论文检索

Zhexin Zhang, Jiale Cheng, Hao Sun, Jiawen Deng, Minlie Huang

Safety detection has been an increasingly important topic in recent years and it has become even more necessary to develop reliable safety detection systems with the rapid development of large language models. However, currently available safety detection systems have limitations in terms of their versatility and interpretability. In this paper, we first introduce InstructSafety, a safety detection framework that unifies 7 common sub-tasks for safety detection. These tasks are unified into a similar form through different instructions. We then conduct a comprehensive survey of existing safety detection datasets and process 39 human-annotated datasets for instruction tuning. We also construct adversarial samples to enhance the model’s robustness. After fine-tuning Flan-T5 on the collected data, we have developed Safety-Flan-T5, a multidimensional and explainable safety detector. We conduct comprehensive experiments on a variety of datasets and tasks, and demonstrate the strong performance of Safety-Flan-T5 in comparison to supervised baselines and served APIs (Perspective API, ChatGPT and InstructGPT). We will release the processed data, fine-tuned Safety-Flan-T5 and related code for public use.

Kankan Zhou, Eason Lai, Wei Bin Au Yeong, Kyriakos Mouratidis, Jing Jiang

Humans possess a strong capability for reasoning beyond common sense. For example, given an unconventional image of a goldfish laying on the table next to an empty fishbowl, a human would effortlessly determine that the fish is not inside the fishbowl. The case, however, may be different for a vision-language model, whose reasoning could gravitate towards the common scenario that the fish is inside the bowl, despite the visual input. In this paper, we introduce a novel probing dataset named ROME (reasoning beyond commonsense knowledge) to evaluate whether the state-of-the-art pre-trained vision-language models have the reasoning capability to correctly interpret counter-intuitive content. ROME contains images that defy commonsense knowledge with regards to color, shape, material, size and positional relation. Experiments on the state-of-the-art pre-trained vision-language models reveal that most of these models are still largely incapable of interpreting counter-intuitive scenarios. We hope that ROME will spur further investigations on reasoning beyond commonsense knowledge in vision-language research.

Jingheng Ye, Yinghui Li, Yangning Li, Hai-Tao Zheng

Data Augmentation through generating pseudo data has been proven effective in mitigating the challenge of data scarcity in the field of Grammatical Error Correction (GEC). Various augmentation strategies have been widely explored, most of which are motivated by two heuristics, i.e., increasing the distribution similarity and diversity of pseudo data. However, the underlying mechanism responsible for the effectiveness of these strategies remains poorly understood. In this paper, we aim to clarify how data augmentation improves GEC models. To this end, we introduce two interpretable and computationally efficient measures: Affinity and Diversity. Our findings indicate that an excellent GEC data augmentation strategy characterized by high Affinity and appropriate Diversity can better improve the performance of GEC models. Based on this observation, we propose MixEdit, a data augmentation approach that strategically and dynamically augments realistic data, without requiring extra monolingual corpora. To verify the correctness of our findings and the effectiveness of the proposed MixEdit, we conduct experiments on mainstream English and Chinese GEC datasets. The results show that MixEdit substantially improves GEC models and is complementary to traditional data augmentation methods. All the source codes of MixEdit are released at https://github.com/THUKElab/MixEdit.

Shih-Chieh Dai, Aiping Xiong, Lun-Wei Ku

Thematic analysis (TA) has been widely used for analyzing qualitative data in many disciplines and fields. To ensure reliable analysis, the same piece of data is typically assigned to at least two human coders. Moreover, to produce meaningful and useful analysis, human coders develop and deepen their data interpretation and coding over multiple iterations, making TA labor-intensive and time-consuming. Recently the emerging field of large language models (LLMs) research has shown that LLMs have the potential replicate human-like behavior in various tasks: in particular, LLMs outperform crowd workers on text-annotation tasks, suggesting an opportunity to leverage LLMs on TA. We propose a human–LLM collaboration framework (i.e., LLM-in-the-loop) to conduct TA with in-context learning (ICL). This framework provides the prompt to frame discussions with a LLM (e.g., GPT-3.5) to generate the final codebook for TA. We demonstrate the utility of this framework using survey datasets on the aspects of the music listening experience and the usage of a password manager. Results of the two case studies show that the proposed framework yields similar coding quality to that of human coders but reduces TA’s labor and time demands.

Marta R. Costa-jussà, Eric Smith, Christophe Ropers, Daniel Licht, Jean Maillard, Javier Ferrando, Carlos Escolano

Machine Translation systems can produce different types of errors, some of which are characterized as critical or catastrophic due to the specific negative impact that they can have on users. In this paper we focus on one type of critical error: added toxicity. We evaluate and analyze added toxicity when translating a large evaluation dataset (HOLISTICBIAS, over 472k sentences, covering 13 demographic axes) from English into 164 languages. An automatic toxicity evaluation shows that added toxicity across languages varies from 0% to 5%. The output languages with the most added toxicity tend to be low-resource ones, and the demographic axes with the most added toxicity include sexual orientation, gender and sex, and ability. We also perform human evaluation on a subset of 8 translation directions, confirming the prevalence of true added toxicity. We use a measurement of the amount of source contribution to the translation, where a low source contribution implies hallucination, to interpret what causes toxicity. Making use of the input attributions allows us to explain toxicity, because the source contributions significantly correlate with toxicity for 84% of languages studied. Given our findings, our recommendations to reduce added toxicity are to curate training data to avoid mistranslations, mitigate hallucination and check unstable translations.

Yurun Song, Junchen Zhao, Spencer Koehler, Amir Abdullah, Ian Harris

Intent detection is a major task in Natural Language Understanding (NLU) and is the component of dialogue systems for interpreting users’ intentions based on their utterances. Many works have explored detecting intents by assuming that each utterance represents only a single intent. Such systems have achieved very good results; however, intent detection is a far more challenging task in typical real-world scenarios, where each user utterance can be highly complex and express multiple intents. Therefore, in this paper, we propose PCMID, a novel Multi-Intent Detection framework enabled by Prototypical Contrastive Learning under a supervised setting. The PCMID model can learn multiple semantic representations of a given user utterance under the context of different intent labels in an optimized semantic space. Our experiments show that PCMID achieves the current state-of-the-art performance on both multiple public benchmark datasets and a private real-world dataset for the multi-intent detection task.

Raymond Li, Gabriel Murray, Giuseppe Carenini

In this work, we propose a method that combines two popular research areas by injecting linguistic structures into pre-trained language models in the parameter-efficient fine-tuning (PEFT) setting. In our approach, parallel adapter modules encoding different linguistic structures are combined using a novel Mixture-of-Linguistic-Experts architecture, where Gumbel-Softmax gates are used to determine the importance of these modules at each layer of the model. To reduce the number of parameters, we first train the model for a fixed small number of steps before pruning the experts based on their important scores. Our experiment results with three different pre-trained models show that our approach can outperform state-of-the-art PEFT methods with a comparable number of parameters. In addition, we provide additional analysis to examine the experts selected by each model at each layer to provide insights for future studies.

Xulang Zhang, Rui Mao, Kai He, Erik Cambria

Sentiment analysis is a task that highly depends on the understanding of word senses. Traditional neural network models are black boxes that represent word senses as vectors that are uninterpretable for humans. On the other hand, the application of Word Sense Disambiguation (WSD) systems in downstream tasks poses challenges regarding i) which words need to be disambiguated, and ii) how to model explicit word senses into easily understandable terms for a downstream model. This work proposes a neurosymbolic framework that incorporates WSD by identifying and paraphrasing ambiguous words to improve the accuracy of sentiment predictions. The framework allows us to understand which words are paraphrased into which semantically unequivocal words, thus enabling a downstream task model to gain both accuracy and interpretability. To better fine-tune a lexical substitution model for WSD on a downstream task without ground-truth word sense labels, we leverage dynamic rewarding to jointly train sentiment analysis and lexical substitution models. Our framework proves to effectively improve the performance of sentiment analysis on corpora from different domains.

Denis McInerney, Geoffrey Young, Jan-Willem van de Meent, Byron Wallace

We propose CHiLL (Crafting High-Level Latents), an approach for natural-language specification of features for linear models. CHiLL prompts LLMs with expert-crafted queries to generate interpretable features from health records. The resulting noisy labels are then used to train a simple linear classifier. Generating features based on queries to an LLM can empower physicians to use their domain expertise to craft features that are clinically meaningful for a downstream task of interest, without having to manually extract these from raw EHR. We are motivated by a real-world risk prediction task, but as a reproducible proxy, we use MIMIC-III and MIMIC-CXR data and standard predictive tasks (e.g., 30-day readmission) to evaluate this approach. We find that linear models using automatically extracted features are comparably performant to models using reference features, and provide greater interpretability than linear models using “Bag-of-Words” features. We verify that learned feature weights align well with clinical expectations.

Dung Thai, Dhruv Agarwal, Mudit Chaudhary, Wenlong Zhao, Rajarshi Das, Jay-Yoon Lee, Hannaneh Hajishirzi, Manzil Zaheer, Andrew McCallum

We present an accurate and interpretable method for answer extraction in machine reading comprehension that is reminiscent of case-based reasoning (CBR) from classical AI. Our method (CBR-MRC) builds upon the hypothesis that contextualized answers to similar questions share semantic similarities with each other. Given a test question, CBR-MRC first retrieves a set of similar cases from a nonparametric memory and then predicts an answer by selecting the span in the test context that is most similar to the contextualized representations of answers in the retrieved cases. The semi-parametric nature of our approach allows it to attribute a prediction to the specific set of evidence cases, making it a desirable choice for building reliable and debuggable QA systems. We show that CBR-MRC provides high accuracy comparable with large reader models and outperforms baselines by 11.5 and 8.4 EM on NaturalQuestions and NewsQA, respectively. Further, we demonstrate the ability of CBR-MRC in identifying not just the correct answer tokens but also the span with the most relevant supporting evidence. Lastly, we observe that contexts for certain question types show higher lexical diversity than others and find that CBR-MRC is robust to these variations while performance using fully-parametric methods drops.

Chenglei Si, Weijia Shi, Chen Zhao, Luke Zettlemoyer, Jordan Boyd-Graber

While recent large language models (LLMs) improve on various question answering (QA) datasets, it remains difficult for a single model to generalize across question types that require distinct reasoning abilities. We provide empirical evidence that state-of-the-art LLMs suffer from poor generalizability on reasoning types beyond those seen in the prompt. To remedy this, we propose a Mixture-of-Reasoning-Experts (MORE) framework that ensembles diverse specialized language models. We specialize the backbone language model with prompts optimized for different reasoning categories, including factual, multihop, mathematical, and commonsense reasoning. Our key insight is to leverage agreement among the specialized experts to select the best answer for each question, or to abstain from answering. This gives MORE higher accuracy than any single specialized model on a collection of 12 QA datasets from four reasoning types. Beyond generalizability, the interpretable design of MORE improves selective question answering results compared to baselines without incorporating inter-expert agreement. This framework is also more interpretable and useful to human consumers of QA outputs. Our human study confirms that presenting expert predictions and the answer selection process helps annotators more accurately calibrate when to trust the system’s output. We release all code and data to facilitate future work.

Ningyuan Li, Haihong E, Shi Li, Mingzhi Sun, Tianyu Yao, Meina Song, Yong Wang, Haoran Luo

Temporal knowledge graph (TKG) has been proved to be an effective way for modeling dynamic facts in real world. Many efforts have been devoted into predicting future events i.e. extrapolation, on TKGs. Recently, rule-based knowledge graph completion methods which are considered to be more interpretable than embedding-based methods, have been transferred to temporal knowledge graph extrapolation. However, rule-based models suffer from temporal redundancy when leveraged under dynamic settings, which results in inaccurate rule confidence calculation. In this paper, we define the problem of temporal redundancy and propose TR-Rules which solves the temporal redundancy issues through a simple but effective strategy. Besides, to capture more information lurking in TKGs, apart from cyclic rules, TR-Rules also mines and properly leverages acyclic rules, which has not been explored by existing models. Experimental results on three benchmarks show that TR-Rules achieves state-of-the-art performance. Ablation study shows the impact of temporal redundancy and demonstrates the performance of acyclic rules is much more promising due to its higher sensitivity to the number of sampled walks during learning stage.

Albert Webson, Alyssa Loo, Qinan Yu, Ellie Pavlick

Prompts have been the center of progress in advancing language models’ zero-shot and few-shot performance. However, recent work finds that models can perform surprisingly well when given intentionally irrelevant or misleading prompts. Such results may be interpreted as evidence that model behavior is not “human like’. In this study, we challenge a central assumption in such work: that humans would perform badly when given pathological instructions. We find that humans are able to reliably ignore irrelevant instructions and thus, like models, perform well on the underlying task despite an apparent lack of signal regarding the task they are being asked to do. However, when given deliberately misleading instructions, humans follow the instructions faithfully, whereas models do not. Thus, our conclusion is mixed with respect to prior work. We argue against the earlier claim that high performance with irrelevant prompts constitutes evidence against models’ instruction understanding, but we reinforce the claim that models’ failure to follow misleading instructions raises concerns. More broadly, we caution that future research should not idealize human behaviors as a monolith and should not train or evaluate models to mimic assumptions about these behaviors without first validating humans’ behaviors empirically.

Jing Huang, Diyi Yang

Humans produce and consume language in a particular cultural context, which includes knowledge about specific norms and practices. A listener’s awareness of the cultural context is critical for interpreting the speaker’s meaning. A simple expression like *I didn’t leave a tip* implies a strong sense of dissatisfaction when tipping is assumed to be the norm. As NLP systems reach users from different cultures, achieving culturally aware language understanding becomes increasingly important. However, current research has focused on building cultural knowledge bases without studying how such knowledge leads to contextualized interpretations of texts. In this work, we operationalize cultural variations in language understanding through a natural language inference (NLI) task that surfaces cultural variations as label disagreement between annotators from different cultural groups. We introduce the first Culturally Aware Natural Language Inference (CALI) dataset with 2.7K premise-hypothesis pairs annotated by two cultural groups located in the U.S. and India. With CALI, we categorize how cultural norms affect language understanding and present an evaluation framework to assess at which levels large language models are culturally aware. Our dataset is available at https://github.com/SALT-NLP/CulturallyAwareNLI.

Shufan Wang, Sébastien Jean, Sailik Sengupta, James Gung, Nikolaos Pappas, Yi Zhang

In executable task-oriented semantic parsing, the system aims to translate users’ utterances in natural language to machine-interpretable programs (API calls) that can be executed according to pre-defined API specifications. With the popularity of Large Language Models (LLMs), in-context learning offers a strong baseline for such scenarios, especially in data-limited regimes. However, LLMs are known to hallucinate and therefore pose a formidable challenge in constraining generated content. Thus, it remains uncertain if LLMs can effectively perform task-oriented utterance-to-API generation, where respecting the API’s structural and task-specific constraints is crucial. In this work, we seek to measure, analyze and mitigate such constraints violations. First, we identify the categories of various constraints in obtaining API-semantics from task-oriented utterances, and define fine-grained metrics that complement traditional ones. Second, we leverage these metrics to conduct a detailed error analysis of constraints violations seen in state-of-the-art LLMs, which motivates us to investigate two popular mitigation strategies– Semantic-Retrieval of Demonstrations (SRD) and API-aware Constrained Decoding (API-CD). Our experiments show that these strategies are effective at reducing constraints violations and improving the quality of the generated API calls, but require careful consideration given their implementation complexity and latency.

Giulio Starace, Konstantinos Papakostas, Rochelle Choenni, Apostolos Panagiotopoulos, Matteo Rosati, Alina Leidinger, Ekaterina Shutova

Large Language Models (LLMs) exhibit impressive performance on a range of NLP tasks, due to the general-purpose linguistic knowledge acquired during pretraining. Existing model interpretability research (Tenney et al., 2019) suggests that a linguistic hierarchy emerges in the LLM layers, with lower layers better suited to solving syntactic tasks and higher layers employed for semantic processing. Yet, little is known about how encodings of different linguistic phenomena interact within the models and to what extent processing of linguistically-related categories relies on the same, shared model representations. In this paper, we propose a framework for testing the joint encoding of linguistic categories in LLMs. Focusing on syntax, we find evidence of joint encoding both at the same (related part-of-speech (POS) classes) and different (POS classes and related syntactic dependency relations) levels of linguistic hierarchy. Our cross-lingual experiments show that the same patterns hold across languages in multilingual LLMs.

Miao Li, Eduard Hovy, Jey Lau

We present PeerSum, a novel dataset for generating meta-reviews of scientific papers. The meta-reviews can be interpreted as abstractive summaries of reviews, multi-turn discussions and the paper abstract. These source documents have a rich inter-document relationship with an explicit hierarchical conversational structure, cross-references and (occasionally) conflicting information. To introduce the structural inductive bias into pre-trained language models, we introduce RAMMER (Relationship-aware Multi-task Meta-review Generator), a model that uses sparse attention based on the conversational structure and a multi-task training objective that predicts metadata features (e.g., review ratings). Our experimental results show that RAMMER outperforms other strong baseline models in terms of a suite of automatic evaluation metrics. Further analyses, however, reveal that RAMMER and other models struggle to handle conflicts in source documents, suggesting meta-review generation is a challenging task and a promising avenue for further research.

Youngwoo Kim, Razieh Rahimi, James Allan

To properly explain sentence pairs that provide contradictory (different) information for different conditions, we introduce the task of conditional natural language inference (Cond-NLI) and focus on automatically extracting contradictory aspects and their conditions from a sentence pair. Cond-NLI can help to provide a full spectrum of information, such as when there are multiple answers to a question each addressing a specific condition, or reviews with different opinions for different conditions. We show that widely-used feature-attribution explanation models are not suitable for finding conditions, especially when sentences are long and are written independently. We propose a simple yet effective model for the original NLI task that can successfully extract conditions while not requiring token-level annotations. Our model enhances the interpretability of the NLI task while maintaining comparable accuracy. To evaluate models for the Cond-NLI, we build and release a token-level annotated dataset BioClaim which contains potentially contradictory claims from the biomedical domain. Our experiments show that our proposed model outperforms the full cross-encoder and other baselines in extracting conditions. It also performs on-par with GPT-3 which has an order of magnitude more parameters and trained on a huge amount of data.

Ruosen Li, Xinya Du

Neural models, including large language models (LLMs), achieve superior performance on multi-hop question-answering. To elicit reasoning capabilities from LLMs, recent works propose using the chain-of-thought (CoT) mechanism to generate both the reasoning chain and the answer, which enhances the model’s capabilities in conducting multi-hop reasoning. However, several challenges still remain: such as struggling with inaccurate reasoning, hallucinations, and lack of interpretability. On the other hand, information extraction (IE) identifies entities, relations, and events grounded to the text. The extracted structured information can be easily interpreted by humans and machines (Grishman, 2019). In this work, we investigate constructing and leveraging extracted semantic structures (graphs) for multi-hop question answering, especially the reasoning process. Empirical results and human evaluations show that our framework: generates more faithful reasoning chains and substantially improves the QA performance on two benchmark datasets. Moreover, the extracted structures themselves naturally provide grounded explanations that are preferred by humans, as compared to the generated reasoning chains and saliency-based explanations.

Tyler Loakman, Aaron Maladry, Chenghua Lin

Human evaluation in often considered to be the gold standard method of evaluating a Natural Language Generation system. However, whilst its importance is accepted by the community at large, the quality of its execution is often brought into question. In this position paper, we argue that the generation of more esoteric forms of language - humour, irony and sarcasm - constitutes a subdomain where the characteristics of selected evaluator panels are of utmost importance, and every effort should be made to report demographic characteristics wherever possible, in the interest of transparency and replicability. We support these claims with an overview of each language form and an analysis of examples in terms of how their interpretation is affected by different participant variables. We additionally perform a critical survey of recent works in NLG to assess how well evaluation procedures are reported in this subdomain, and note a severe lack of open reporting of evaluator demographic information, and a significant reliance on crowdsourcing platforms for recruitment.