ACL 2023 论文列表

Multi-modal emotion recognition has gained increasing attention in recent years due to its widespread applications and the advances in multi-modal learning approaches. However, previous studies primarily focus on developing models that exploit the unification of multiple modalities. In this paper, we propose that maintaining modality independence is beneficial for the model performance. According to this principle, we construct a dataset, and devise a multi-modal transformer model. The new dataset, CHinese Emotion Recognition dataset with Modality-wise Annotions, abbreviated as CHERMA, provides uni-modal labels for each individual modality, and multi-modal labels for all modalities jointly observed. The model consists of uni-modal transformer modules that learn representations for each modality, and a multi-modal transformer module that fuses all modalities. All the modules are supervised by their corresponding labels separately, and the forward information flow is uni-directionally from the uni-modal modules to the multi-modal module. The supervision strategy and the model architecture guarantee each individual modality learns its representation independently, and meanwhile the multi-modal module aggregates all information. Extensive empirical results demonstrate that our proposed scheme outperforms state-of-the-art alternatives, corroborating the importance of modality independence in multi-modal emotion recognition. The dataset and codes are availabel at https://github.com/sunjunaimer/LFMIM

The use of positional embeddings in transformer language models is widely accepted. However, recent research has called into question the necessity of such embeddings. We further extend this inquiry by demonstrating that a randomly initialized and frozen transformer language model, devoid of positional embeddings, inherently encodes strong positional information through the shrinkage of self-attention variance. To quantify this variance, we derive the underlying distribution of each step within a transformer layer. Through empirical validation using a fully pretrained model, we show that the variance shrinkage effect still persists after extensive gradient updates. Our findings serve to justify the decision to discard positional embeddings and thus facilitate more efficient pretraining of transformer language models.

Jointly fine-tuning a Pre-trained Language Model (PLM) on a pre-defined set of tasks with in-context instructions has been proven to improve its generalization performance, allowing us to build a universal language model that can be deployed across task boundaries. In this work, we explore for the first time whether this attractive property of in-context instruction learning can be extended to a scenario in which tasks are fed to the target PLM in a sequential manner. The primary objective of so-called lifelong in-context instruction learning is to improve the target PLM’s instance- and task-level generalization performance as it observes more tasks. DynaInst, the proposed method to lifelong in-context instruction learning, achieves noticeable improvements in both types of generalization, nearly reaching the upper bound performance obtained through joint training.

Automated coherence metrics constitute an important and popular way to evaluate topic models. Previous works present a mixed picture of their presumed correlation with human judgement. In this paper, we conduct a large-scale correlation analysis of coherence metrics. We propose a novel sampling approach to mine topics for the purpose of metric evaluation, and conduct the analysis via three large corpora showing that certain automated coherence metrics are correlated. Moreover, we extend the analysis to measure topical differences between corpora. Lastly, we examine the reliability of human judgement by conducting an extensive user study, which is designed as an amalgamation of different proxy tasks to derive a finer insight into the human decision-making processes. Our findings reveal some correlation between automated coherence metrics and human judgement, especially for generic corpora.

E-commerce websites (e.g. Amazon, Alibaba) have a plethora of structured and unstructured information (text and images) present on the product pages. Sellers often don’t label or mislabel values of the attributes (e.g. color, size etc.) for their products. Automatically identifying these attribute values from an eCommerce product page that contains both text and images is a challenging task, especially when the attribute value is not explicitly mentioned in the catalog. In this paper, we present a scalable solution for this problem where we pose attribute extraction problem as a question-answering task, which we solve using MXT, that consists of three key components: (i) MAG (Multimodal Adaptation Gate), (ii) Xception network, and (iii) T5 encoder-decoder. Our system consists of a generative model that generates attribute-values for a given product by using both textual and visual characteristics (e.g. images) of the product. We show that our system is capable of handling zero-shot attribute prediction (when attribute value is not seen in training data) and value-absent prediction (when attribute value is not mentioned in the text) which are missing in traditional classification-based and NER-based models respectively. We have trained our models using distant supervision, removing dependency on human labeling, thus making them practical for real-world applications. With this framework, we are able to train a single model for 1000s of (product-type, attribute) pairs, thus reducing the overhead of training and maintaining separate models. Extensive experiments on two real world datasets (total 57 attributes) show that our framework improves the absolute recall@90P by 10.16% and 6.9 from the existing state of the art models. In a popular e-commerce store, we have productionized our models that cater to 12K (product-type, attribute) pairs, and have extracted 150MM attribute values.

Large language models (LLMs) have led to a series of breakthroughs in natural language processing (NLP), partly owing to the massive amounts of world knowledge they memorize during pretraining. While many downstream applications provide the model with an informational context to aid its underlying task, how the model’s world knowledge interacts with the factual information presented in the context remains under explored. As a desirable behavior, an LLM should give precedence to the context whenever it contains task-relevant information that conflicts with the model’s memorized knowledge. This enables model predictions to be grounded in the context, which then facilitates updating specific model predictions without frequently retraining the model. By contrast, when the context is irrelevant to the task, the model should ignore it and fall back on its internal knowledge. In this paper, we undertake a first joint study of the aforementioned two properties, namely controllability and robustness, in the context of LLMs. We demonstrate that state-of-the-art T5 and PaLM models (both pretrained and finetuned) could exhibit low controllability and robustness that does not improve with increasing the model size. As a solution, we propose a simple yet effective method – knowledge aware finetuning (KAFT) – to strengthen both controllability and robustness by injecting counterfactual and irrelevant contexts to standard supervised datasets. Our comprehensive evaluation showcases the utility of KAFT across model architectures and sizes.

Document retrieval is a key stage of standard Web search engines. Existing dual-encoder dense retrievers obtain representations for questions and documents independently, allowing for only shallow interactions between them. To overcome this limitation, recent autoregressive search engines replace the dual-encoder architecture by directly generating identifiers for relevant documents in the candidate pool. However, the training cost of such autoregressive search engines rises sharply as the number of candidate documents increases. In this paper, we find that large language models (LLMs) can follow human instructions to directly generate URLs for document retrieval. Surprisingly, when providing a few Query-URL pairs as in-context demonstrations, LLMs can generate Web URLs where nearly 90% of the corresponding documents contain correct answers to open-domain questions. In this way, LLMs can be thought of as built-in search engines, since they have not been explicitly trained to map questions to document identifiers. Experiments demonstrate that our method can consistently achieve better retrieval performance than existing retrieval approaches by a significant margin on three open-domain question answering benchmarks, under both zero and few-shot settings. The code for this work can be found at https://github.com/Ziems/llm-url.

The task of generating code from a natural language description, or NL2Code, is considered a pressing and significant challenge in code intelligence. Thanks to the rapid development of pre-training techniques, surging large language models are being proposed for code, sparking the advances in NL2Code. To facilitate further research and applications in this field, in this paper, we present a comprehensive survey of 27 existing large language models for NL2Code, and also review benchmarks and metrics. We provide an intuitive comparison of all existing models on the HumanEval benchmark. Through in-depth observation and analysis, we provide some insights and conclude that the key factors contributing to the success of large language models for NL2Code are “Large Size, Premium Data, Expert Tuning”. In addition, we discuss challenges and opportunities regarding the gap between models and humans. We also create a website https://nl2code.github.io to track the latest progress through crowd-sourcing. To the best of our knowledge, this is the first survey of large language models for NL2Code, and we believe it will contribute to the ongoing development of the field.

Large language models (LLMs) have recently shown great potential for in-context learning, where LLMs learn a new task simply by conditioning on a few input-label pairs (prompts). Despite their potential, our understanding of the factors influencing end-task performance and the robustness of in-context learning remains limited. This paper aims to bridge this knowledge gap by investigating the reliance of LLMs on shortcuts or spurious correlations within prompts. Through comprehensive experiments on classification and extraction tasks, we reveal that LLMs are “lazy learners” that tend to exploit such shortcuts. Additionally, we uncover a surprising finding that larger models are more likely to utilize shortcuts in prompts during inference. Our findings provide a new perspective on evaluating robustness in in-context learning and pose new challenges for detecting and mitigating the use of shortcuts in prompts.

Recent works have shown that chain-of-thought (CoT) prompting can elicit language models to solve complex reasoning tasks, step-by-step. However, prompt-based CoT methods are dependent on very large models such as GPT-3 175B which are prohibitive to deploy at scale. In this paper, we use these large models as reasoning teachers to enable complex reasoning in smaller models and reduce model size requirements by several orders of magnitude. We propose Fine-tune-CoT, a method that generates reasoning samples from very large teacher models to fine-tune smaller models. We evaluate our method on a wide range of public models and complex tasks. We find that Fine-tune-CoT enables substantial reasoning capability in small models, far outperforming prompt-based baselines and even the teacher model in many tasks. Additionally, we extend our method by leveraging the teacher model’s ability to generate multiple distinct rationales for each original sample. Enriching the fine-tuning data with such diverse reasoning results in a substantial performance boost across datasets, even for very small models. We conduct ablations and sample studies to understand the emergence of reasoning capabilities of student models. Our code implementation and data are available at https://github.com/itsnamgyu/reasoning-teacher.

While a large body of literature suggests that large language models (LLMs) acquire rich linguistic representations, little is known about whether they adapt to linguistic biases in a human-like way. The present study probes this question by asking whether LLMs display human-like referential biases using stimuli and procedures from real psycholinguistic experiments. Recent psycholinguistic studies suggest that humans adapt their referential biases with recent exposure to referential patterns; closely replicating three relevant psycholinguistic experiments from Johnson & Arnold (2022) in an in-context learning (ICL) framework, we found that InstructGPT adapts its pronominal interpretations in response to the frequency of referential patterns in the local discourse, though in a limited fashion: adaptation was only observed relative to syntactic but not semantic biases. By contrast, FLAN-UL2 fails to generate meaningful patterns. Our results provide further evidence that contemporary LLMs discourse representations are sensitive to syntactic patterns in the local context but less so to semantic patterns. Our data and code are available at https://github.com/zkx06111/llm_priming.

Leveraging an established exercise in negotiation education, we build a novel dataset for studying how the use of language shapes bilateral bargaining. Our dataset extends existing work in two ways: 1) we recruit participants via behavioral labs instead of crowdsourcing platforms and allow participants to negotiate through audio, enabling more naturalistic interactions; 2) we add a control setting where participants negotiate only through alternating, written numeric offers. Despite the two contrasting forms of communication, we find that the average agreed prices of the two treatments are identical. But when subjects can talk, fewer offers are exchanged, negotiations finish faster, the likelihood of reaching agreement rises, and the variance of prices at which subjects agree drops substantially. We further propose a taxonomy of speech acts in negotiation and enrich the dataset with annotated speech acts. Our work also reveals linguistic signals that are predictive of negotiation outcomes.

Targeted syntactic evaluations of language models ask whether models show stable preferences for syntactically acceptable content over minimal-pair unacceptable inputs. Our best syntactic evaluation datasets, however, provide substantially less linguistic context than models receive during pretraining. This mismatch raises an important question: how robust are models’ syntactic judgements across different contexts? In this paper, we vary the input contexts based on: length, the types of syntactic phenomena it contains, and whether or not there are grammatical violations. We find that model judgements are generally robust when placed in randomly sampled linguistic contexts, but are unstable when contexts match the test stimuli in syntactic structure. Among all tested models (GPT-2 and five variants of OPT), we find that model performance is affected when we provided contexts with matching syntactic structure: performance significantly improves when contexts are acceptable, and it significantly declines when they are unacceptable. This effect is amplified by the length of the context, except for unrelated inputs. We show that these changes in model performance are not explainable by acceptability-preserving syntactic perturbations. This sensitivity to highly specific syntactic features of the context can only be explained by the models’ implicit in-context learning abilities.

During language acquisition, children follow a typical sequence of learning stages, whereby they first learn to categorize phonemes before they develop their lexicon and eventually master increasingly complex syntactic structures. However, the computational principles that lead to this learning trajectory remain largely unknown. To investigate this, we here compare the learning trajectories of deep language models to those of human children. Specifically, we test whether, during its training, GPT-2 exhibits stages of language acquisition comparable to those observed in children aged between 18 months and 6 years. For this, we train 48 GPT-2 models from scratch and evaluate their syntactic and semantic abilities at each training step, using 96 probes curated from the BLiMP, Zorro and BIG-Bench benchmarks. We then compare these evaluations with the behavior of 54 children during language production. Our analyses reveal three main findings. First, similarly to children, the language models tend to learn linguistic skills in a systematic order. Second, this learning scheme is parallel: the language tasks that are learned last improve from the very first training steps. Third, some – but not all – learning stages are shared between children and these language models. Overall, these results shed new light on the principles of language acquisition, and highlight important divergences in how humans and modern algorithms learn to process natural language.

Automatic text simplification systems help to reduce textual information barriers on the internet. However, for languages other than English, only few parallel data to train these systems exists. We propose a two-step approach to overcome this data scarcity issue. First, we fine-tuned language models on a corpus of German Easy Language, a specific style of German. Then, we used these models as decoders in a sequence-to-sequence simplification task. We show that the language models adapt to the style characteristics of Easy Language and output more accessible texts. Moreover, with the style-specific pre-training, we reduced the number of trainable parameters in text simplification models. Hence, less parallel data is sufficient for training. Our results indicate that pre-training on unaligned data can reduce the required parallel data while improving the performance on downstream tasks.

Societal biases present in pre-trained large language models are a critical issue as these models have been shown to propagate biases in countless downstream applications, rendering them unfair towards specific groups of people. Since large-scale retraining of these models from scratch is both time and compute-expensive, a variety of approaches have been previously proposed that de-bias a pre-trained model. While the majority of current state-of-the-art debiasing methods focus on changes to the training regime, in this paper, we propose data intervention strategies as a powerful yet simple technique to reduce gender bias in pre-trained models. Specifically, we empirically show that by fine-tuning a pre-trained model on only 10 debiased (intervened) training examples, the tendency to favor any gender is significantly reduced. Since our proposed method only needs a few training examples, we argue that our few-shot de-biasing approach is highly feasible and practical. Through extensive experimentation, we show that our de-biasing technique performs better than competitive state-of-the-art baselines with minimal loss in language modeling ability.

Language models (LMs) often generate incoherent outputs: they refer to events and entity states that are incompatible with the state of the world described in inputs. We introduce SITUATIONSUPERVISION, a family of approaches for improving coherence in LMs by training them to construct and condition on explicit representations of entities and their states. SITUATIONSUPERVISION has two components: an *auxiliary situation modeling* task that trains models to predict entity state representations in context, and a *latent state inference* procedure that imputes these states from partially annotated training data. SITUATIONSUPERVISION can be applied via fine-tuning (by supervising LMs to encode state variables in their hidden representations) and prompting (by inducing LMs to interleave textual descriptions of entity states with output text). In both cases, it requires only a small number of state annotations to produce substantial coherence improvements (up to an 16% reduction in errors), showing that standard LMs can be efficiently adapted to explicitly model language and aspects of its meaning.

Investigating whether pre-trained language models (LMs) can function as knowledge bases (KBs) has raised wide research interests recently. However, existing works focus on simple, triple-based, relational KBs, but omit more sophisticated, logic-based, conceptualised KBs such as OWL ontologies. To investigate an LM’s knowledge of ontologies, we propose OntoLAMA, a set of inference-based probing tasks and datasets from ontology subsumption axioms involving both atomic and complex concepts. We conduct extensive experiments on ontologies of different domains and scales, and our results demonstrate that LMs encode relatively less background knowledge of Subsumption Inference (SI) than traditional Natural Language Inference (NLI) but can improve on SI significantly when a small number of samples are given. We will open-source our code and datasets.

Transformer-based Language Models (LMs) have achieved impressive results on natural language understanding tasks, but they can also generate toxic text such as insults, threats, and profanity, limiting their real-world applications. To overcome this issue, a few text generation approaches aim to detoxify toxic texts using additional LMs or perturbations. However, previous methods require excessive memory, computations, and time which are serious bottlenecks in their real-world application. To address such limitations, we propose an effective yet efficient method for language detoxification using an attribute-discriminative latent space. Specifically, we project the latent space of an original Transformer LM onto a discriminative latent space that well-separates texts by their attributes using a projection block and an attribute discriminator. This allows the LM to control the text generation to be non-toxic with minimal memory and computation overhead. We validate our model, Attribute-Discriminative Language Model (ADLM) on detoxified language and dialogue generation tasks, on which our method significantly outperforms baselines both in performance and efficiency.