Large Language Models (LLMs) have demonstrated exceptional proficiency in instruction-following, making them increasingly integral to various applications. However, this capability introduces the risk of prompt injection attacks, where malicious instructions are embedded in the input to trigger unintended actions or content. Understanding the robustness of LLMs against such attacks is critical for ensuring their safe deployment. In this work, we establish a benchmark to evaluate the robustness of instruction-following LLMs against prompt injection attacks, assessing their ability to discern which instructions to follow and which to disregard. Through extensive experiments with leading instruction-following LLMs, we reveal significant vulnerabilities, particularly in models that mis-follow injected instructions. Our results show that certain models are excessively inclined to prioritize embedded instructions in prompts, often focusing on the latter parts of the prompt without fully understanding the overall context. Conversely, models that exhibit stronger contextual understanding and instruction-following capabilities tend to be more easily compromised by injected instructions. These findings highlight the need to balance improving LLMs’ instruction-following abilities with enhancing their overall comprehension of prompts, to prevent mis-following inappropriate instructions. We hope our analysis provides valuable insights into these vulnerabilities, contributing to the development of more robust solutions in the future.

Public companies in the US are required to publish annual reports that detail their recent financial performance, present the current state of ongoing business operations, and discuss future prospects. However, they typically contain over 25,000 words across all sections, large amounts of industry and legal jargon, and a high percentage of boilerplate content that does not change much year-to-year. These unique characteristics present challenges for many generic pretrained language models because it is likely that only a small percentage of the long report that reflects salient information contains meaningful signal about the future prospects of the company. In this work, we curate a large-scale dataset of paired financial reports and introduce two novel, challenging tasks of predicting long-horizon company risk and correlation that evaluate the ability of the model to recognize cross-document relationships with complex, nuanced signals. We explore and present a comprehensive set of methods and experiments, and establish strong baselines designed to learn to identify subtle similarities and differences between long documents. Furthermore, we demonstrate that it is possible to predict company risk and correlation solely from the text of their financial reports and further that modeling the cross-document interactions at a fine-grained level provides significant benefit. Finally, we probe the best performing model through quantitative and qualitative interpretability methods to reveal some insight into the underlying task signal.

There is a variety of multimodal data pertinent to public companies, spanning from accounting statements, macroeconomic statistics, earnings conference calls, and financial reports. These diverse modalities capture the state of firms from a variety of different perspectives but requires complex interactions to reconcile in the formation of accurate financial predictions. The commonality between these different modalities is that they all represent a time series, typically observed for a particular firm at a quarterly horizon, providing the ability to model trends and variations of company data over time. However, the time series of these diverse modalities contains varying temporal and cross-channel patterns that are challenging to model without the appropriate inductive biases. In this work, we design a novel multimodal time series prediction task that includes numerical financial results, macroeconomic states, and long financial documents to predict next quarter’s company earnings relative to analyst expectations. We explore a variety of approaches for this novel setting, establish strong unimodal baselines, and propose a multimodal model that exhibits state-of-the-art performance on this unique task. We demonstrate that each modality contains unique information and that the best performing model requires careful fusion of the different modalities in a multi-stage training approach. To better understand model behavior, we conduct a variety of probing experiments, reveal insights into the value of different modalities, and demonstrate the practical utility of our proposed method in a simulated trading setting.

As a type of figurative language, an East Asian idiom condenses rich cultural background into only a few characters. Translating such idioms is challenging for human translators, who often resort to choosing a context-aware translation from an existing list of candidates. However, compiling a dictionary of candidate translations demands much time and creativity even for expert translators. To alleviate such burden, we evaluate if GPT-4 can help generate high-quality translations. Based on automatic evaluations of faithfulness and creativity, we first identify Pareto-optimal prompting strategies that can outperform translation engines from Google and DeepL. Then, at a low cost, our context-aware translations can achieve far more high-quality translations per idiom than the human baseline. We open-source all code and data to facilitate further research.

As a type of figurative language, an East Asian idiom condenses rich cultural background into only a few characters. Translating such idioms is challenging for human translators, who often resort to choosing a context-aware translation from an existing list of candidates. However, compiling a dictionary of candidate translations demands much time and creativity even for expert translators. To alleviate such burden, we evaluate if GPT-4 can help generate high-quality translations. Based on automatic evaluations of faithfulness and creativity, we first identify Pareto-optimal prompting strategies that can outperform translation engines from Google and DeepL. Then, at a low cost, our context-aware translations can achieve far more high-quality translations per idiom than the human baseline. We open-source all code and data to facilitate further research.

Large language models (LLMs) are increasingly prevalent in conversational systems due to their advanced understanding and generative capabilities in general contexts. However, their effectiveness in task-oriented dialogues (TOD), which requires not only response generation but also effective dialogue state tracking (DST) within specific tasks and domains, remains less satisfying. In this work, we propose a novel approach FnCTOD for solving DST with LLMs through function calling. This method improves zero-shot DST, allowing adaptation to diverse domains without extensive data collection or model tuning. Our experimental results demonstrate that our approach achieves exceptional performance with both modestly sized open-source and also proprietary LLMs: with in-context prompting it enables various 7B or 13B parameter models to surpass the previous state-of-the-art (SOTA) achieved by ChatGPT, and improves ChatGPT’s performance beating the SOTA by 5.6% average joint goal accuracy (JGA). Individual model results for GPT-3.5 and GPT-4 are boosted by 4.8% and 14%, respectively. We also show that by fine-tuning on a small collection of diverse task-oriented dialogues, we can equip modestly sized models, specifically a 13B parameter LLaMA2-Chat model, with function-calling capabilities and DST performance comparable to ChatGPT while maintaining their chat capabilities. We have made the code publicly available at https://github.com/facebookresearch/FnCTOD.

Recent work has shown that large pretrained Language Models (LMs) can not only perform remarkably well on a range of Natural Language Processing (NLP) tasks but also start improving on reasoning tasks such as arithmetic induction, symbolic manipulation, and commonsense reasoning with increasing size of models. However, it is still unclear what the underlying capabilities of these LMs are. Surprisingly, we find that these models have limitations on certain basic symbolic manipulation tasks such as copy, reverse, and addition. When the total number of symbols or repeating symbols increases, the model performance drops quickly. We investigate the potential causes behind this phenomenon and examine a set of possible methods, including explicit positional markers, fine-grained computation steps, and LMs with callable programs. Experimental results show that none of these techniques can solve the simplest addition induction problem completely. In the end, we introduce LMs with tutor, which demonstrates every single step of teaching. LMs with tutor is able to deliver 100% accuracy in situations of OOD and repeating symbols, shedding new insights on the boundary of large LMs in induction.

Answering complex questions often requires reasoning over knowledge graphs (KGs). State-of-the-art methods often utilize entities in questions to retrieve local subgraphs, which are then fed into KG encoder, e.g. graph neural networks (GNNs), to model their local structures and integrated into language models for question answering. However, this paradigm constrains retrieved knowledge in local subgraphs and discards more diverse triplets buried in KGs that are disconnected but useful for question answering. In this paper, we propose a simple yet effective method to first retrieve the most relevant triplets from KGs and then rerank them, which are then concatenated with questions to be fed into language models. Extensive results on both CommonsenseQA and OpenbookQA datasets show that our method can outperform state-of-the-art up to 4.6% absolute accuracy.

There is a multitude of textual data relevant to the financial markets, spanning genres such as financial news, earnings conference calls, and social media posts. Earnings conference calls are one of the most important to information flow as they reflect a direct communication between company executives, financial analysts, and large shareholders. Since these calls contain content that is forward-looking in nature, they can be used to forecast the future performance of the company relative to market expectations. However, they typically contain over 5,000 words of text and large amounts of industry jargon. This length and domain-specific language present problems for many generic pretrained language models. In this work, we introduce a novel task of predicting earnings surprises from earnings call transcripts and contribute a new long document dataset that tests financial understanding with complex signals. We explore a variety of approaches for this long document classification task and establish some strong baselines. Furthermore, we demonstrate that it is possible to predict companies’ future earnings surprises from solely the text of their conference calls with reasonable accuracy. Finally, we probe the models through different interpretability methods and reveal some intuitive explanations of the linguistic features captured that go beyond traditional sentiment analysis.

Building dialogue systems requires a large corpus of annotated dialogues. Such datasets are usually created via crowdsourcing, which is expensive and time-consuming. In this paper, we propose Dialogic, a novel dialogue simulation method based on large language model in-context learning to automate dataset creation. Seeded with a few annotated dialogues, Dialogic automatically selects in-context examples for demonstration and prompts GPT-3 to generate new dialogues and annotations in a controllable way. Our method can rapidly expand a small set of dialogue data with minimum or zero human involvement and parameter update and is thus much more cost-efficient and time-saving than crowdsourcing. Experimental results on the MultiWOZ dataset demonstrate that training a model on the simulated dialogues leads to even better performance than using the same amount of human-generated dialogues under the challenging low-resource settings, with as few as 85 dialogues as a seed. When the full training set is given, our method can still serve as an effective data augmentation method to further improve performance. Human evaluation results also show that our simulated dialogues have near-human fluency and annotation accuracy. The code and data are available at https://github.com/Leezekun/dialogic.

To reduce the toxic degeneration in a pretrained Language Model (LM), previous work on Language Model detoxification has focused on reducing the toxicity of the generation itself (self-toxicity) without consideration of the context. As a result, a type of implicit offensive language where the generations support the offensive language in the context is ignored. Different from the LM controlling tasks in previous work, where the desired attributes are fixed for generation, the desired stance of the generation depends on the offensiveness of the context. Therefore, we propose a novel control method to do context-dependent detoxification with the stance taken into consideration. We introduce meta prefixes to learn the contextualized stance control strategy and to generate the stance control prefix according to the input context. The generated stance prefix is then combined with the toxicity control prefix to guide the response generation. Experimental results show that our proposed method can effectively learn the context-dependent stance control strategies while keeping a low self-toxicity of the underlying LM.

Existing work on automated hate speech classification assumes that the dataset is fixed and the classes are pre-defined. However, the amount of data in social media increases every day, and the hot topics changes rapidly, requiring the classifiers to be able to continuously adapt to new data without forgetting the previously learned knowledge. This ability, referred to as lifelong learning, is crucial for the real-word application of hate speech classifiers in social media. In this work, we propose lifelong learning of hate speech classification on social media. To alleviate catastrophic forgetting, we propose to use Variational Representation Learning (VRL) along with a memory module based on LB-SOINN (Load-Balancing Self-Organizing Incremental Neural Network). Experimentally, we show that combining variational representation learning and the LB-SOINN memory module achieves better performance than the commonly-used lifelong learning techniques.

Task-adaptive pre-training (TAPT) and Self-training (ST) have emerged as the major semi-supervised approaches to improve natural language understanding (NLU) tasks with massive amount of unlabeled data. However, it’s unclear whether they learn similar representations or they can be effectively combined. In this paper, we show that TAPT and ST can be complementary with simple TFS protocol by following TAPT -> Finetuning -> Self-training (TFS) process. Experimental results show that TFS protocol can effectively utilize unlabeled data to achieve strong combined gains consistently across six datasets covering sentiment classification, paraphrase identification, natural language inference, named entity recognition and dialogue slot classification. We investigate various semi-supervised settings and consistently show that gains from TAPT and ST can be strongly additive by following TFS procedure. We hope that TFS could serve as an important semi-supervised baseline for future NLP studies.

Data-to-text generation has recently attracted substantial interests due to its wide applications. Existing methods have shown impressive performance on an array of tasks. However, they rely on a significant amount of labeled data for each task, which is costly to acquire and thus limits their application to new tasks and domains. In this paper, we propose to leverage pre-training and transfer learning to address this issue. We propose a knowledge-grounded pre-training (KGPT), which consists of two parts, 1) a general knowledge-grounded generation model to generate knowledge-enriched text. 2) a pre-training paradigm on a massive knowledge-grounded text corpus crawled from the web. The pre-trained model can be fine-tuned on various data-to-text generation tasks to generate task-specific text. We adopt three settings, namely fully-supervised, zero-shot, few-shot to evaluate its effectiveness. Under the fully-supervised setting, our model can achieve remarkable gains over the known baselines. Under zero-shot setting, our model without seeing any examples achieves over 30 ROUGE-L on WebNLG while all other baselines fail. Under the few-shot setting, our model only needs about one-fifteenth as many labeled examples to achieve the same level of performance as baseline models. These experiments consistently prove the strong generalization ability of our proposed framework.

Pre-trained embeddings such as word embeddings and sentence embeddings are fundamental tools facilitating a wide range of downstream NLP tasks. In this work, we investigate how to learn a general-purpose embedding of textual relations, defined as the shortest dependency path between entities. Textual relation embedding provides a level of knowledge between word/phrase level and sentence level, and we show that it can facilitate downstream tasks requiring relational understanding of the text. To learn such an embedding, we create the largest distant supervision dataset by linking the entire English ClueWeb09 corpus to Freebase. We use global co-occurrence statistics between textual and knowledge base relations as the supervision signal to train the embedding. Evaluation on two relational understanding tasks demonstrates the usefulness of the learned textual relation embedding. The data and code can be found at https://github.com/czyssrs/GloREPlus

Semantically controlled neural response generation on limited-domain has achieved great performance. However, moving towards multi-domain large-scale scenarios are shown to be difficult because the possible combinations of semantic inputs grow exponentially with the number of domains. To alleviate such scalability issue, we exploit the structure of dialog acts to build a multi-layer hierarchical graph, where each act is represented as a root-to-leaf route on the graph. Then, we incorporate such graph structure prior as an inductive bias to build a hierarchical disentangled self-attention network, where we disentangle attention heads to model designated nodes on the dialog act graph. By activating different (disentangled) heads at each layer, combinatorially many dialog act semantics can be modeled to control the neural response generation. On the large-scale Multi-Domain-WOZ dataset, our model can yield a significant improvement over the baselines on various automatic and human evaluation metrics.

With the rapid development in deep learning, deep neural networks have been widely adopted in many real-life natural language applications. Under deep neural networks, a pre-defined vocabulary is required to vectorize text inputs. The canonical approach to select pre-defined vocabulary is based on the word frequency, where a threshold is selected to cut off the long tail distribution. However, we observed that such a simple approach could easily lead to under-sized vocabulary or over-sized vocabulary issues. Therefore, we are interested in understanding how the end-task classification accuracy is related to the vocabulary size and what is the minimum required vocabulary size to achieve a specific performance. In this paper, we provide a more sophisticated variational vocabulary dropout (VVD) based on variational dropout to perform vocabulary selection, which can intelligently select the subset of the vocabulary to achieve the required performance. To evaluate different algorithms on the newly proposed vocabulary selection problem, we propose two new metrics: Area Under Accuracy-Vocab Curve and Vocab Size under X% Accuracy Drop. Through extensive experiments on various NLP classification tasks, our variational framework is shown to significantly outperform the frequency-based and other selection baselines on these metrics.

We study the problem of textual relation embedding with distant supervision. To combat the wrong labeling problem of distant supervision, we propose to embed textual relations with global statistics of relations, i.e., the co-occurrence statistics of textual and knowledge base relations collected from the entire corpus. This approach turns out to be more robust to the training noise introduced by distant supervision. On a popular relation extraction dataset, we show that the learned textual relation embedding can be used to augment existing relation extraction models and significantly improve their performance. Most remarkably, for the top 1,000 relational facts discovered by the best existing model, the precision can be improved from 83.9% to 89.3%.

Inferring missing links in knowledge graphs (KG) has attracted a lot of attention from the research community. In this paper, we tackle a practical query answering task involving predicting the relation of a given entity pair. We frame this prediction problem as an inference problem in a probabilistic graphical model and aim at resolving it from a variational inference perspective. In order to model the relation between the query entity pair, we assume that there exists an underlying latent variable (paths connecting two nodes) in the KG, which carries the equivalent semantics of their relations. However, due to the intractability of connections in large KGs, we propose to use variation inference to maximize the evidence lower bound. More specifically, our framework (Diva) is composed of three modules, i.e. a posterior approximator, a prior (path finder), and a likelihood (path reasoner). By using variational inference, we are able to incorporate them closely into a unified architecture and jointly optimize them to perform KG reasoning. With active interactions among these sub-modules, Diva is better at handling noise and coping with more complex reasoning scenarios. In order to evaluate our method, we conduct the experiment of the link prediction task on multiple datasets and achieve state-of-the-art performances on both datasets.

The recent advance in deep learning and semantic parsing has significantly improved the translation accuracy of natural language questions to structured queries. However, further improvement of the existing approaches turns out to be quite challenging. Rather than solely relying on algorithmic innovations, in this work, we introduce DialSQL, a dialogue-based structured query generation framework that leverages human intelligence to boost the performance of existing algorithms via user interaction. DialSQL is capable of identifying potential errors in a generated SQL query and asking users for validation via simple multi-choice questions. User feedback is then leveraged to revise the query. We design a generic simulator to bootstrap synthetic training dialogues and evaluate the performance of DialSQL on the WikiSQL dataset. Using SQLNet as a black box query generation tool, DialSQL improves its performance from 61.3% to 69.0% using only 2.4 validation questions per dialogue.

Task-oriented dialog systems are becoming pervasive, and many companies heavily rely on them to complement human agents for customer service in call centers. With globalization, the need for providing cross-lingual customer support becomes more urgent than ever. However, cross-lingual support poses great challenges—it requires a large amount of additional annotated data from native speakers. In order to bypass the expensive human annotation and achieve the first step towards the ultimate goal of building a universal dialog system, we set out to build a cross-lingual state tracking framework. Specifically, we assume that there exists a source language with dialog belief tracking annotations while the target languages have no annotated dialog data of any form. Then, we pre-train a state tracker for the source language as a teacher, which is able to exploit easy-to-access parallel data. We then distill and transfer its own knowledge to the student state tracker in target languages. We specifically discuss two types of common parallel resources: bilingual corpus and bilingual dictionary, and design different transfer learning strategies accordingly. Experimentally, we successfully use English state tracker as the teacher to transfer its knowledge to both Italian and German trackers and achieve promising results.

WikiSQL is a newly released dataset for studying the natural language sequence to SQL translation problem. The SQL queries in WikiSQL are simple: Each involves one relation and does not have any join operation. Despite of its simplicity, none of the publicly reported structured query generation models can achieve an accuracy beyond 62%, which is still far from enough for practical use. In this paper, we ask two questions, “Why is the accuracy still low for such simple queries?” and “What does it take to achieve 100% accuracy on WikiSQL?” To limit the scope of our study, we focus on the WHERE clause in SQL. The answers will help us gain insights about the directions we should explore in order to further improve the translation accuracy. We will then investigate alternative solutions to realize the potential ceiling performance on WikiSQL. Our proposed solution can reach up to 88.6% condition accuracy on the WikiSQL dataset.

The existing factoid QA systems often lack a post-inspection component that can help models recover from their own mistakes. In this work, we propose to crosscheck the corresponding KB relations behind the predicted answers and identify potential inconsistencies. Instead of developing a new model that accepts evidences collected from these relations, we choose to plug them back to the original questions directly and check if the revised question makes sense or not. A bidirectional LSTM is applied to encode revised questions. We develop a scoring mechanism over the revised question encodings to refine the predictions of a base QA system. This approach can improve the F1 score of STAGG (Yih et al., 2015), one of the leading QA systems, from 52.5% to 53.9% on WEBQUESTIONS data.

Existing studies on semantic parsing mainly focus on the in-domain setting. We formulate cross-domain semantic parsing as a domain adaptation problem: train a semantic parser on some source domains and then adapt it to the target domain. Due to the diversity of logical forms in different domains, this problem presents unique and intriguing challenges. By converting logical forms into canonical utterances in natural language, we reduce semantic parsing to paraphrasing, and develop an attentive sequence-to-sequence paraphrase model that is general and flexible to adapt to different domains. We discover two problems, small micro variance and large macro variance, of pre-trained word embeddings that hinder their direct use in neural networks, and propose standardization techniques as a remedy. On the popular Overnight dataset, which contains eight domains, we show that both cross-domain training and standardized pre-trained word embeddings can bring significant improvement.