, we propose a simple but effective approach to extend M-BERT E-MBERT so it can benefit any new language, and show that our approach aids languages that are already in M-BERT as well. We perform an extensive set of experiments with Named Entity Recognition (NER) on 27 languages, only 16 of which are in M-BERT, and show an average increase of about 6% F1 on M-BERT languages and 23% F1 increase on new languages. We release models and code at http://cogcomp.org/page/publication_view/912.
Everyone makes mistakes. So do human annotators when curating labels for named entity recognition (NER). Such label mistakes might hurt model training and interfere model comparison. In this study, we dive deep into one of the widely-adopted NER benchmark datasets, CoNLL03 NER. We are able to identify label mistakes in about 5.38% test sentences, which is a significant ratio considering that the state-of-the-art test F1 score is already around 93%. Therefore, we manually correct these label mistakes and form a cleaner test set. Our re-evaluation of popular models on this corrected test set leads to more accurate assessments, compared to those on the original test set. More importantly, we propose a simple yet effective framework, CrossWeigh, to handle label mistakes during NER model training. Specifically, it partitions the training data into several folds and train independent NER models to identify potential mistakes in each fold. Then it adjusts the weights of training data accordingly to train the final NER model. Extensive experiments demonstrate significant improvements of plugging various NER models into our proposed framework on three datasets. All implementations and corrected test set are available at our Github repo https://github.com/ZihanWangKi/CrossWeigh.