Computation and Language (cs.CL)

Abstract: Language models often exhibit behaviors that improve performance on a pre-training objective but harm performance on downstream tasks. We propose a novel approach to removing undesirable behaviors by ablating a small number of causal pathways between model components, with the intention of disabling the computational circuit responsible for the bad behavior. Given a small dataset of inputs where the model behaves poorly, we learn to ablate a small number of important causal pathways. In the setting of reducing GPT-2 toxic language generation, we find ablating just 12 of the 11.6K causal edges mitigates toxic generation with minimal degradation of performance on other inputs.

Abstract: Many computational linguistic methods have been proposed to study the information content of languages. We consider two interesting research questions: 1) how is information distributed over long documents, and 2) how does content reduction, such as token selection and text summarization, affect the information density in long documents. We present four criteria for information density estimation for long documents, including surprisal, entropy, uniform information density, and lexical density. Among those criteria, the first three adopt the measures from information theory. We propose an attention-based word selection method for clinical notes and study machine summarization for multiple-domain documents. Our findings reveal the systematic difference in information density of long text in various domains. Empirical results on automated medical coding from long clinical notes show the effectiveness of the attention-based word selection method.

Abstract: The rapid development of Large Language Models (LLMs) and the emergence of novel abilities with scale have necessitated the construction of holistic, diverse and challenging benchmarks such as HELM and BIG-bench. However, at the moment, most of these benchmarks focus only on performance in English and evaluations that include Southeast Asian (SEA) languages are few in number. We therefore propose BHASA, a holistic linguistic and cultural evaluation suite for LLMs in SEA languages. It comprises three components: (1) a NLP benchmark covering eight tasks across Natural Language Understanding (NLU), Generation (NLG) and Reasoning (NLR) tasks, (2) LINDSEA, a linguistic diagnostic toolkit that spans the gamut of linguistic phenomena including syntax, semantics and pragmatics, and (3) a cultural diagnostics dataset that probes for both cultural representation and sensitivity. For this preliminary effort, we implement the NLP benchmark only for Indonesian, Vietnamese, Thai and Tamil, and we only include Indonesian and Tamil for LINDSEA and the cultural diagnostics dataset. As GPT-4 is purportedly one of the best-performing multilingual LLMs at the moment, we use it as a yardstick to gauge the capabilities of LLMs in the context of SEA languages. Our initial experiments on GPT-4 with BHASA find it lacking in various aspects of linguistic capabilities, cultural representation and sensitivity in the targeted SEA languages. BHASA is a work in progress and will continue to be improved and expanded in the future.

Abstract: The cross-lingual transfer is a promising technique to solve tasks in less-resourced languages. In this empirical study, we compare two fine-tuning approaches combined with zero-shot and full-shot learning approaches for large language models in a cross-lingual setting. As fine-tuning strategies, we compare parameter-efficient adapter methods with fine-tuning of all parameters. As cross-lingual transfer strategies, we compare the intermediate-training (\textit{IT}) that uses each language sequentially and cross-lingual validation (\textit{CLV}) that uses a target language already in the validation phase of fine-tuning. We assess the success of transfer and the extent of catastrophic forgetting in a source language due to cross-lingual transfer, i.e., how much previously acquired knowledge is lost when we learn new information in a different language. The results on two different classification problems, hate speech detection and product reviews, each containing datasets in several languages, show that the \textit{IT} cross-lingual strategy outperforms \textit{CLV} for the target language. Our findings indicate that, in the majority of cases, the \textit{CLV} strategy demonstrates superior retention of knowledge in the base language (English) compared to the \textit{IT} strategy, when evaluating catastrophic forgetting in multiple cross-lingual transfers.

Abstract: Media houses reporting on public figures, often come with their own biases stemming from their respective worldviews. A characterization of these underlying patterns helps us in better understanding and interpreting news stories. For this, we need diverse or subjective summarizations, which may not be amenable for classifying into predefined class labels. This work proposes a zero-shot approach for non-extractive or generative characterizations of person entities from a corpus using GPT-2. We use well-articulated articles from several well-known news media houses as a corpus to build a sound argument for this approach. First, we fine-tune a GPT-2 pre-trained language model with a corpus where specific person entities are characterized. Second, we further fine-tune this with demonstrations of person entity characterizations, created from a corpus of programmatically constructed characterizations. This twice fine-tuned model is primed with manual prompts consisting of entity names that were not previously encountered in the second fine-tuning, to generate a simple sentence about the entity. The results were encouraging, when compared against actual characterizations from the corpus.

Abstract: We present a hybrid approach to the automated measurement of vagueness and subjectivity in texts. We first introduce the expert system VAGO, we illustrate it on a small benchmark of fact vs. opinion sentences, and then test it on the larger French press corpus FreSaDa to confirm the higher prevalence of subjective markers in satirical vs. regular texts. We then build a neural clone of VAGO, based on a BERT-like architecture, trained on the symbolic VAGO scores obtained on FreSaDa. Using explainability tools (LIME), we show the interest of this neural version for the enrichment of the lexicons of the symbolic version, and for the production of versions in other languages.

Abstract: Text-to-image diffusion models, e.g. Stable Diffusion (SD), lately have shown remarkable ability in high-quality content generation, and become one of the representatives for the recent wave of transformative AI. Nevertheless, such advance comes with an intensifying concern about the misuse of this generative technology, especially for producing copyrighted or NSFW (i.e. not safe for work) images. Although efforts have been made to filter inappropriate images/prompts or remove undesirable concepts/styles via model fine-tuning, the reliability of these safety mechanisms against diversified problematic prompts remains largely unexplored. In this work, we propose Prompting4Debugging (P4D) as a debugging and red-teaming tool that automatically finds problematic prompts for diffusion models to test the reliability of a deployed safety mechanism. We demonstrate the efficacy of our P4D tool in uncovering new vulnerabilities of SD models with safety mechanisms. Particularly, our result shows that around half of prompts in existing safe prompting benchmarks which were originally considered "safe" can actually be manipulated to bypass many deployed safety mechanisms, including concept removal, negative prompt, and safety guidance. Our findings suggest that, without comprehensive testing, the evaluations on limited safe prompting benchmarks can lead to a false sense of safety for text-to-image models.

Abstract: We present the first shared task for detecting and analyzing code-switching in Guarani and Spanish, GUA-SPA at IberLEF 2023. The challenge consisted of three tasks: identifying the language of a token, NER, and a novel task of classifying the way a Spanish span is used in the code-switched context. We annotated a corpus of 1500 texts extracted from news articles and tweets, around 25 thousand tokens, with the information for the tasks. Three teams took part in the evaluation phase, obtaining in general good results for Task 1, and more mixed results for Tasks 2 and 3.

Abstract: This paper presents a novel approach to address the Entity Recognition and Linking Challenge at NLPCC 2015. The task involves extracting named entity mentions from short search queries and linking them to entities within a reference Chinese knowledge base. To tackle this problem, we first expand the existing knowledge base and utilize external knowledge to identify candidate entities, thereby improving the recall rate. Next, we extract features from the candidate entities and utilize Support Vector Regression and Multiple Additive Regression Tree as scoring functions to filter the results. Additionally, we apply rules to further refine the results and enhance precision. Our method is computationally efficient and achieves an F1 score of 0.535.

Abstract: Simultaneous machine translation (SiMT) outputs translation while reading the source sentence. Unlike conventional sequence-to-sequence (seq2seq) training, existing SiMT methods adopt the prefix-to-prefix (prefix2prefix) training, where the model predicts target tokens based on partial source tokens. However, the prefix2prefix training diminishes the ability of the model to capture global information and introduces forced predictions due to the absence of essential source information. Consequently, it is crucial to bridge the gap between the prefix2prefix training and seq2seq training to enhance the translation capability of the SiMT model. In this paper, we propose a novel method that glances future in curriculum learning to achieve the transition from the seq2seq training to prefix2prefix training. Specifically, we gradually reduce the available source information from the whole sentence to the prefix corresponding to that latency. Our method is applicable to a wide range of SiMT methods and experiments demonstrate that our method outperforms strong baselines.

Abstract: News recommender systems play an increasingly influential role in shaping information access within democratic societies. However, tailoring recommendations to users' specific interests can result in the divergence of information streams. Fragmented access to information poses challenges to the integrity of the public sphere, thereby influencing democracy and public discourse. The Fragmentation metric quantifies the degree of fragmentation of information streams in news recommendations. Accurate measurement of this metric requires the application of Natural Language Processing (NLP) to identify distinct news events, stories, or timelines. This paper presents an extensive investigation of various approaches for quantifying Fragmentation in news recommendations. These approaches are evaluated both intrinsically, by measuring performance on news story clustering, and extrinsically, by assessing the Fragmentation scores of different simulated news recommender scenarios. Our findings demonstrate that agglomerative hierarchical clustering coupled with SentenceBERT text representation is substantially better at detecting Fragmentation than earlier implementations. Additionally, the analysis of simulated scenarios yields valuable insights and recommendations for stakeholders concerning the measurement and interpretation of Fragmentation.

Abstract: Reasoning presents a significant and challenging issue for Large Language Models (LLMs). The predominant focus of research has revolved around developing diverse prompting strategies to guide and structure the reasoning processes of LLMs. However, these approaches based on decoder-only causal language models often operate the input question in a single forward pass, potentially missing the rich, back-and-forth interactions inherent in human reasoning. Scant attention has been paid to a critical dimension, i.e., the input question itself embedded within the prompts. In response, we introduce a deceptively simple yet highly effective prompting strategy, termed question "re-reading". Drawing inspiration from human learning and problem-solving, re-reading entails revisiting the question information embedded within input prompts. This approach aligns seamlessly with the cognitive principle of reinforcement, enabling LLMs to extract deeper insights, identify intricate patterns, establish more nuanced connections, and ultimately enhance their reasoning capabilities across various tasks. Experiments conducted on a series of reasoning benchmarks serve to underscore the effectiveness and generality of our method. Moreover, our findings demonstrate that our approach seamlessly integrates with various language models, though-eliciting prompting methods, and ensemble techniques, further underscoring its versatility and compatibility in the realm of LLMs.

Abstract: Prior work has shown that the ordering in which concepts are shown to a commonsense generator plays an important role, affecting the quality of the generated sentence. However, it remains a challenge to determine the optimal ordering of a given set of concepts such that a natural sentence covering all the concepts could be generated from a pretrained generator. To understand the relationship between the ordering of the input concepts and the quality of the generated sentences, we conduct a systematic study considering multiple language models (LMs) and concept ordering strategies. We find that BART-large model consistently outperforms all other LMs considered in this study when fine-tuned using the ordering of concepts as they appear in CommonGen training data as measured using multiple evaluation metrics. Moreover, the larger GPT3-based large language models (LLMs) variants do not necessarily outperform much smaller LMs on this task, even when fine-tuned on task-specific training data. Interestingly, human annotators significantly reorder input concept sets when manually writing sentences covering those concepts, and this ordering provides the best sentence generations independently of the LM used for the generation, outperforming a probabilistic concept ordering baseline

Abstract: Automatic related work generation must ground their outputs to the content of the cited papers to avoid non-factual hallucinations, but due to the length of scientific documents, existing abstractive approaches have conditioned only on the cited paper \textit{abstracts}. We demonstrate that the abstract is not always the most appropriate input for citation generation and that models trained in this way learn to hallucinate. We propose to condition instead on the \textit{cited text span} (CTS) as an alternative to the abstract. Because manual CTS annotation is extremely time- and labor-intensive, we experiment with automatic, ROUGE-based labeling of candidate CTS sentences, achieving sufficiently strong performance to substitute for expensive human annotations, and we propose a human-in-the-loop, keyword-based CTS retrieval approach that makes generating citation texts grounded in the full text of cited papers both promising and practical.