Computation and Language (cs.CL)

Abstract: As AI-generated text increasingly resembles human-written content, the ability to detect machine-generated text becomes crucial. To address this challenge, we present GPTWatermark, a robust and high-quality solution designed to ascertain whether a piece of text originates from a specific model. Our approach extends existing watermarking strategies and employs a fixed group design to enhance robustness against editing and paraphrasing attacks. We show that our watermarked language model enjoys strong provable guarantees on generation quality, correctness in detection, and security against evasion attacks. Experimental results on various large language models (LLMs) and diverse datasets demonstrate that our method achieves superior detection accuracy and comparable generation quality in perplexity, thus promoting the responsible use of LLMs.

Abstract: Chinese Text Error Correction (CTEC) aims to detect and correct errors in the input text, which benefits human's daily life and various downstream tasks. Recent approaches mainly employ Pre-trained Language Models (PLMs) to resolve CTEC task and achieve tremendous success. However, previous approaches suffer from issues of over-correction and under-correction, and the former is especially conspicuous in the precision-critical CTEC task. To mitigate the issue of overcorrection, we propose a novel model-agnostic progressive multitask learning framework for CTEC, named ProTEC, which guides a CTEC model to learn the task from easy to difficult. We divide CTEC task into three sub-tasks from easy to difficult: Error Detection, Error Type Identification, and Correction Result Generation. During the training process, ProTEC guides the model to learn text error correction progressively by incorporating these sub-tasks into a multi-task training objective. During the inference process, the model completes these sub-tasks in turn to generate the correction results. Extensive experiments and detailed analyses fully demonstrate the effectiveness and efficiency of our proposed framework.

Abstract: Despite their impressive scale, knowledge bases (KBs), such as Wikidata, still contain significant gaps. Language models (LMs) have been proposed as a source for filling these gaps. However, prior works have focused on prominent entities with rich coverage by LMs, neglecting the crucial case of long-tail entities. In this paper, we present a novel method for LM-based-KB completion that is specifically geared for facts about long-tail entities. The method leverages two different LMs in two stages: for candidate retrieval and for candidate verification and disambiguation. To evaluate our method and various baselines, we introduce a novel dataset, called MALT, rooted in Wikidata. Our method outperforms all baselines in F1, with major gains especially in recall.

Abstract: Large language models (LLMs) often contain misleading content, emphasizing the need to align them with human values to ensure secur AI systems. Reinforcement learning from human feedback (RLHF) has been employed to achieve this alignment by combining a reward model, typically based on Bradley-Terry paired comparison, with an RL algorithm such as Proximal Policy Optimization (PPO) to optimize LLM responses. However, RLHF exhibits complexity, instability, and sensitivity to hyperparameters. In this paper, we propose Preference Ranking Optimization (PRO) as an alternative to PPO for directly aligning LLMs with the Bradley-Terry comparison. PRO extends the pairwise Bradley-Terry comparison to accommodate preference rankings of any length. By iteratively contrasting the likelihood of generating responses, PRO instructs the LLM to prioritize the best response while progressively ranking the remaining responses. In this manner, PRO effectively transforms human alignment into aligning the probability ranking of $n$ responses generated by LLM with the preference ranking of humans towards these responses. Experiments have shown that PRO outperforms existing alignment algorithms, achieving comparable results to ChatGPT and human responses through automatic-based, reward-based, GPT-4, and human evaluations. Furthermore, we demonstrate that longer, more diverse, and higher-quality preference ranking sequences can consistently enhance the performance of human alignment.

Abstract: Relation extraction (RE) is a crucial task in natural language processing (NLP) that aims to identify and classify relationships between entities mentioned in text. In the financial domain, relation extraction plays a vital role in extracting valuable information from financial documents, such as news articles, earnings reports, and company filings. This paper describes our solution to relation extraction on one such dataset REFinD. The dataset was released along with shared task as a part of the Fourth Workshop on Knowledge Discovery from Unstructured Data in Financial Services, co-located with SIGIR 2023. In this paper, we employed OpenAI models under the framework of in-context learning (ICL). We utilized two retrieval strategies to find top K relevant in-context learning demonstrations / examples from training data for a given test example. The first retrieval mechanism, we employed, is a learning-free dense retriever and the other system is a learning-based retriever. We were able to achieve 4th rank on the leaderboard. Our best F1-score is 0.718.

Abstract: Depression is a growing issue in society's mental health that affects all areas of life and can even lead to suicide. Fortunately, prevention programs can be effective in its treatment. In this context, this work proposes an automatic system for detecting depression on social media based on machine learning and natural language processing methods. This paper presents the following contributions: (i) an ensemble learning system that combines several types of text representations for depression detection, including recent advances in the field; (ii) a contextualization schema through topic and affective information; (iii) an analysis of models' energy consumption, establishing a trade-off between classification performance and overall computational costs. To assess the proposed models' effectiveness, a thorough evaluation is performed in two datasets that model depressive text. Experiments indicate that the proposed contextualization strategies can improve the classification and that approaches that use Transformers can improve the overall F-score by 2% while augmenting the energy cost a hundred times. Finally, this work paves the way for future energy-wise systems by considering both the performance classification and the energy consumption.

Abstract: University admission at many highly selective institutions uses a holistic review process, where all aspects of the application, including protected attributes (e.g., race, gender), grades, essays, and recommendation letters are considered, to compose an excellent and diverse class. In this study, we empirically evaluate how influential protected attributes are for predicting admission decisions using a machine learning (ML) model, and in how far textual information (e.g., personal essay, teacher recommendation) may substitute for the loss of protected attributes in the model. Using data from 14,915 applicants to an undergraduate admission office at a selective U.S. institution in the 2022-2023 cycle, we find that the exclusion of protected attributes from the ML model leads to substantially reduced admission-prediction performance. The inclusion of textual information via both a TF-IDF representation and a Latent Dirichlet allocation (LDA) model partially restores model performance, but does not appear to provide a full substitute for admitting a similarly diverse class. In particular, while the text helps with gender diversity, the proportion of URM applicants is severely impacted by the exclusion of protected attributes, and the inclusion of new attributes generated from the textual information does not recover this performance loss.

Abstract: The NL2SQL task involves parsing natural language statements into SQL queries. While most state-of-the-art methods treat NL2SQL as a slot-filling task and use feature representation learning techniques, they overlook explicit correlation features between the SELECT and WHERE clauses and implicit correlation features between sub-tasks within a single clause. To address this issue, we propose the Clause Feature Correlation Decoupling and Coupling (CFCDC) model, which uses a feature representation decoupling method to separate the SELECT and WHERE clauses at the parameter level. Next, we introduce a multi-task learning architecture to decouple implicit correlation feature representation between different SQL tasks in a specific clause. Moreover, we present an improved feature representation coupling module to integrate the decoupled tasks in the SELECT and WHERE clauses and predict the final SQL query. Our proposed CFCDC model demonstrates excellent performance on the WikiSQL dataset, with significant improvements in logic precision and execution accuracy. The source code for the model will be publicly available on GitHub

Abstract: As opposed to general English, many concepts in biomedical terminology have been designed in recent history by biomedical professionals with the goal of being precise and concise. This is often achieved by concatenating meaningful biomedical morphemes to create new semantic units. Nevertheless, most modern biomedical language models (LMs) are pre-trained using standard domain-specific tokenizers derived from large scale biomedical corpus statistics without explicitly leveraging the agglutinating nature of biomedical language. In this work, we first find that standard open-domain and biomedical tokenizers are largely unable to segment biomedical terms into meaningful components. Therefore, we hypothesize that using a tokenizer which segments biomedical terminology more accurately would enable biomedical LMs to improve their performance on downstream biomedical NLP tasks, especially ones which involve biomedical terms directly such as named entity recognition (NER) and entity linking. Surprisingly, we find that pre-training a biomedical LM using a more accurate biomedical tokenizer does not improve the entity representation quality of a language model as measured by several intrinsic and extrinsic measures such as masked language modeling prediction (MLM) accuracy as well as NER and entity linking performance. These quantitative findings, along with a case study which explores entity representation quality more directly, suggest that the biomedical pre-training process is quite robust to instances of sub-optimal tokenization.

Abstract: Task-oriented dialogue research has mainly focused on a few popular languages like English and Chinese, due to the high dataset creation cost for a new language. To reduce the cost, we apply manual editing to automatically translated data. We create a new multilingual benchmark, X-RiSAWOZ, by translating the Chinese RiSAWOZ to 4 languages: English, French, Hindi, Korean; and a code-mixed English-Hindi language. X-RiSAWOZ has more than 18,000 human-verified dialogue utterances for each language, and unlike most multilingual prior work, is an end-to-end dataset for building fully-functioning agents. The many difficulties we encountered in creating X-RiSAWOZ led us to develop a toolset to accelerate the post-editing of a new language dataset after translation. This toolset improves machine translation with a hybrid entity alignment technique that combines neural with dictionary-based methods, along with many automated and semi-automated validation checks. We establish strong baselines for X-RiSAWOZ by training dialogue agents in the zero- and few-shot settings where limited gold data is available in the target language. Our results suggest that our translation and post-editing methodology and toolset can be used to create new high-quality multilingual dialogue agents cost-effectively. Our dataset, code, and toolkit are released open-source.

Abstract: The ability to conduct retrospective analyses of attacks on human rights defenders over time and by location is important for humanitarian organizations to better understand historical or ongoing human rights violations and thus better manage the global impact of such events. We hypothesize that NLP can support such efforts by quickly processing large collections of news articles to detect and summarize the characteristics of attacks on human rights defenders. To that end, we propose a new dataset for detecting Attacks on Human Rights Defenders (HRDsAttack) consisting of crowdsourced annotations on 500 online news articles. The annotations include fine-grained information about the type and location of the attacks, as well as information about the victim(s). We demonstrate the usefulness of the dataset by using it to train and evaluate baseline models on several sub-tasks to predict the annotated characteristics.

Abstract: The Natural Language to SQL (NL2SQL) technique is used to convert natural language queries into executable SQL statements. Typically, slot-filling is employed as a classification method for multi-task cases to achieve this goal. However, slot-filling can result in inaccurate SQL statement generation due to negative migration issues arising from different classification tasks. To overcome this limitation, this study introduces a new approach called Multi-Layer Expert Generate SQL (MLEG-SQL), which utilizes a dedicated multi-task hierarchical network. The lower layer of the network extracts semantic features of natural language statements, while the upper layer builds a specialized expert system for handling specific classification tasks. This hierarchical approach mitigates performance degradation resulting from different task conflicts. The proposed method was evaluated on the WiKSQL dataset and was found to be effective in generating accurate SQL statements.

Abstract: Document-level event extraction is a long-standing challenging information retrieval problem involving a sequence of sub-tasks: entity extraction, event type judgment, and event type-specific multi-event extraction. However, addressing the problem as multiple learning tasks leads to increased model complexity. Also, existing methods insufficiently utilize the correlation of entities crossing different events, resulting in limited event extraction performance. This paper introduces a novel framework for document-level event extraction, incorporating a new data structure called token-event-role and a multi-channel argument role prediction module. The proposed data structure enables our model to uncover the primary role of tokens in multiple events, facilitating a more comprehensive understanding of event relationships. By leveraging the multi-channel prediction module, we transform entity and multi-event extraction into a single task of predicting token-event pairs, thereby reducing the overall parameter size and enhancing model efficiency. The results demonstrate that our approach outperforms the state-of-the-art method by 9.5 percentage points in terms of the F1 score, highlighting its superior performance in event extraction. Furthermore, an ablation study confirms the significant value of the proposed data structure in improving event extraction tasks, further validating its importance in enhancing the overall performance of the framework.

Abstract: Autoregressive language models (LMs) map token sequences to probabilities. The usual practice for computing the probability of any character string (e.g. English sentences) is to first transform it into a sequence of tokens that is scored by the model. However, there are exponentially many token sequences that represent any given string. To truly compute the probability of a string one should marginalize over all tokenizations, which is typically intractable. Here, we analyze whether the practice of ignoring the marginalization is justified. To this end, we devise an importance-sampling-based algorithm that allows us to compute estimates of the marginal probabilities and compare them to the default procedure in a range of state-of-the-art models and datasets. Our results show that the gap in log-likelihood is no larger than 0.5% in most cases, but that it becomes more pronounced for data with long complex words.

Abstract: With the rise of bidirectional encoder representations from Transformer models in natural language processing, the speech community has adopted some of their development methodologies. Therefore, the Wav2Vec models were introduced to reduce the data required to obtain state-of-the-art results. This work leverages this knowledge and improves the performance of the pre-trained speech models by simply replacing the fine-tuning dense layer with a lateral inhibition layer inspired by the biological process. Our experiments on Romanian, a low-resource language, show an average improvement of 12.5% word error rate (WER) using the lateral inhibition layer. In addition, we obtain state-of-the-art results on both the Romanian Speech Corpus and the Robin Technical Acquisition Corpus with 1.78% WER and 29.64% WER, respectively.

Abstract: Classifier-Free Guidance (CFG) has recently emerged in text-to-image generation as a lightweight technique to encourage prompt-adherence in generations. In this work, we demonstrate that CFG can be used broadly as an inference-time technique in pure language modeling. We show that CFG (1) improves the performance of Pythia, GPT-2 and LLaMA-family models across an array of tasks: Q\&A, reasoning, code generation, and machine translation, achieving SOTA on LAMBADA with LLaMA-7B over PaLM-540B; (2) brings improvements equivalent to a model with twice the parameter-count; (3) can stack alongside other inference-time methods like Chain-of-Thought and Self-Consistency, yielding further improvements in difficult tasks; (4) can be used to increase the faithfulness and coherence of assistants in challenging form-driven and content-driven prompts: in a human evaluation we show a 75\% preference for GPT4All using CFG over baseline.

Abstract: A diversity of tasks use language models trained on semantic similarity data. While there are a variety of datasets that capture semantic similarity, they are either constructed from modern web data or are relatively small datasets created in the past decade by human annotators. This study utilizes a novel source, newly digitized articles from off-copyright, local U.S. newspapers, to assemble a massive-scale semantic similarity dataset spanning 70 years from 1920 to 1989 and containing nearly 400M positive semantic similarity pairs. Historically, around half of articles in U.S. local newspapers came from newswires like the Associated Press. While local papers reproduced articles from the newswire, they wrote their own headlines, which form abstractive summaries of the associated articles. We associate articles and their headlines by exploiting document layouts and language understanding. We then use deep neural methods to detect which articles are from the same underlying source, in the presence of substantial noise and abridgement. The headlines of reproduced articles form positive semantic similarity pairs. The resulting publicly available HEADLINES dataset is significantly larger than most existing semantic similarity datasets and covers a much longer span of time. It will facilitate the application of contrastively trained semantic similarity models to a variety of tasks, including the study of semantic change across space and time.

Abstract: Symbolization methods in large language models (LLMs) have been shown effective to improve LLMs' reasoning ability. However, most of these approaches hinge on mapping natural languages to formal languages (e.g., Python, SQL) that are more syntactically complete and free of ambiguity. Although effective, they depart from the natural language itself and deviate from the habits of human thinking, and instead cater more to the execution mindset of computers. In contrast, we hope to simplify natural language by starting from the concept of symbols in linguistics itself, so that LLMs can learn the common formulation and general solution of reasoning problems wrapped in different natural semantics. From this consideration, we propose \textbf{Meta-Reasoning}, which allows LLMs to automatically accomplish semantic-symbol deconstruction, i.e., semantic resolution, to maximally reduce different questions of certain reasoning tasks to similar natural language representation, thus gaining the ability to learn by analogy and facilitating data-efficient in-context learning. Our experiments show that the Meta-Reasoning paradigm saliently enhances LLMs' reasoning performance with fewer demonstrations. They can learn not only reasoning chains but also general solutions to certain types of tasks. In particular, for symbolic reasoning tasks, such as 7-step Tracking Shuffled Objects, GPT-3 (text-davinci-002) achieves over 99% accuracy with only one Meta-Reasoning demonstration, outperforming all current LLMs with the standard chain-of-thought prompting.

Abstract: Large Language Models (LLMs) are trained primarily on minimally processed web text, which exhibits the same wide range of social biases held by the humans who created that content. Consequently, text generated by LLMs can inadvertently perpetuate stereotypes towards marginalized groups, like the LGBTQIA+ community. In this paper, we perform a comparative study of how LLMs generate text describing people with different sexual identities. Analyzing bias in the text generated by an LLM using regard score shows measurable bias against queer people. We then show that a post-hoc method based on chain-of-thought prompting using SHAP analysis can increase the regard of the sentence, representing a promising approach towards debiasing the output of LLMs in this setting.

Abstract: An essential aspect of prioritizing incident tickets for resolution is efficiently labeling tickets with fine-grained categories. However, ticket data is often complex and poses several unique challenges for modern machine learning methods: (1) tickets are created and updated either by machines with pre-defined algorithms or by engineers with domain expertise that share different protocols, (2) tickets receive frequent revisions that update ticket status by modifying all or parts of ticket descriptions, and (3) ticket labeling is time-sensitive and requires knowledge updates and new labels per the rapid software and hardware improvement lifecycle. To handle these issues, we introduce Ticket- BERT which trains a simple yet robust language model for labeling tickets using our proposed ticket datasets. Experiments demonstrate the superiority of Ticket-BERT over baselines and state-of-the-art text classifiers on Azure Cognitive Services. We further encapsulate Ticket-BERT with an active learning cycle and deploy it on the Microsoft IcM system, which enables the model to quickly finetune on newly-collected tickets with a few annotations.

Abstract: Large language models show impressive results on few-shot NLP tasks. However, these models are memory and computation-intensive. Meta-training allows one to leverage smaller models for few-shot generalization in a domain-general and task-agnostic manner; however, these methods alone results in models that may not have sufficient parameterization or knowledge to adapt quickly to a large variety of tasks. To overcome this issue, we propose meta-training with demonstration retrieval, where we use a dense passage retriever to retrieve semantically similar labeled demonstrations to each example for more varied supervision. By separating external knowledge from model parameters, we can use meta-training to train parameter-efficient models that generalize well on a larger variety of tasks. We construct a meta-training set from UnifiedQA and CrossFit, and propose a demonstration bank based on UnifiedQA tasks. To our knowledge, our work is the first to combine retrieval with meta-training, to use DPR models to retrieve demonstrations, and to leverage demonstrations from many tasks simultaneously, rather than randomly sampling demonstrations from the training set of the target task. Our approach outperforms a variety of targeted parameter-efficient and retrieval-augmented few-shot methods on QA, NLI, and text classification tasks (including SQuAD, QNLI, and TREC). Our approach can be meta-trained and fine-tuned quickly on a single GPU.

Abstract: Information extraction (IE) plays very important role in natural language processing (NLP) and is fundamental to many NLP applications that used to extract structured information from unstructured text data. Heuristic-based searching and data-driven learning are two main stream implementation approaches. However, no much attention has been paid to document genre and length influence on IE tasks. To fill the gap, in this study, we investigated the accuracy and generalization abilities of heuristic-based searching and data-driven to perform two IE tasks: named entity recognition (NER) and semantic role labeling (SRL) on domain-specific and generic documents with different length. We posited two hypotheses: first, short documents may yield better accuracy results compared to long documents; second, generic documents may exhibit superior extraction outcomes relative to domain-dependent documents due to training document genre limitations. Our findings reveals that no single method demonstrated overwhelming performance in both tasks. For named entity extraction, data-driven approaches outperformed symbolic methods in terms of accuracy, particularly in short texts. In the case of semantic roles extraction, we observed that heuristic-based searching method and data-driven based model with syntax representation surpassed the performance of pure data-driven approach which only consider semantic information. Additionally, we discovered that different semantic roles exhibited varying accuracy levels with the same method. This study offers valuable insights for downstream text mining tasks, such as NER and SRL, when addressing various document features and genres.

Abstract: Sentence-level relation extraction (RE) aims to identify the relationship between 2 entities given a contextual sentence. While there have been many attempts to solve this problem, the current solutions have a lot of room to improve. In this paper, we approach the task of relationship extraction in the financial dataset REFinD. Our approach incorporates typed entity markers representations and various models finetuned on the dataset, which has allowed us to achieve an F1 score of 69.65% on the validation set. Through this paper, we discuss various approaches and possible limitations.

Abstract: We study the ability of transformer-based language models (LMs) to understand social media language. Social media (SM) language is distinct from standard written language, yet existing benchmarks fall short of capturing LM performance in this socially, economically, and politically important domain. We quantify the degree to which social media language differs from conventional language and conclude that the difference is significant both in terms of token distribution and rate of linguistic shift. Next, we introduce a new benchmark for Social MedIa Language Evaluation (SMILE) that covers four SM platforms and eleven tasks. Finally, we show that learning a tokenizer and pretraining on a mix of social media and conventional language yields an LM that outperforms the best similar-sized alternative by 4.2 points on the overall SMILE score.

Abstract: Many self-supervised speech models (S3Ms) have been introduced over the last few years, producing performance and data efficiency improvements for a variety of speech tasks. Evidence is emerging that different S3Ms encode linguistic information in different layers, and also that some S3Ms appear to learn phone-like sub-word units. However, the extent to which these models capture larger linguistic units, such as words, and where word-related information is encoded, remains unclear. In this study, we conduct several analyses of word segment representations extracted from different layers of three S3Ms: wav2vec2, HuBERT, and WavLM. We employ canonical correlation analysis (CCA), a lightweight analysis tool, to measure the similarity between these representations and word-level linguistic properties. We find that the maximal word-level linguistic content tends to be found in intermediate model layers, while some lower-level information like pronunciation is also retained in higher layers of HuBERT and WavLM. Syntactic and semantic word attributes have similar layer-wise behavior. We also find that, for all of the models tested, word identity information is concentrated near the center of each word segment. We then test the layer-wise performance of the same models, when used directly with no additional learned parameters, on several tasks: acoustic word discrimination, word segmentation, and semantic sentence similarity. We find similar layer-wise trends in performance, and furthermore, find that when using the best-performing layer of HuBERT or WavLM, it is possible to achieve performance on word segmentation and sentence similarity that rivals more complex existing approaches.

Abstract: We consider the questions of whether or not large language models (LLMs) have beliefs, and, if they do, how we might measure them. First, we evaluate two existing approaches, one due to Azaria and Mitchell (2023) and the other to Burns et al. (2022). We provide empirical results that show that these methods fail to generalize in very basic ways. We then argue that, even if LLMs have beliefs, these methods are unlikely to be successful for conceptual reasons. Thus, there is still no lie-detector for LLMs. After describing our empirical results we take a step back and consider whether or not we should expect LLMs to have something like beliefs in the first place. We consider some recent arguments aiming to show that LLMs cannot have beliefs. We show that these arguments are misguided. We provide a more productive framing of questions surrounding the status of beliefs in LLMs, and highlight the empirical nature of the problem. We conclude by suggesting some concrete paths for future work.