Large Language Models (LLMs) have emerged as a significant advancement in the field of Natural Language Processing (NLP), demonstrating remarkable capabilities in language generation and other language-centric tasks. Despite their evaluation across a multitude of analytical and reasoning tasks in various scientific domains, a comprehensive exploration of their knowledge and understanding within the realm of natural language tasks in the telecommunications domain is still needed. This study, therefore, seeks to evaluate the knowledge and understanding capabilities of LLMs within this domain. To achieve this, we conduct an exhaustive zero-shot evaluation of four prominent LLMs-Llama-2, Falcon, Mistral, and Zephyr. These models require fewer resources than ChatGPT, making them suitable for resource-constrained environments. Their performance is compared with state-of-the-art, fine-tuned models. To the best of our knowledge, this is the first work to extensively evaluate and compare the understanding of LLMs across multiple language-centric tasks in this domain. Our evaluation reveals that zero-shot LLMs can achieve performance levels comparable to the current state-of-the-art fine-tuned models. This indicates that pretraining on extensive text corpora equips LLMs with a degree of specialization, even within the telecommunications domain. We also observe that no single LLM consistently outperforms others, and the performance of different LLMs can fluctuate. Although their performance lags behind fine-tuned models, our findings underscore the potential of LLMs as a valuable resource for understanding various aspects of this field that lack large annotated data.
Large Language Models (LLMs) are often trained with safety guards intended to prevent harmful text generation. However, such safety training can be removed by fine-tuning the LLM on harmful datasets. While this emerging threat (harmful fine-tuning attacks) has been characterized by previous work, there is little understanding of how we should proceed in constructing and validating defenses against these attacks especially in the case where defenders would not have control of the fine-tuning process. We introduce a formal framework based on the training budget of an attacker which we call "Immunization" conditions. Using a formal characterisation of the harmful fine-tuning problem, we provide a thorough description of what a successful defense must comprise of and establish a set of guidelines on how rigorous defense research that gives us confidence should proceed.
Effective feature representations play a critical role in enhancing the performance of text generation models that rely on deep neural networks. However, current approaches suffer from several drawbacks, such as the inability to capture the deep semantics of language and sensitivity to minor input variations, resulting in significant changes in the generated text. In this paper, we present a novel solution to these challenges by employing a mixture of experts, multiple encoders, to offer distinct perspectives on the emotional state of the user's utterance while simultaneously enhancing performance. We propose an end-to-end model architecture called ASEM that performs emotion analysis on top of sentiment analysis for open-domain chatbots, enabling the generation of empathetic responses that are fluent and relevant. In contrast to traditional attention mechanisms, the proposed model employs a specialized attention strategy that uniquely zeroes in on sentiment and emotion nuances within the user's utterance. This ensures the generation of context-rich representations tailored to the underlying emotional tone and sentiment intricacies of the text. Our approach outperforms existing methods for generating empathetic embeddings, providing empathetic and diverse responses. The performance of our proposed model significantly exceeds that of existing models, enhancing emotion detection accuracy by 6.2% and lexical diversity by 1.4%.
Email continues to be a pivotal and extensively utilized communication medium within professional and commercial domains. Nonetheless, the prevalence of spam emails poses a significant challenge for users, disrupting their daily routines and diminishing productivity. Consequently, accurately identifying and filtering spam based on content has become crucial for cybersecurity. Recent advancements in natural language processing, particularly with large language models like ChatGPT, have shown remarkable performance in tasks such as question answering and text generation. However, its potential in spam identification remains underexplored. To fill in the gap, this study attempts to evaluate ChatGPT's capabilities for spam identification in both English and Chinese email datasets. We employ ChatGPT for spam email detection using in-context learning, which requires a prompt instruction with (or without) a few demonstrations. We also investigate how the number of demonstrations in the prompt affects the performance of ChatGPT. For comparison, we also implement five popular benchmark methods, including naive Bayes, support vector machines (SVM), logistic regression (LR), feedforward dense neural networks (DNN), and BERT classifiers. Through extensive experiments, the performance of ChatGPT is significantly worse than deep supervised learning methods in the large English dataset, while it presents superior performance on the low-resourced Chinese dataset. This study provides insights into the potential and limitations of ChatGPT for spam identification, highlighting its potential as a viable solution for resource-constrained language domains.
Large Language Models (LLMs) are widely used to evaluate natural language generation tasks as automated metrics. However, the likelihood, a measure of LLM's plausibility for a sentence, can vary due to superficial differences in sentences, such as word order and sentence structure. It is therefore possible that there might be a likelihood bias if LLMs are used for evaluation: they might overrate sentences with higher likelihoods while underrating those with lower likelihoods. In this paper, we investigate the presence and impact of likelihood bias in LLM-based evaluators. We also propose a method to mitigate the likelihood bias. Our method utilizes highly biased instances as few-shot examples for in-context learning. Our experiments in evaluating the data-to-text and grammatical error correction tasks reveal that several LLMs we test display a likelihood bias. Furthermore, our proposed method successfully mitigates this bias, also improving evaluation performance (in terms of correlation of models with human scores) significantly.
The topic-to-essay generation task is a challenging natural language generation task that aims to generate paragraph-level text with high semantic coherence based on a given set of topic words. Previous work has focused on the introduction of external knowledge, ignoring the insufficient generated text diversity. In order to improve the generation diversity, we propose a novel copy mechanism model with a content selection module that integrates rich semantic knowledge from the language model into the decoder. Furthermore, we introduce the improved prefix tuning method to train the model, enabling it to adapt to varying input complexities. In addition, we have contributed a new Chinese dataset for TEG tasks. Experimental results demonstrate that the proposed model can improve the generated text diversity by 35\% to 59\% compared to the state-of-the-art method, while maintaining a high level of topic consistency.
While Large Language Models (LLMs) demonstrate impressive capabilities in text generation, we find that their ability has yet to be generalized to music, humanity's creative language. We introduce ChatMusician, an open-source LLM that integrates intrinsic musical abilities. It is based on continual pre-training and finetuning LLaMA2 on a text-compatible music representation, ABC notation, and the music is treated as a second language. ChatMusician can understand and generate music with a pure text tokenizer without any external multi-modal neural structures or tokenizers. Interestingly, endowing musical abilities does not harm language abilities, even achieving a slightly higher MMLU score. Our model is capable of composing well-structured, full-length music, conditioned on texts, chords, melodies, motifs, musical forms, etc, surpassing GPT-4 baseline. On our meticulously curated college-level music understanding benchmark, MusicTheoryBench, ChatMusician surpasses LLaMA2 and GPT-3.5 on zero-shot setting by a noticeable margin. Our work reveals that LLMs can be an excellent compressor for music, but there remains significant territory to be conquered. We release our 4B token music-language corpora MusicPile, the collected MusicTheoryBench, code, model and demo in GitHub.
Generative artificial intelligence (AI) and large language models (LLMs) have gained rapid popularity through publicly available tools such as ChatGPT. The adoption of LLMs for personal and professional use is fueled by the natural interactions between human users and computer applications such as ChatGPT, along with powerful summarization and text generation capabilities. Given the widespread use of such generative AI tools, in this work we investigate how these tools can be deployed in a non-safety critical, strategic traffic flow management setting. Specifically, we train an LLM, CHATATC, based on a large historical data set of Ground Delay Program (GDP) issuances, spanning 2000-2023 and consisting of over 80,000 GDP implementations, revisions, and cancellations. We test the query and response capabilities of CHATATC, documenting successes (e.g., providing correct GDP rates, durations, and reason) and shortcomings (e.g,. superlative questions). We also detail the design of a graphical user interface for future users to interact and collaborate with the CHATATC conversational agent.
Impressive results have been achieved in natural language processing (NLP) tasks through the training of large language models (LLMs). However, these models occasionally produce toxic content such as insults, threats, and profanity in response to certain prompts, thereby constraining their practical utility. To tackle this issue, various finetuning-based and decoding-based approaches have been utilized to mitigate toxicity. However, these methods typically necessitate additional costs such as high-quality training data or auxiliary models. In this paper, we propose fine-grained detoxification via instance-level prefixes (FGDILP) to mitigate toxic text without additional cost. Specifically, FGDILP contrasts the contextualized representation in attention space using a positive prefix-prepended prompt against multiple negative prefix-prepended prompts at the instance level. This allows for constructing fine-grained subtoxicity vectors, which enables collaborative detoxification by fusing them to correct the normal generation process when provided with a raw prompt. We validate that FGDILP enables controlled text generation with regard to toxicity at both the utterance and context levels. Our method surpasses prompt-based baselines in detoxification, although at a slight cost to generation fluency and diversity.
Counterfactual generation lies at the core of various machine learning tasks, including image translation and controllable text generation. This generation process usually requires the identification of the disentangled latent representations, such as content and style, that underlie the observed data. However, it becomes more challenging when faced with a scarcity of paired data and labeling information. Existing disentangled methods crucially rely on oversimplified assumptions, such as assuming independent content and style variables, to identify the latent variables, even though such assumptions may not hold for complex data distributions. For instance, food reviews tend to involve words like tasty, whereas movie reviews commonly contain words such as thrilling for the same positive sentiment. This problem is exacerbated when data are sampled from multiple domains since the dependence between content and style may vary significantly over domains. In this work, we tackle the domain-varying dependence between the content and the style variables inherent in the counterfactual generation task. We provide identification guarantees for such latent-variable models by leveraging the relative sparsity of the influences from different latent variables. Our theoretical insights enable the development of a doMain AdapTive counTerfactual gEneration model, called (MATTE). Our theoretically grounded framework achieves state-of-the-art performance in unsupervised style transfer tasks, where neither paired data nor style labels are utilized, across four large-scale datasets. Code is available at https://github.com/hanqi-qi/Matte.git
Large Vision-Language Models (LVLMs) are susceptible to object hallucinations, an issue in which their generated text contains non-existent objects, greatly limiting their reliability and practicality. Current approaches often rely on the model's token likelihoods or other internal information, instruction tuning on additional datasets, or incorporating complex external tools. We first perform empirical analysis on sentence-level LVLM hallucination, finding that CLIP similarity to the image acts as a stronger and more robust indicator of hallucination compared to token likelihoods. Motivated by this, we introduce our CLIP-Guided Decoding (CGD) approach, a straightforward but effective training-free approach to reduce object hallucination at decoding time. CGD uses CLIP to guide the model's decoding process by enhancing visual grounding of generated text with the image. Experiments demonstrate that CGD effectively mitigates object hallucination across multiple LVLM families while preserving the utility of text generation. Codes are available at https://github.com/d-ailin/CLIP-Guided-Decoding.
Understanding the inner workings of large language models (LLMs) is crucial for advancing their theoretical foundations and real-world applications. While the attention mechanism and multi-layer perceptrons (MLPs) have been studied independently, their interactions remain largely unexplored. This study investigates how attention heads and next-token neurons interact in LLMs to predict new words. We propose a methodology to identify next-token neurons, find prompts that highly activate them, and determine the upstream attention heads responsible. We then generate and evaluate explanations for the activity of these attention heads in an automated manner. Our findings reveal that some attention heads recognize specific contexts relevant to predicting a token and activate a downstream token-predicting neuron accordingly. This mechanism provides a deeper understanding of how attention heads work with MLP neurons to perform next-token prediction. Our approach offers a foundation for further research into the intricate workings of LLMs and their impact on text generation and understanding.
The paper will briefly present the development history of transformer-based language models for the Serbian language. Several new models for text generation and vectorization, trained on the resources of the Society for Language Resources and Technologies, will also be presented. Ten selected vectorization models for Serbian, including two new ones, will be compared on four natural language processing tasks. Paper will analyze which models are the best for each selected task, how does their size and the size of their training sets affect the performance on those tasks, and what is the optimal setting to train the best language models for the Serbian language.
In the domain of Natural Language Processing (NLP), Large Language Models (LLMs) have demonstrated promise in text-generation tasks. However, their educational applications, particularly for domain-specific queries, remain underexplored. This study investigates LLMs' capabilities in educational scenarios, focusing on concept graph recovery and question-answering (QA). We assess LLMs' zero-shot performance in creating domain-specific concept graphs and introduce TutorQA, a new expert-verified NLP-focused benchmark for scientific graph reasoning and QA. TutorQA consists of five tasks with 500 QA pairs. To tackle TutorQA queries, we present CGLLM, a pipeline integrating concept graphs with LLMs for answering diverse questions. Our results indicate that LLMs' zero-shot concept graph recovery is competitive with supervised methods, showing an average 3% F1 score improvement. In TutorQA tasks, LLMs achieve up to 26% F1 score enhancement. Moreover, human evaluation and analysis show that CGLLM generates answers with more fine-grained concepts.
Preference learning (PL) with large language models (LLMs) aims to align the LLMs' generations with human preferences. Previous work on reinforcement learning from human feedback (RLHF) has demonstrated promising results in in-distribution PL. However, due to the difficulty of obtaining human feedback, discretely training reward models for every encountered distribution is challenging. Thus, out-of-distribution (OOD) PL is practically useful for enhancing the generalization ability of LLMs with limited preference feedback. This work addresses OOD PL by optimizing a general reward model through a meta-learning approach. During meta-training, a bilevel optimization algorithm is utilized to learn a reward model capable of guiding policy learning to align with human preferences across various distributions. When encountering a test distribution, the meta-test procedure conducts regularized policy optimization using the learned reward model for PL. We theoretically demonstrate the convergence rate of the bilevel optimization algorithm under reasonable assumptions. Additionally, we conduct experiments on two text generation tasks across 20 held-out domains and outperform a variety of strong baselines across various evaluation metrics.
Large language models (LLMs) may generate text that lacks consistency with human knowledge, leading to factual inaccuracies or \textit{hallucination}. Existing research for evaluating the factuality of LLMs involves extracting fact claims using an LLM and verifying them against a predefined fact source. However, these evaluation metrics are task-specific, and not scalable, and the substitutability of fact sources in different tasks is under-explored. To address these challenges, we categorize four available fact sources: human-written evidence, reference documents, search engine results, and LLM knowledge, along with five text generation tasks containing six representative datasets. Then, we propose \texttt{UFO}, an LLM-based unified and flexible evaluation framework to verify facts against plug-and-play fact sources. We implement five evaluation scenarios based on this framework. Experimental results show that for most QA tasks, human-written evidence and reference documents are crucial, and they can substitute for each other in retrieval-augmented QA tasks. In news fact generation tasks, search engine results and LLM knowledge are essential. Our dataset and code are available at \url{https://github.com/WaldenRUC/UFO}.
The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions (MCQ) to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model's diverse response styles such as starting with "Sure" or refusing to answer. Consequently, MCQ evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output as well and ii) caution against relying solely on first-token evaluation.
As large-scale language models become the standard for text generation, there is a greater need to tailor the generations to be more or less concise, targeted, and informative, depending on the audience/application. Existing control approaches primarily adjust the semantic (e.g., emotion, topics), structural (e.g., syntax tree, parts-of-speech), and lexical (e.g., keyword/phrase inclusion) properties of text, but are insufficient to accomplish complex objectives such as pacing which control the complexity and readability of the text. In this paper, we introduce CEV-LM - a lightweight, semi-autoregressive language model that utilizes constrained edit vectors to control three complementary metrics (speed, volume, and circuitousness) that quantify the shape of text (e.g., pacing of content). We study an extensive set of state-of-the-art CTG models and find that CEV-LM provides significantly more targeted and precise control of these three metrics while preserving semantic content, using less training data, and containing fewer parameters.
Recent diffusion-based generative models show promise in their ability to generate text images, but limitations in specifying the styles of the generated texts render them insufficient in the realm of typographic design. This paper proposes a typographic text generation system to add and modify text on typographic designs while specifying font styles, colors, and text effects. The proposed system is a novel combination of two off-the-shelf methods for diffusion models, ControlNet and Blended Latent Diffusion. The former functions to generate text images under the guidance of edge conditions specifying stroke contours. The latter blends latent noise in Latent Diffusion Models (LDM) to add typographic text naturally onto an existing background. We first show that given appropriate text edges, ControlNet can generate texts in specified fonts while incorporating effects described by prompts. We further introduce text edge manipulation as an intuitive and customizable way to produce texts with complex effects such as ``shadows'' and ``reflections''. Finally, with the proposed system, we successfully add and modify texts on a predefined background while preserving its overall coherence.
Speculative decoding is a widely used method that accelerates the generation process of large language models (LLMs) with no compromise in model performance. It achieves this goal by using an existing smaller model for drafting and then employing the target LLM to verify the draft in a low-cost parallel manner. Under such a drafting-verification framework, drafting efficiency has become a bottleneck in the final speedup of speculative decoding. Therefore, generating longer drafts at less cost can lead to better decoding speedup. To achieve this, we introduce Ouroboros, which can generate draft phrases to parallelize the drafting process and meanwhile lengthen drafts in a training-free manner. The experimental results on various typical text generation tasks show that Ouroboros can achieve speedups of up to $2.8\times$ over speculative decoding and $3.9\times$ over vanilla decoding, without fine-tuning draft and target models. The source code of Ouroboros is available at https://github.com/thunlp/Ouroboros.