We analyze the masked language modeling pretraining objective function from the perspective of the distributional hypothesis. We investigate whether better sample efficiency and the better generalization capability of models pretrained with masked language modeling can be attributed to the semantic similarity encoded in the pretraining data's distributional property. Via a synthetic dataset, our analysis suggests that distributional property indeed leads to the better sample efficiency of pretrained masked language models, but does not fully explain the generalization capability. We also conduct analyses over two real-world datasets and demonstrate that the distributional property does not explain the generalization ability of pretrained natural language models either. Our results illustrate our limited understanding of model pretraining and provide future research directions.
We present an overview of the BLP Sentiment Shared Task, organized as part of the inaugural BLP 2023 workshop, co-located with EMNLP 2023. The task is defined as the detection of sentiment in a given piece of social media text. This task attracted interest from 71 participants, among whom 29 and 30 teams submitted systems during the development and evaluation phases, respectively. In total, participants submitted 597 runs. However, a total of 15 teams submitted system description papers. The range of approaches in the submitted systems spans from classical machine learning models, fine-tuning pre-trained models, to leveraging Large Language Model (LLMs) in zero- and few-shot settings. In this paper, we provide a detailed account of the task setup, including dataset development and evaluation setup. Additionally, we provide a brief overview of the systems submitted by the participants. All datasets and evaluation scripts from the shared task have been made publicly available for the research community, to foster further research in this domain.
Federated learning (FL) has become a key component in various language modeling applications such as machine translation, next-word prediction, and medical record analysis. These applications are trained on datasets from many FL participants that often include privacy-sensitive data, such as healthcare records, phone/credit card numbers, login credentials, etc. Although FL enables computation without necessitating clients to share their raw data, determining the extent of privacy leakage in federated language models is challenging and not straightforward. Moreover, existing attacks aim to extract data regardless of how sensitive or naive it is. To fill this research gap, we introduce two novel findings with regard to leaking privacy-sensitive user data from federated large language models. Firstly, we make a key observation that model snapshots from the intermediate rounds in FL can cause greater privacy leakage than the final trained model. Secondly, we identify that privacy leakage can be aggravated by tampering with a model's selective weights that are specifically responsible for memorizing the sensitive training data. We show how a malicious client can leak the privacy-sensitive data of some other users in FL even without any cooperation from the server. Our best-performing method improves the membership inference recall by 29% and achieves up to 71% private data reconstruction, evidently outperforming existing attacks with stronger assumptions of adversary capabilities.
Numerous studies have highlighted the privacy risks associated with pretrained large language models. In contrast, our research offers a unique perspective by demonstrating that pretrained large language models can effectively contribute to privacy preservation. We propose a locally differentially private mechanism called DP-Prompt, which leverages the power of pretrained large language models and zero-shot prompting to counter author de-anonymization attacks while minimizing the impact on downstream utility. When DP-Prompt is used with a powerful language model like ChatGPT (gpt-3.5), we observe a notable reduction in the success rate of de-anonymization attacks, showing that it surpasses existing approaches by a considerable margin despite its simpler design. For instance, in the case of the IMDB dataset, DP-Prompt (with ChatGPT) perfectly recovers the clean sentiment F1 score while achieving a 46\% reduction in author identification F1 score against static attackers and a 26\% reduction against adaptive attackers. We conduct extensive experiments across six open-source large language models, ranging up to 7 billion parameters, to analyze various effects of the privacy-utility tradeoff.
Recommendation system (RS) plays significant roles in matching users information needs for Internet applications, and it usually utilizes the vanilla neural network as the backbone to handle embedding details. Recently, the large language model (LLM) has exhibited emergent abilities and achieved great breakthroughs both in the CV and NLP communities. Thus, it is logical to incorporate RS with LLM better, which has become an emerging research direction. Although some existing works have made their contributions to this issue, they mainly consider the single key situation (e.g. historical interactions), especially in sequential recommendation. The situation of multiple key-value data is simply neglected. This significant scenario is mainstream in real practical applications, where the information of users (e.g. age, occupation, etc) and items (e.g. title, category, etc) has more than one key. Therefore, we aim to implement sequential recommendations based on multiple key-value data by incorporating RS with LLM. In particular, we instruct tuning a prevalent open-source LLM (Llama 7B) in order to inject domain knowledge of RS into the pre-trained LLM. Since we adopt multiple key-value strategies, LLM is hard to learn well among these keys. Thus the general and innovative shuffle and mask strategies, as an innovative manner of data argument, are designed. To demonstrate the effectiveness of our approach, extensive experiments are conducted on the popular and suitable dataset MovieLens which contains multiple keys-value. The experimental results demonstrate that our approach can nicely and effectively complete this challenging issue.
Despite their groundbreaking performance for many generative modeling tasks, diffusion models have fallen short on discrete data domains such as natural language. Crucially, standard diffusion models rely on the well-established theory of score matching, but efforts to generalize this to discrete structures have not yielded the same empirical gains. In this work, we bridge this gap by proposing score entropy, a novel loss that naturally extends score matching to discrete spaces, integrates seamlessly to build discrete diffusion models, and significantly boosts performance. Experimentally, we test our Score Entropy Discrete Diffusion models (SEDD) on standard language modeling tasks. For comparable model sizes, SEDD beats existing language diffusion paradigms (reducing perplexity by $25$-$75$\%) and is competitive with autoregressive models, in particular outperforming GPT-2. Furthermore, compared to autoregressive mdoels, SEDD generates faithful text without requiring distribution annealing techniques like temperature scaling (around $6$-$8\times$ better generative perplexity than un-annealed GPT-2), can trade compute and quality (similar quality with $32\times$ fewer network evaluations), and enables controllable infilling (matching nucleus sampling quality while enabling other strategies besides left to right prompting).
Inflation is a highly favoured theory for the early Universe. It is compatible with current observations of the cosmic microwave background and large scale structure and is a driver in the quest to detect primordial gravitational waves. It is also, given the current quality of the data, highly under-determined with a large number of candidate implementations. We use a new method in symbolic regression to generate all possible simple scalar field potentials for one of two possible basis sets of operators. Treating these as single-field, slow-roll inflationary models we then score them with an information-theoretic metric ("minimum description length") that quantifies their efficiency in compressing the information in current data. We explore two possible priors on the parameter space of potentials, one related to the functions' structural complexity and one that uses a Katz back-off language model to prefer functions that may be theoretically motivated. This enables us to identify the inflaton potentials that optimally balance simplicity with accuracy at explaining current data, which may subsequently find theoretical motivation. Our exploratory study opens the door to extraction of fundamental physics directly from data, and may be augmented with more refined theoretical priors in the quest for a complete understanding of the early Universe.
D scene graphs are an emerging 3D scene representation, that models both the objects present in the scene as well as their relationships. However, learning 3D scene graphs is a challenging task because it requires not only object labels but also relationship annotations, which are very scarce in datasets. While it is widely accepted that pre-training is an effective approach to improve model performance in low data regimes, in this paper, we find that existing pre-training methods are ill-suited for 3D scene graphs. To solve this issue, we present the first language-based pre-training approach for 3D scene graphs, whereby we exploit the strong relationship between scene graphs and language. To this end, we leverage the language encoder of CLIP, a popular vision-language model, to distill its knowledge into our graph-based network. We formulate a contrastive pre-training, which aligns text embeddings of relationships (subject-predicate-object triplets) and predicted 3D graph features. Our method achieves state-of-the-art results on the main semantic 3D scene graph benchmark by showing improved effectiveness over pre-training baselines and outperforming all the existing fully supervised scene graph prediction methods by a significant margin. Furthermore, since our scene graph features are language-aligned, it allows us to query the language space of the features in a zero-shot manner. In this paper, we show an example of utilizing this property of the features to predict the room type of a scene without further training.
Facial Expression Recognition (FER) is a crucial task in affective computing, but its conventional focus on the seven basic emotions limits its applicability to the complex and expanding emotional spectrum. To address the issue of new and unseen emotions present in dynamic in-the-wild FER, we propose a novel vision-language model that utilises sample-level text descriptions (i.e. captions of the context, expressions or emotional cues) as natural language supervision, aiming to enhance the learning of rich latent representations, for zero-shot classification. To test this, we evaluate using zero-shot classification of the model trained on sample-level descriptions on four popular dynamic FER datasets. Our findings show that this approach yields significant improvements when compared to baseline methods. Specifically, for zero-shot video FER, we outperform CLIP by over 10\% in terms of Weighted Average Recall and 5\% in terms of Unweighted Average Recall on several datasets. Furthermore, we evaluate the representations obtained from the network trained using sample-level descriptions on the downstream task of mental health symptom estimation, achieving performance comparable or superior to state-of-the-art methods and strong agreement with human experts. Namely, we achieve a Pearson's Correlation Coefficient of up to 0.85 on schizophrenia symptom severity estimation, which is comparable to human experts' agreement. The code is publicly available at: https://github.com/NickyFot/EmoCLIP.
Psychiatrists diagnose mental disorders via the linguistic use of patients. Still, due to data privacy, existing passive mental health monitoring systems use alternative features such as activity, app usage, and location via mobile devices. We propose FedTherapist, a mobile mental health monitoring system that utilizes continuous speech and keyboard input in a privacy-preserving way via federated learning. We explore multiple model designs by comparing their performance and overhead for FedTherapist to overcome the complex nature of on-device language model training on smartphones. We further propose a Context-Aware Language Learning (CALL) methodology to effectively utilize smartphones' large and noisy text for mental health signal sensing. Our IRB-approved evaluation of the prediction of self-reported depression, stress, anxiety, and mood from 46 participants shows higher accuracy of FedTherapist compared with the performance with non-language features, achieving 0.15 AUROC improvement and 8.21% MAE reduction.
The integration of path reasoning with language modeling in recommender systems has shown promise for enhancing explainability but often struggles with the authenticity of the explanations provided. Traditional models modify their architecture to produce entities and relations alternately--for example, employing separate heads for each in the model--which does not ensure the authenticity of paths reflective of actual Knowledge Graph (KG) connections. This misalignment can lead to user distrust due to the generation of corrupted paths. Addressing this, we introduce PEARLM (Path-based Explainable-Accurate Recommender based on Language Modelling), which innovates with a Knowledge Graph Constraint Decoding (KGCD) mechanism. This mechanism ensures zero incidence of corrupted paths by enforcing adherence to valid KG connections at the decoding level, agnostic of the underlying model architecture. By integrating direct token embedding learning from KG paths, PEARLM not only guarantees the generation of plausible and verifiable explanations but also highly enhances recommendation accuracy. We validate the effectiveness of our approach through a rigorous empirical assessment, employing a newly proposed metric that quantifies the integrity of explanation paths. Our results demonstrate a significant improvement over existing methods, effectively eliminating the generation of inaccurate paths and advancing the state-of-the-art in explainable recommender systems.
This paper introduces the Cross-lingual Fact Extraction and VERification (XFEVER) dataset designed for benchmarking the fact verification models across different languages. We constructed it by translating the claim and evidence texts of the Fact Extraction and VERification (FEVER) dataset into six languages. The training and development sets were translated using machine translation, whereas the test set includes texts translated by professional translators and machine-translated texts. Using the XFEVER dataset, two cross-lingual fact verification scenarios, zero-shot learning and translate-train learning, are defined, and baseline models for each scenario are also proposed in this paper. Experimental results show that the multilingual language model can be used to build fact verification models in different languages efficiently. However, the performance varies by language and is somewhat inferior to the English case. We also found that we can effectively mitigate model miscalibration by considering the prediction similarity between the English and target languages. The XFEVER dataset, code, and model checkpoints are available at https://github.com/nii-yamagishilab/xfever.
Language models trained on large-scale corpus often generate content that is harmful, toxic, or contrary to human preferences, making their alignment with human values a critical concern. Reinforcement learning from human feedback (RLHF) with algorithms like PPO is a prevalent approach for alignment but is often complex, unstable, and resource-intensive. Recently, ranking-based alignment methods have emerged, offering stability and effectiveness by replacing the RL framework with supervised fine-tuning, but they are costly due to the need for annotated data. Considering that existing large language models (LLMs) like ChatGPT are already relatively well-aligned and cost-friendly, researchers have begun to align the language model with human preference from AI feedback. The common practices, which unidirectionally distill the instruction-following responses from LLMs, are constrained by their bottleneck. Thus we introduce CycleAlign to distill alignment capabilities from parameter-invisible LLMs (black-box) to a parameter-visible model (white-box) in an iterative manner. With in-context learning (ICL) as the core of the cycle, the black-box models are able to rank the model-generated responses guided by human-craft instruction and demonstrations about their preferences. During iterative interaction, the white-box models also have a judgment about responses generated by them. Consequently, the agreement ranking could be viewed as a pseudo label to dynamically update the in-context demonstrations and improve the preference ranking ability of black-box models. Through multiple interactions, the CycleAlign framework could align the white-box model with the black-box model effectively in a low-resource way. Empirical results illustrate that the model fine-tuned by CycleAlign remarkably exceeds existing methods, and achieves the state-of-the-art performance in alignment with human value.
The quickly-expanding nature of published medical literature makes it challenging for clinicians and researchers to keep up with and summarize recent, relevant findings in a timely manner. While several closed-source summarization tools based on large language models (LLMs) now exist, rigorous and systematic evaluations of their outputs are lacking. Furthermore, there is a paucity of high-quality datasets and appropriate benchmark tasks with which to evaluate these tools. We address these issues with four contributions: we release Clinfo.ai, an open-source WebApp that answers clinical questions based on dynamically retrieved scientific literature; we specify an information retrieval and abstractive summarization task to evaluate the performance of such retrieval-augmented LLM systems; we release a dataset of 200 questions and corresponding answers derived from published systematic reviews, which we name PubMed Retrieval and Synthesis (PubMedRS-200); and report benchmark results for Clinfo.ai and other publicly available OpenQA systems on PubMedRS-200.
Two of the central factors believed to underpin human sentence processing difficulty are expectations and retrieval from working memory. A recent attempt to create a unified cognitive model integrating these two factors relied on the parallels between the self-attention mechanism of transformer language models and cue-based retrieval theories of working memory in human sentence processing (Ryu and Lewis 2021). While Ryu and Lewis show that attention patterns in specialized attention heads of GPT-2 are consistent with similarity-based interference, a key prediction of cue-based retrieval models, their method requires identifying syntactically specialized attention heads, and makes the cognitively implausible assumption that hundreds of memory retrieval operations take place in parallel. In the present work, we develop a recurrent neural language model with a single self-attention head, which more closely parallels the memory system assumed by cognitive theories. We show that our model's single attention head captures semantic and syntactic interference effects observed in human experiments.
While large language models demonstrate remarkable capabilities, they often present challenges in terms of safety, alignment with human values, and stability during training. Here, we focus on two prevalent methods used to align these models, Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). SFT is simple and robust, powering a host of open-source models, while RLHF is a more sophisticated method used in top-tier models like ChatGPT but also suffers from instability and susceptibility to reward hacking. We propose a novel approach, Supervised Iterative Learning from Human Feedback (SuperHF), which seeks to leverage the strengths of both methods. Our hypothesis is two-fold: that the reward model used in RLHF is critical for efficient data use and model generalization and that the use of Proximal Policy Optimization (PPO) in RLHF may not be necessary and could contribute to instability issues. SuperHF replaces PPO with a simple supervised loss and a Kullback-Leibler (KL) divergence prior. It creates its own training data by repeatedly sampling a batch of model outputs and filtering them through the reward model in an online learning regime. We then break down the reward optimization problem into three components: robustly optimizing the training rewards themselves, preventing reward hacking-exploitation of the reward model that degrades model performance-as measured by a novel METEOR similarity metric, and maintaining good performance on downstream evaluations. Our experimental results show SuperHF exceeds PPO-based RLHF on the training objective, easily and favorably trades off high reward with low reward hacking, improves downstream calibration, and performs the same on our GPT-4 based qualitative evaluation scheme all the while being significantly simpler to implement, highlighting SuperHF's potential as a competitive language model alignment technique.
Large language models (LLMs) have shown great potential to solve varieties of natural language processing (NLP) tasks, including mathematical reasoning. In this work, we present SkyMath, a large language model for mathematics with 13 billion parameters. By applying self-compare fine-tuning, we have enhanced mathematical reasoning abilities of Skywork-13B-Base remarkably. On GSM8K, SkyMath outperforms all known open-source models of similar size and has established a new SOTA performance.
A term in a corpus is said to be ``bursty'' (or overdispersed) when its occurrences are concentrated in few out of many documents. In this paper, we propose Residual Inverse Collection Frequency (RICF), a statistical significance test inspired heuristic for quantifying term burstiness. The chi-squared test is, to our knowledge, the sole test of statistical significance among existing term burstiness measures. Chi-squared test term burstiness scores are computed from the collection frequency statistic (i.e., the proportion that a specified term constitutes in relation to all terms within a corpus). However, the document frequency of a term (i.e., the proportion of documents within a corpus in which a specific term occurs) is exploited by certain other widely used term burstiness measures. RICF addresses this shortcoming of the chi-squared test by virtue of its term burstiness scores systematically incorporating both the collection frequency and document frequency statistics. We evaluate the RICF measure on a domain-specific technical terminology extraction task using the GENIA Term corpus benchmark, which comprises 2,000 annotated biomedical article abstracts. RICF generally outperformed the chi-squared test in terms of precision at k score with percent improvements of 0.00% (P@10), 6.38% (P@50), 6.38% (P@100), 2.27% (P@500), 2.61% (P@1000), and 1.90% (P@5000). Furthermore, RICF performance was competitive with the performances of other well-established measures of term burstiness. Based on these findings, we consider our contributions in this paper as a promising starting point for future exploration in leveraging statistical significance testing in text analysis.
Since ChatGPT released its API for public use, the number of applications built on top of commercial large language models (LLMs) increase exponentially. One popular usage of such models is leveraging its in-context learning ability and generating responses given user queries leveraging knowledge obtained by retrieval augmentation. One problem of deploying commercial retrieval-augmented LLMs is the cost due to the additionally retrieved context that largely increases the input token size of the LLMs. To mitigate this, we propose a token compression scheme that includes two methods: summarization compression and semantic compression. The first method applies a T5-based model that is fine-tuned by datasets generated using self-instruct containing samples with varying lengths and reduce token size by doing summarization. The second method further compresses the token size by removing words with lower impact on the semantic. In order to adequately evaluate the effectiveness of the proposed methods, we propose and utilize a dataset called Food-Recommendation DB (FRDB) focusing on food recommendation for women around pregnancy period or infants. Our summarization compression can reduce 65% of the retrieval token size with further 0.3% improvement on the accuracy; semantic compression provides a more flexible way to trade-off the token size with performance, for which we can reduce the token size by 20% with only 1.6% of accuracy drop.
The Transformer architecture is crucial for numerous AI models, but it still faces challenges in long-range language modeling. Though several specific transformer architectures have been designed to tackle issues of long-range dependencies, existing methods like Transformer-XL are plagued by a high percentage of ineffective memories. In this study, we present a plug-and-play strategy, known as TRAining-free Memory Selection (TRAMS), that selects tokens participating in attention calculation based on one simple metric. This strategy allows us to keep tokens that are likely to have a high attention score with the current queries and ignore the other ones. We have tested our approach on the word-level benchmark (WikiText-103) and the character-level benchmark (enwik8), and the results indicate an improvement without having additional training or adding additional parameters.