ACL 2023 论文列表

We present a simple and unified approach for both continuous and discontinuous constituency parsing via autoregressive span selection. Constituency parsing aims to produce a set of non-crossing spans so that they can form a constituency parse tree. We sort gold spans using a predefined order and leverage a pointer network to autoregressively select spans by that order. To deal with discontinuous spans, we consecutively select their subspans from left to right, label all but last subspans with special discontinuous labels and the last subspan as the whole discontinuous spans’ labels. We use simple heuristic to output valid trees so that our approach is able to predict all possible continuous and discontinuous constituency trees without sacrificing data coverage and without the need to use expensive chart-based parsing algorithms. Experiments on multiple continuous and discontinuous benchmarks show that our model achieves state-of-the-art or competitive performance.

As a critical step to achieve human-like chatbots, empathetic response generation has attained increasing interests. Previous attempts are incomplete and not sufficient enough to elicit empathy because they only stay on the initial stage of empathy to automatically sense and simulate the feelings and thoughts of others via other-awareness. However, they ignore to include self-awareness to consider the own views of the self in their responses, which is a crucial process to achieve the empathy. To this end, we propose to generate Empathetic response with explicit Self-Other Awareness (EmpSOA). Specifically, three stages, self-other differentiation, self-other modulation and self-other generation, are devised to clearly maintain, regulate and inject the self-other aware information into the process of empathetic response generation. Both automatic and human evaluations on the benchmark dataset demonstrate the superiority of EmpSOA to generate more empathetic responses. Our source code will be publicly available.

A key missing capacity of current language models (LMs) is grounding to real-world environments. Most existing work for grounded language understanding uses LMs to directly generate plans that can be executed in the environment to achieve the desired effects. It thereby casts the burden of ensuring grammaticality, faithfulness, and controllability all on the LMs. We propose Pangu, a generic framework for grounded language understanding that capitalizes on the discriminative ability of LMs instead of their generative ability. Pangu consists of a symbolic agent and a neural LM working in a concerted fashion: The agent explores the environment to incrementally construct valid plans, and the LM evaluates the plausibility of the candidate plans to guide the search process. A case study on the challenging problem of knowledge base question answering (KBQA), which features a massive environment, demonstrates the remarkable effectiveness and flexibility of Pangu: A BERT-base LM is sufficient for setting a new record on standard KBQA datasets, and larger LMs further bring substantial gains.Pangu also enables, for the first time, effective few-shot in-context learning for KBQA with large LMs such as Codex.

Despite tremendous advancements in dialogue systems, stable evaluation still requires human judgments producing notoriously high-variance metrics due to their inherent subjectivity. Moreover, methods and labels in dialogue evaluation are not fully standardized, especially for open-domain chats, with a lack of work to compare and assess the validity of those approaches. The use of inconsistent evaluation can misinform the performance of a dialogue system, which becomes a major hurdle to enhance it. Thus, a dimensional evaluation of chat-oriented open-domain dialogue systems that reliably measures several aspects of dialogue capabilities is desired. This paper presents a novel human evaluation method to estimate the rates of many{pasted macro ‘LN’} dialogue system behaviors. Our method is used to evaluate four state-of-the-art open-domain dialogue systems and compared with existing approaches. The analysis demonstrates that our behavior method is more suitable than alternative Likert-style or comparative approaches for dimensional evaluation of these systems.

Due to the democratization of e-commerce, many product companies are listing their goods for online shopping. For periodic buying within a domain such as Grocery, consumers are generally inclined to buy certain brands of products. Due to a large non-English speaking population in India, we observe a significant percentage of code-mix Hinglish search queries e.g., sasta atta. An intuitive approach to dealing with code-mix queries is to train an encoder-decoder model to translate the query to English to perform the search. However, the problem becomes non-trivial when the brand names themselves have Hinglish names and possibly have a literal English translation. In such queries, only the context (non-brand name) Hinglish words needs to be translated. In this paper, we propose a simple yet effective modification to the transformer training to preserve/correct Grocery brand names in the output while selectively translating the context words. To achieve this, we use an additional dataset of popular Grocery brand names. Brand names are added as tokens to the model vocabulary, and the token embeddings are randomly initialized. Further, we introduce a Brand loss in training the translation model. Brand loss is a cross entropy loss computed using a denoising auto-encoder objective with brand name data. We warm-start the training from a public pre-trained checkpoint (such as BART/T5) and further adapt it for query translation using the domain data. The proposed model is generic and can be used with English as well as code-mix Hinglish queries alleviating the need for language detection. To reduce the latency of the model for the production deployment, we use knowledge distillation and quantization. Experimental evaluation indicates that the proposed approach improves translation results by preserving/correcting English/Hinglish brand names. After positive results with A/B testing, the model is currently deployed in production.

The end-to-end task-oriented dialogue system has achieved great success in recent years. Most of these dialogue systems need to accommodate multi-domain dialogue in real-world scenarios. However, due to the high cost of dialogue data annotation and the scarcity of labeled dialogue data, existing methods are difficult to extend to new domains. Therefore, it is important to use limited data to construct multi-domain dialogue systems. To solve this problem, we propose a novel domain attention module. It use the distributional signatures to construct a multi-domain dialogue system effectively with limited data, which has strong extensibility. We also define a adjacent n-gram pattern to explore potential patterns for dialogue entities. Experimental results show that our approach outperforms the baseline models on most metrics. In the few-shot scenario, we show our method get a great improvement compared with previous methods while keeping smaller model scale.

Production deployments in complex systems require ML architectures to be highly efficient and usable against multiple tasks. Particularly demanding are classification problems in which data arrives in a streaming fashion and each class is presented separately. Recent methods with stochastic gradient learning have been shown to struggle in such setups or have limitations like memory buffers, and being restricted to specific domains that disable its usage in real-world scenarios. For this reason, we present a fully differentiable architecture based on the Mixture of Experts model, that enables the training of high-performance classifiers when examples from each class are presented separately. We conducted exhaustive experiments that proved its applicability in various domains and ability to learn online in production environments. The proposed technique achieves SOTA results without a memory buffer and clearly outperforms the reference methods.

Lifelong learning (LL) is an important ability for NLP models to learn new tasks continuously. Architecture-based approaches are reported to be effective implementations for LL models. However, it is non-trivial to extend previous approaches to domain incremental LL scenarios since they either require access to task identities in the testing phase or cannot handle samples from unseen tasks. In this paper, we propose Diana: a dynamic architecture-based lifelong learning model that tries to learn a sequence of tasks with a prompt-enhanced language model. Four types of hierarchically organized prompts are used in Diana to capture knowledge from different granularities. Specifically, we dedicate task-level prompts to capture task-specific knowledge to retain high LL performances and maintain instance-level prompts to learn knowledge shared across input samples to improve the model’s generalization performance. Moreover, we dedicate separate prompts to explicitly model unseen tasks and introduce a set of prompt key vectors to facilitate knowledge sharing between tasks. Extensive experiments demonstrate that Diana outperforms state-of-the-art LL models, especially in handling unseen tasks.

Applying neural models injected with in-domain user and product information to learn review representations of unseen or anonymous users incurs an obvious obstacle in content-based recommender systems. For the generalization of the in-domain classifier, most existing models train an extra plain-text model for the unseen domain. Without incorporating historical user and product information, such a schema makes unseen and anonymous users dissociate from the recommender system. To simultaneously learn the review representation of both existing and unseen users, this study proposed a switch knowledge distillation for domain generalization. A generalization-switch (GSwitch) model was initially applied to inject user and product information by flexibly encoding both domain-invariant and domain-specific features. By turning the status ON or OFF, the model introduced a switch knowledge distillation to learn a robust review representation that performed well for either existing or anonymous unseen users. The empirical experiments were conducted on IMDB, Yelp-2013, and Yelp-2014 by masking out users in test data as unseen and anonymous users. The comparative results indicate that the proposed method enhances the generalization capability of several existing baseline models. For reproducibility, the code for this paper is available at: https://github.com/yoyo-yun/DG_RRR.

Domain generalization is hitherto an underexplored area applied in abstractive summarization. Moreover, most existing works on domain generalization have sophisticated training algorithms. In this paper, we propose a lightweight, weight averaging based, Domain Aligned Prefix Averaging approach to domain generalization for abstractive summarization. Given a number of source domains, our method first trains a prefix for each one of them. These source prefixes generate summaries for a small number of target domain documents. The similarity of the generated summaries to their corresponding source documents is used for calculating weights required to average source prefixes. In DAPA, prefix tuning allows for lightweight finetuning, and weight averaging allows for the computationally efficient addition of new source domains. When evaluated on four diverse summarization domains, DAPA shows comparable or better performance against the baselines demonstrating the effectiveness of its prefix averaging scheme.

Conceptual metaphors present a powerful cognitive vehicle to transfer knowledge structures from a source to a target domain. Prior neural approaches focus on detecting whether natural language sequences are metaphoric or literal. We believe that to truly probe metaphoric knowledge in pre-trained language models, their capability to detect this transfer should be investigated. To this end, this paper proposes to probe the ability of GPT-3 to detect metaphoric language and predict the metaphor’s source domain without any pre-set domains. We experiment with different training sample configurations for fine-tuning and few-shot prompting on two distinct datasets. When provided 12 few-shot samples in the prompt, GPT-3 generates the correct source domain for a new sample with an accuracy of 65.15% in English and 34.65% in Spanish. GPT’s most common error is a hallucinated source domain for which no indicator is present in the sentence. Other common errors include identifying a sequence as literal even though a metaphor is present and predicting the wrong source domain based on specific words in the sequence that are not metaphorically related to the target domain.

Document-level multi-event extraction aims to extract the structural information from a given document automatically. Most recent approaches usually involve two steps: (1) modeling entity interactions; (2) decoding entity interactions into events. However, such approaches ignore a global view of inter-dependency of multiple events. Moreover, an event is decoded by iteratively merging its related entities as arguments, which might suffer from error propagation and is computationally inefficient. In this paper, we propose an alternative approach for document-level multi-event extraction with event proxy nodes and Hausdorff distance minimization. The event proxy nodes, representing pseudo-events, are able to build connections with other event proxy nodes, essentially capturing global information. The Hausdorff distance makes it possible to compare the similarity between the set of predicted events and the set of ground-truth events. By directly minimizing Hausdorff distance, the model is trained towards the global optimum directly, which improves performance and reduces training time. Experimental results show that our model outperforms previous state-of-the-art method in F1-score on two datasets with only a fraction of training time.

Document-level event argument extraction aims to identify event arguments beyond sentence level, where a significant challenge is to model long-range dependencies. Focusing on this challenge, we present a new chain reasoning paradigm for the task, which can generate decomposable first-order logic rules for reasoning. This paradigm naturally captures long-range interdependence due to the chains’ compositional nature, which also improves interpretability by explicitly modeling the reasoning process. We introduce T-norm fuzzy logic for optimization, which permits end-to-end learning and shows promise for integrating the expressiveness of logical reasoning with the generalization of neural networks. In experiments, we show that our approach outperforms previous methods by a significant margin on two standard benchmarks (over 6 points in F1).Moreover, it is data-efficient in low-resource scenarios and robust enough to defend against adversarial attacks.

Neural sequence-to-sequence models have been very successful at tasks in phonology and morphology that seemingly require a capacity for intricate linguistic generalisations. In this paper, we perform a detailed breakdown of the power of such models to capture various phonological generalisations and to benefit from exposure to one phonological rule to infer the behaviour of another similar rule. We present two types of experiments, one of which establishes the efficacy of the transformer model on 29 different processes. The second experiment type follows a priming and held-out case split where our model is exposed to two (or more) phenomena; one which is used as a primer to make the model aware of a linguistic category (e.g. voiceless stops) and a second one which contains a rule with a withheld case that the model is expected to infer (e.g. word-final devoicing with a missing training example such as b→p) results show that the transformer model can successfully model all 29 phonological phenomena considered, regardless of perceived process difficulty. We also show that the model can generalise linguistic categories and structures, such as vowels and syllables, through priming processes.

When people think of everyday things like an egg, they typically have a mental image associated with it. This allows them to correctly judge, for example, that “the yolk surrounds the shell” is a false statement. Do language models similarly have a coherent picture of such everyday things? To investigate this, we propose a benchmark dataset consisting of 100 everyday things, their parts, and the relationships between these parts, expressed as 11,720 “X relation Y?” true/false questions. Using these questions as probes, we observe that state-of-the-art pre-trained language models (LMs) like GPT-3 and Macaw have fragments of knowledge about these everyday things, but do not have fully coherent “parts mental models” (54-59% accurate, 19-43% conditional constraint violation). We propose an extension where we add a constraint satisfaction layer on top of the LM’s raw predictions to apply commonsense constraints. As well as removing inconsistencies, we find that this also significantly improves accuracy (by 16-20%), suggesting how the incoherence of the LM’s pictures of everyday things can be significantly reduced.

Argument summarisation is a promising but currently under-explored field. Recent work has aimed to provide textual summaries in the form of concise and salient short texts, i.e., key points (KPs), in a task known as Key Point Analysis (KPA). One of the main challenges in KPA is finding high-quality key point candidates from dozens of arguments even in a small corpus. Furthermore, evaluating key points is crucial in ensuring that the automatically generated summaries are useful. Although automatic methods for evaluating summarisation have considerably advanced over the years, they mainly focus on sentence-level comparison, making it difficult to measure the quality of a summary (a set of KPs) as a whole. Aggravating this problem is the fact that human evaluation is costly and unreproducible. To address the above issues, we propose a two-step abstractive summarisation framework based on neural topic modelling with an iterative clustering procedure, to generate key points which are aligned with how humans identify key points. Our experiments show that our framework advances the state of the art in KPA, with performance improvement of up to 14 (absolute) percentage points, in terms of both ROUGE and our own proposed evaluation metrics. Furthermore, we evaluate the generated summaries using a novel set-based evaluation toolkit. Our quantitative analysis demonstrates the effectiveness of our proposed evaluation metrics in assessing the quality of generated KPs. Human evaluation further demonstrates the advantages of our approach and validates that our proposed evaluation metric is more consistent with human judgment than ROUGE scores.

Do question answering (QA) modeling improvements (e.g., choice of architecture and training procedure) hold consistently across the diverse landscape of QA benchmarks? To study this question, we introduce the notion of concurrence—two benchmarks have high concurrence on a set of modeling approaches if they rank the modeling approaches similarly. We measure the concurrence between 32 QA benchmarks on a set of 20 diverse modeling approaches and find that human-constructed benchmarks have high concurrence amongst themselves, even if their passage and question distributions are very different. Surprisingly, even downsampled human-constructed benchmarks (i.e., collecting less data) and programmatically-generated benchmarks (e.g., cloze-formatted examples) have high concurrence with human-constructed benchmarks. These results indicate that, despite years of intense community focus on a small number of benchmarks, the modeling improvements studied hold broadly.

Ontological knowledge, which comprises classes and properties and their relationships, is integral to world knowledge. It is significant to explore whether Pretrained Language Models (PLMs) know and understand such knowledge. However, existing PLM-probing studies focus mainly on factual knowledge, lacking a system- atic probing of ontological knowledge. In this paper, we focus on probing whether PLMs store ontological knowledge and have a semantic un- derstanding of the knowledge rather than rote memorization of the surface form. To probe whether PLMs know ontological knowledge, we investigate how well PLMs memorize: (1) types of entities; (2) hierarchical relationships among classes and properties, e.g., Person is a subclass of Animal and Member of Sports Team is a subproperty of Member of ; (3) domain and range constraints of properties, e.g., the subject of Member of Sports Team should be a Person and the object should be a Sports Team. To further probe whether PLMs truly understand ontological knowledge beyond memorization, we comprehensively study whether they can reliably perform logical reasoning with given knowledge according to ontological entailment rules. Our probing results show that PLMs can memorize certain ontological knowledge and utilize implicit knowledge in reasoning. How- ever, both the memorizing and reasoning per- formances are less than perfect, indicating in- complete knowledge and understanding.

Recent works on instruction tuning (IT) have achieved great performance with zero-shot generalizability to unseen tasks. With additional context (e.g., task definition, examples) provided to models for fine-tuning, they achieved much higher performance than untuned models. Despite impressive performance gains, what models learn from IT remains understudied. In this work, we analyze how models utilize instructions during IT by comparing model training with altered vs. original instructions. Specifically, we create simplified task definitions by removing all semantic components and only leaving the output space information, and delusive examples that contain incorrect input-output mapping. Our experiments show that models trained on simplified task definition or delusive examples can achieve comparable performance to the ones trained on the original instructions and examples. Furthermore, we introduce a random baseline to perform zeroshot classification tasks, and find it achieves similar performance (42.6% exact-match) as IT does (43% exact-match) in low resource setting, while both methods outperform naive T5 significantly (30% per exact-match). Our analysis provides evidence that the impressive performance gain of current IT models can come from picking up superficial patterns, such as learning the output format and guessing. Our study highlights the urgent need for more reliable IT methods and evaluation.