ACL 2023 论文列表

Prompt-tuning has become an increasingly popular parameter-efficient method for adapting large pretrained language models to downstream tasks. However, both discrete prompting and continuous prompting assume fixed prompts for all data samples within a task, neglecting the fact that inputs vary greatly in some tasks such as open-domain dialogue generation. In this paper, we present a novel, instance-specific prompt-tuning algorithm for dialogue generation. Specifically, we generate prompts based on instance-level control code, rather than the conversation history, to explore their impact on controlled dialogue generation. Experiments on popular open-domain dialogue datasets, evaluated on both automated metrics and human evaluation, demonstrate that our method is superior to prompting baselines and comparable to fine-tuning with only 5%-6% of total parameters.

We propose a method for unsupervised opinion summarization that encodes sentences from customer reviews into a hierarchical discrete latent space, then identifies common opinions based on the frequency of their encodings. We are able to generate both abstractive summaries by decoding these frequent encodings, and extractive summaries by selecting the sentences assigned to the same frequent encodings. Our method is attributable, because the model identifies sentences used to generate the summary as part of the summarization process. It scales easily to many hundreds of input reviews, because aggregation is performed in the latent space rather than over long sequences of tokens. We also demonstrate that our appraoch enables a degree of control, generating aspect-specific summaries by restricting the model to parts of the encoding space that correspond to desired aspects (e.g., location or food). Automatic and human evaluation on two datasets from different domains demonstrates that our method generates summaries that are more informative than prior work and better grounded in the input reviews.

Most research on stylized image captioning aims to generate style-specific captions using unpaired text, and has achieved impressive performance for simple styles like positive and negative. However, unlike previous single-sentence captions whose style is mostly embodied in distinctive words or phrases, real-world styles are likely to be implied at the syntactic and discourse levels. In this work, we introduce a new task of Stylized Visual Storytelling (SVST), which aims to describe a photo stream with stylized stories that are more expressive and attractive. We propose a multitasking memory-augmented framework called StyleVSG, which is jointly trained on factual visual storytelling data and unpaired style corpus, achieving a trade-off between style accuracy and visual relevance. Particularly for unpaired stylized text, StyleVSG learns to reconstruct the stylistic story from roughly parallel visual inputs mined with the CLIP model, avoiding problems caused by random mapping in previous methods. Furthermore, a memory module is designed to preserve the consistency and coherence of generated stories. Experiments show that our method can generate attractive and coherent stories with different styles such as fairy tale, romance, and humor. The overall performance of our StyleVSG surpasses state-of-the-art methods on both automatic and human evaluation metrics.

In simultaneous speech translation (SimulST), effective policies that determine when to write partial translations are crucial to reach high output quality with low latency. Towards this objective, we propose EDAtt (Encoder-Decoder Attention), an adaptive policy that exploits the attention patterns between audio source and target textual translation to guide an offline-trained ST model during simultaneous inference. EDAtt exploits the attention scores modeling the audio-translation relation to decide whether to emit a partial hypothesis or wait for more audio input. This is done under the assumption that, if attention is focused towards the most recently received speech segments, the information they provide can be insufficient to generate the hypothesis (indicating that the system has to wait for additional audio input). Results on en->de, es show that EDAtt yields better results compared to the SimulST state of the art, with gains respectively up to 7 and 4 BLEU points for the two languages, and with a reduction in computational-aware latency up to 1.4s and 0.7s compared to existing SimulST policies applied to offline-trained models.

Although the incorporation of pre-trained language models (PLMs) significantly pushes the research frontier of multi-turn response selection, it brings a new issue of heavy computation costs. To alleviate this problem and make the PLM-based response selection model both effective and efficient, we propose an inference framework together with a post-training strategy that builds upon any pre-trained transformer-based response selection models to accelerate inference by progressively selecting and eliminating unimportant content under the guidance of context-response dual-attention. Specifically, at each transformer layer, we first identify the importance of each word based on context-to-response and response-to-context attention, then select a number of unimportant words to be eliminated following a retention configuration derived from evolutionary search while passing the rest of the representations into deeper layers. To mitigate the training-inference gap posed by content elimination, we introduce a post-training strategy where we use knowledge distillation to force the model with progressively eliminated content to mimic the predictions of the original model with no content elimination. Experiments on three benchmarks indicate that our method can effectively speeds-up SOTA models without much performance degradation and shows a better trade-off between speed and performance than previous methods.

Annotating long-document question answering (long-document QA) pairs is time-consuming and expensive. To alleviate the problem, it might be possible to generate long-document QA pairs via unsupervised question answering (UQA) methods. However, existing UQA tasks are based on short documents, and can hardly incorporate long-range information. To tackle the problem, we propose a new task, named unsupervised long-document question answering (ULQA), aiming to generate high-quality long-document QA instances in an unsupervised manner. Besides, we propose AttenWalker, a novel unsupervised method to aggregate and generate answers with long-range dependency so as to construct long-document QA pairs. Specifically, AttenWalker is composed of three modules, i.e. span collector, span linker and answer aggregator. Firstly, the span collector takes advantage of constituent parsing and reconstruction loss to select informative candidate spans for constructing answers. Secondly, with the help of the attention graph of a pre-trained long-document model, potentially interrelated text spans (that might be far apart) could be linked together via an attention-walking algorithm. Thirdly, in the answer aggregator, linked spans are aggregated into the final answer via the mask-filling ability of a pre-trained model. Extensive experiments show that AttenWalker outperforms previous methods on NarrativeQA and Qasper. In addition, AttenWalker also shows strong performance in the few-shot learning setting.

Attribute extraction aims to identify attribute names and the corresponding values from descriptive texts, which is the foundation for extensive downstream applications such as knowledge graph construction, search engines, and e-Commerce. In previous studies, attribute extraction is generally treated as a classification problem for predicting attribute types or a sequence tagging problem for labeling attribute values, where two paradigms, i.e., closed-world and open-world assumption, are involved. However, both of these paradigms have limitations in terms of real-world applications. And prior studies attempting to integrate these paradigms through ensemble, pipeline, and co-training models, still face challenges like cascading errors, high computational overhead, and difficulty in training. To address these existing problems, this paper presents Attribute Tree, a unified formulation for real-world attribute extraction application, where closed-world, open-world, and semi-open attribute extraction tasks are modeled uniformly. Then a text-to-tree generation model, AtTGen, is proposed to learn annotations from different scenarios efficiently and consistently. Experiments demonstrate that our proposed paradigm well covers various scenarios for real-world applications, and the model achieves state-of-the-art, outperforming existing methods by a large margin on three datasets. Our code, pretrained model, and datasets are available at https://github.com/lsvih/AtTGen.

In natural language processing (NLP), deep neural networks (DNNs) could model complex interactions between context and have achieved impressive results on a range of NLP tasks. Prior works on feature interaction attribution mainly focus on studying symmetric interaction that only explains the additional influence of a set of words in combination, which fails to capture asymmetric influence that contributes to model prediction. In this work, we propose an asymmetric feature interaction attribution explanation model that aims to explore asymmetric higher-order feature interactions in the inference of deep neural NLP models. By representing our explanation with an directed interaction graph, we experimentally demonstrate interpretability of the graph to discover asymmetric feature interactions. Experimental results on two sentiment classification datasets show the superiority of our model against the state-of-the-art feature interaction attribution methods in identifying influential features for model predictions.

Many NLP tasks require to automatically identify the most significant words in a text. In this work, we derive word significance from models trained to solve semantic task: Natural Language Inference and Paraphrase Identification. Using an attribution method aimed to explain the predictions of these models, we derive importance scores for each input token. We evaluate their relevance using a so-called cross-task evaluation: Analyzing the performance of one model on an input masked according to the other model’s weight, we show that our method is robust with respect to the choice of the initial task. Additionally, we investigate the scores from the syntax point of view and observe interesting patterns, e.g. words closer to the root of a syntactic tree receive higher importance scores. Altogether, these observations suggest that our method can be used to identify important words in sentences without any explicit word importance labeling in training.

Extractive opinion summarization extracts sentences from users’ reviews to represent the prevalent opinions about a product or service. However, the extracted sentences can be redundant and may miss some important aspects, especially for centroid-based extractive summarization models (Radev et al., 2004). To alleviate these issues, we introduce TokenCluster– a method for unsupervised extractive opinion summarization that automatically identifies the aspects described in the review sentences and then extracts sentences based on their aspects. It identifies the underlying aspects of the review sentences using roots of noun phrases and adjectives appearing in them. Empirical evaluation shows that TokenCluster improves aspect coverage in summaries and achieves strong performance on multiple opinion summarization datasets, for both general and aspect-specific summarization. We also perform extensive ablation and human evaluation studies to validate the design choices of our method. The implementation of our work is available at https://github.com/leehaoyuan/TokenCluster

Existing dialogue models may encounter scenarios which are not well-represented in the training data, and as a result generate responses that are unnatural, inappropriate, or unhelpful. We propose the “Ask an Expert” framework in which the model is trained with access to an “expert” which it can consult at each turn. Advice is solicited via a structured dialogue with the expert, and the model is optimized to selectively utilize (or ignore) it given the context and dialogue history. In this work the expert takes the form of an LLM.We evaluate this framework in a mental health support domain, where the structure of the expert conversation is outlined by pre-specified prompts which reflect a reasoning strategy taught to practitioners in the field. Blenderbot models utilizing “Ask an Expert” show quality improvements across all expert sizes, including those with fewer parameters than the dialogue model itself. Our best model provides a ~10% improvement over baselines, approaching human-level scores on “engingingness” and “helpfulness” metrics.

Effective argumentation is essential towards a purposeful conversation with a satisfactory outcome. For example, persuading someone to reconsider smoking might involve empathetic, well founded arguments based on facts and expert opinions about its ill-effects and the consequences on one’s family. However, the automatic generation of high-quality factual arguments can be challenging. Addressing existing controllability issues can make the recent advances in computational models for argument generation a potential solution. In this paper, we introduce ArgU: a neural argument generator capable of producing factual arguments from input facts and real-world concepts that can be explicitly controlled for stance and argument structure using Walton’s argument scheme-based control codes. Unfortunately, computational argument generation is a relatively new field and lacks datasets conducive to training. Hence, we have compiled and released an annotated corpora of 69,428 arguments spanning six topics and six argument schemes, making it the largest publicly available corpus for identifying argument schemes; the paper details our annotation and dataset creation framework. We further experiment with an argument generation strategy that establishes an inference strategy by generating an “argument template” before actual argument generation. Our results demonstrate that it is possible to automatically generate diverse arguments exhibiting different inference patterns for the same set of facts by using control codes based on argument schemes and stance.

Argument Quality Detection is an emerging field in NLP which has seen significant recent development. However, existing datasets in this field suffer from a lack of quality, quantity and diversity of topics and arguments, specifically the presence of vague arguments that are not persuasive in nature. In this paper, we leverage a combined experience of 10+ years of Parliamentary Debating to create a dataset that covers significantly more topics and has a wide range of sources to capture more diversity of opinion. With 34,890 high-quality argument-analysis pairs (a term we introduce in this paper), this is also the largest dataset of its kind to our knowledge. In addition to this contribution, we introduce an innovative argument scoring system based on instance-level annotator reliability and propose a quantitative model of scoring the relevance of arguments to a range of topics.

Large language models (LLMs) have demonstrated powerful capabilities in both text understanding and generation. Companies have begun to offer Embedding as a Service (EaaS) based on these LLMs, which can benefit various natural language processing (NLP) tasks for customers. However, previous studies have shown that EaaS is vulnerable to model extraction attacks, which can cause significant losses for the owners of LLMs, as training these models is extremely expensive. To protect the copyright of LLMs for EaaS, we propose an Embedding Watermark method called {pasted macro ‘METHOD’} that implants backdoors on embeddings. Our method selects a group of moderate-frequency words from a general text corpus to form a trigger set, then selects a target embedding as the watermark, and inserts it into the embeddings of texts containing trigger words as the backdoor. The weight of insertion is proportional to the number of trigger words included in the text. This allows the watermark backdoor to be effectively transferred to EaaS-stealer’s model for copyright verification while minimizing the adverse impact on the original embeddings’ utility. Our extensive experiments on various datasets show that our method can effectively protect the copyright of EaaS models without compromising service quality. Our code is available at https://github.com/yjw1029/EmbMarker.

In this paper, we explore the following question: Are synonym substitution attacks really synonym substitution attacks (SSAs)?We approach this question by examining how SSAs replace words in the original sentence and show that there are still unresolved obstacles that make current SSAs generate invalid adversarial samples. We reveal that four widely used word substitution methods generate a large fraction of invalid substitution words that are ungrammatical or do not preserve the original sentence’s semantics. Next, we show that the semantic and grammatical constraints used in SSAs for detecting invalid word replacements are highly insufficient in detecting invalid adversarial samples.

Recent results in image classification and extractive question answering have observed that pre-trained models trained on less in-distribution data have better out-ofdistribution performance. However, it is unclear how broadly these trends hold. We conduct a large empirical study across three tasks, three broadly-applicable modeling interventions (increasing model size, using a different adaptation method, and pre-training on more data), and 14 diverse datasets to investigate the relationship between sample efficiency (amount of data needed to reach a given ID accuracy) and robustness (how models fare on OOD evaluation). We find that higher sample efficiency is only correlated with better average OOD robustness on some modeling interventions and tasks, but not others. On individual datasets, models with lower sample efficiency can even be more robust. These results suggest that general-purpose methods for improving sample efficiency are unlikely to yield universal OOD robustness improvements, since such improvements are highly dataset- and task-dependent. Even in an era of large, multi-purpose pre-trained models, task-specific decisions may often be necessary for OOD generalization.

Model ensemble has been in widespread use for Grammatical Error Correction (GEC), boosting model performance. We hypothesize that model ensemble based on the perplexity (PPL) computed by pre-trained language models (PLMs) should benefit the GEC system. To this end, we explore several ensemble strategies based on strong PLMs with four sophisticated single models. However, the performance does not improve but even gets worse after the PLM-based ensemble. This surprising result sets us doing a detailed analysis on the data and coming up with some insights on GEC. The human references of correct sentences is far from sufficient in the test data, and the gap between a correct sentence and an idiomatic one is worth our attention. Moreover, the PLM-based ensemble strategies provide an effective way to extend and improve GEC benchmark data. Our source code is available at https://github.com/JamyDon/PLM-based-CGEC-Model-Ensemble.

Knowledge graphs (KGs) facilitate a wide variety of applications. Despite great efforts in creation and maintenance, even the largest KGs are far from complete. Hence, KG completion (KGC) has become one of the most crucial tasks for KG research. Recently, considerable literature in this space has centered around the use of Message Passing (Graph) Neural Networks (MPNNs), to learn powerful embeddings. The success of these methods is naturally attributed to the use of MPNNs over simpler multi-layer perceptron (MLP) models, given their additional message passing (MP) component. In this work, we find that surprisingly, simple MLP models are able to achieve comparable performance to MPNNs, suggesting that MP may not be as crucial as previously believed. With further exploration, we show careful scoring function and loss function design has a much stronger influence on KGC model performance. This suggests a conflation of scoring function design, loss function design, and MP in prior work, with promising insights regarding the scalability of state-of-the-art KGC methods today, as well as careful attention to more suitable MP designs for KGC tasks tomorrow.

Among the remarkable emergent capabilities of large language models (LMs) is free-text rationalization; beyond certain scale, large LMs are capable of generating seemingly useful rationalizations, which in turn, can dramatically enhance their performances on leaderboards. This phenomenon raises a question: can machine generated rationales also be useful for humans, especially when lay humans try to answer questions based on those machine rationales? We observe that human utility of existing rationales is far from satisfactory and expensive to estimate with human studies. Existing metrics like task performance of the LM generating the rationales or similarity between generated and gold rationales are not good indicators of their human utility. While we observe that certain properties of rationales like conciseness and novelty are correlated with their human utility, estimating them without human involvement is challenging. We show that, by estimating a rationale’s helpfulness in answering similar unseen instances, we can measure its human utility to a better extent. We also translate this finding into an automated score, Gen-U, that we propose, which can help improve LMs’ ability to generate rationales with better human utility, while maintaining most of its task performance. Lastly, we release all code and collected data with this project.