EMNLP 2023 论文列表

We study model extraction attacks in natural language processing (NLP) where attackers aim to steal victim models by repeatedly querying the open Application Programming Interfaces (APIs). Recent works focus on limited-query budget settings and adopt random sampling or active learning-based sampling strategies on publicly available, unannotated data sources. However, these methods often result in selected queries that lack task relevance and data diversity, leading to limited success in achieving satisfactory results with low query costs. In this paper, we propose MeaeQ (Model extraction attack with efficient Queries), a straightforward yet effective method to address these issues. Specifically, we initially utilize a zero-shot sequence inference classifier, combined with API service information, to filter task-relevant data from a public text corpus instead of a problem domain-specific dataset. Furthermore, we employ a clustering-based data reduction technique to obtain representative data as queries for the attack. Extensive experiments conducted on four benchmark datasets demonstrate that MeaeQ achieves higher functional similarity to the victim model than baselines while requiring fewer queries.

Language models (LMs) with less than 100B parameters are known to perform poorly on chain-of-thought (CoT) reasoning in contrast to large LMs when solving unseen tasks. In this work, we aim to equip smaller LMs with the step-by-step reasoning capability by instruction tuning with CoT rationales. In order to achieve this goal, we first introduce a new instruction-tuning dataset called the CoT Collection, which augments the existing Flan Collection (including only 9 CoT tasks) with additional 1.84 million rationales across 1,060 tasks. We show that CoT fine-tuning Flan-T5 (3B \& 11B) with CoT Collection enables smaller LMs to have better CoT capabilities on unseen tasks. On the BIG-Bench-Hard (BBH) benchmark, we report an average improvement of +4.34% (Flan-T5 3B) and +2.60% (Flan-T5 11B), in terms of zero-shot task accuracy. Furthermore, we show that instruction tuning with CoT Collection allows LMs to possess stronger few-shot learning capabilities on 4 domain-specific tasks, resulting in an improvement of +2.24% (Flan-T5 3B) and +2.37% (Flan-T5 11B), even outperforming ChatGPT utilizing demonstrations until the max length by a +13.98% margin. Our code, the CoT Collection data, and model checkpoints are publicly available.

Reasoning over spans of tokens from different parts of the input is essential for natural language understanding (NLU) tasks such as fact-checking (FC), machine reading comprehension (MRC) or natural language inference (NLI). However, existing highlight-based explanations primarily focus on identifying individual important features or interactions only between adjacent tokens or tuples of tokens. Most notably, there is a lack of annotations capturing the human decision-making process with respect to the necessary interactions for informed decision-making in such tasks. To bridge this gap, we introduce SpanEx, a multi-annotator dataset of human span interaction explanations for two NLU tasks: NLI and FC. We then investigate the decision-making processes of multiple fine-tuned large language models in terms of the employed connections between spans in separate parts of the input and compare them to the human reasoning processes. Finally, we present a novel community detection based unsupervised method to extract such interaction explanations. We make the code and the dataset available on [Github](https://github.com/copenlu/spanex). The dataset is also available on [Huggingface datasets](https://huggingface.co/datasets/copenlu/spanex).

Multilingual Language Models offer a way to incorporate multiple languages in one model and utilize cross-language transfer learning to improve performance for different Natural Language Processing (NLP) tasks. Despite progress in multilingual models, not all languages are supported as well, particularly in low-resource settings. In this work, we investigate the linguistic representation of different languages in multilingual models. We start by asking the question which languages are supported in popular multilingual models and which languages are left behind. Then, for included languages, we look at models' learned representations based on language family and dialect and try to understand how models' learned representations for (1) seen and (2) unseen languages vary across different language groups. In addition, we test and analyze performance on downstream tasks such as text generation and Named Entity Recognition. We observe from our experiments that community-centered models---models that focus on languages of a given family or geographical location and are built by communities who speak them---perform better at distinguishing between languages in the same family for low-resource languages. Our paper contributes to the literature in understanding multilingual models and their shortcomings and offers insights on potential ways to improve them.

This paper focuses on using natural language descriptions to enhance predictive models in the chemistry field. Conventionally, chemoinformatics models are trained with extensive structured data manually extracted from the literature. In this paper, we introduce TextReact, a novel method that directly augments predictive chemistry with texts retrieved from the literature. TextReact retrieves text descriptions relevant for a given chemical reaction, and then aligns them with the molecular representation of the reaction. This alignment is enhanced via an auxiliary masked LM objective incorporated in the predictor training. We empirically validate the framework on two chemistry tasks: reaction condition recommendation and one-step retrosynthesis. By leveraging text retrieval, TextReact significantly outperforms state-of-the-art chemoinformatics models trained solely on molecular data.

Quality estimation models have been developed to assess the corrections made by grammatical error correction (GEC) models when the reference or gold-standard corrections are not available. An ideal quality estimator can be utilized to combine the outputs of multiple GEC systems by choosing the best subset of edits from the union of all edits proposed by the GEC base systems. However, we found that existing GEC quality estimation models are not good enough in differentiating good corrections from bad ones, resulting in a low F0.5 score when used for system combination. In this paper, we propose GRECO, a new state-of-the-art quality estimation model that gives a better estimate of the quality of a corrected sentence, as indicated by having a higher correlation to the F0.5 score of a corrected sentence. It results in a combined GEC system with a higher F0.5 score. We also propose three methods for utilizing GEC quality estimation models for system combination with varying generality: model-agnostic, model-agnostic with voting bias, and model-dependent method. The combined GEC system outperforms the state of the art on the CoNLL-2014 test set and the BEA-2019 test set, achieving the highest F0.5 scores published to date.

Recently, contrastive learning has become a key component in fine-tuning code search models for software development efficiency and effectiveness. It pulls together positive code snippets while pushing negative samples away given search queries. Among contrastive learning, InfoNCE is the most widely used loss function due to its better performance. However, the following problems in negative samples of InfoNCE may deteriorate its representation learning: 1) The existence of false negative samples in large code corpora due to duplications. 2). The failure to explicitly differentiate between the potential relevance of negative samples. As an example, a bubble sorting algorithm example is less ``negative'' than a file saving function for the quick sorting algorithm query. In this paper, we tackle the above problems by proposing a simple yet effective Soft-InfoNCE loss that inserts weight terms into InfoNCE. In our proposed loss function, we apply three methods to estimate the weights of negative pairs and show that the vanilla InfoNCE loss is a special case of Soft-InfoNCE. Theoretically, we analyze the effects of Soft-InfoNCE on controlling the distribution of learnt code representations and on deducing a more precise mutual information estimation. We furthermore discuss the superiority of proposed loss functions with other design alternatives. Extensive experiments demonstrate the effectiveness of Soft-InfoNCE and weights estimation methods under state-of-the-art code search models on a large-scale public dataset consisting of six programming languages.

Despite its relevance, the maturity of NLP for social media pales in comparison with general-purpose models, metrics and benchmarks. This fragmented landscape makes it hard for the community to know, for instance, given a task, which is the best performing model and how it compares with others. To alleviate this issue, we introduce a unified benchmark for NLP evaluation in social media, SuperTweetEval, which includes a heterogeneous set of tasks and datasets combined, adapted and constructed from scratch. We benchmarked the performance of a wide range of models on SuperTweetEval and our results suggest that, despite the recent advances in language modelling, social media remains challenging.

Unsupervised neural machine translation has recently made remarkable strides, achieving impressive results with the exclusive use of monolingual corpora. Nonetheless, these methods still exhibit fundamental flaws, such as confusing similar words. A straightforward remedy to rectify this drawback is to employ bilingual dictionaries, however, high-quality bilingual dictionaries can be costly to obtain. To overcome this limitation, we propose a method that incorporates images at the word level to augment the lexical mappings. Specifically, our method inserts visual representations into the model, modifying the corresponding embedding layer information. Besides, a visible matrix is adopted to isolate the impact of images on other unrelated words. Experiments on the Multi30k dataset with over 300,000 self-collected images validate the effectiveness in generating more accurate word translation, achieving an improvement of up to $+$2.81 BLEU score, which is comparable or even superior to using bilingual dictionaries.

People rely heavily on context to enrich meaning beyond what is literally said, enabling concise but effective communication. To interact successfully and naturally with people, user-facing artificial intelligence systems will require similar skills in pragmatics: relying on various types of context --- from shared linguistic goals and conventions, to the visual and embodied world --- to use language effectively. We survey existing grounded settings and pragmatic modeling approaches and analyze how the task goals, environmental contexts, and communicative affordances in each work enrich linguistic meaning. We present recommendations for future grounded task design to naturally elicit pragmatic phenomena, and suggest directions that focus on a broader range of communicative contexts and affordances.

Question answering plays a pivotal role in human daily life because it involves our acquisition of knowledge about the world. However, due to the dynamic and ever-changing nature of real-world facts, the answer can be completely different when the time constraint in the question changes. Recently, Large Language Models (LLMs) have shown remarkable intelligence in question answering, while our experiments reveal that the aforementioned problems still pose a significant challenge to existing LLMs. This can be attributed to the LLMs' inability to perform rigorous reasoning based on surface-level text semantics. To overcome this limitation, rather than requiring LLMs to directly answer the question, we propose a novel approach where we reframe the $\textbf{Q}$uestion $\textbf{A}$nswering task $\textbf{a}$s $\textbf{P}$rogramming ($\textbf{QAaP}$). Concretely, by leveraging modern LLMs' superior capability in understanding both natural language and programming language, we endeavor to harness LLMs to represent diversely expressed text as well-structured code and select the best matching answer from multiple candidates through programming. We evaluate our QAaP framework on several time-sensitive question answering datasets and achieve decent improvement, up to $14.5$% over strong baselines.

In code vulnerability detection tasks, a detector trained on a label-rich source domain fails to provide accurate prediction on new or unseen target domains due to the lack of labeled training data on target domains. Previous studies mainly utilize domain adaptation to perform cross-domain vulnerability detection. But they ignore the negative effect of private semantic characteristics of the target domain for domain alignment, which easily causes the problem of negative transfer. In addition, these methods forcibly reduce the distribution discrepancy between domains and do not take into account the interference of irrelevant target instances for distributional domain alignment, which leads to the problem of excessive alignment. To address the above issues, we propose a novel cross-domain code vulnerability detection framework named MNCRI. Specifically, we introduce mutual nearest neighbor contrastive learning to align the source domain and target domain geometrically, which could align the common semantic characteristics of two domains and separate out the private semantic characteristics of each domain. Furthermore, we introduce an instance re-weighting scheme to alleviate the problem of excessive alignment. This scheme dynamically assign different weights to instances, reducing the contribution of irrelevant instances so as to achieve better domain alignment. Finally, extensive experiments demonstrate that MNCRI significantly outperforms state-of-the-art cross-domain code vulnerability detection methods by a large margin.

The research study of detecting multiple intents and filling slots is becoming more popular because of its relevance to complicated real-world situations. Recent advanced approaches, which are joint models based on graphs, might still face two potential issues: (i) the uncertainty introduced by constructing graphs based on preliminary intents and slots, which may transfer intent-slot correlation information to incorrect label node destinations, and (ii) direct incorporation of multiple intent labels for each token w.r.t. token-level intent voting might potentially lead to incorrect slot predictions, thereby hurting the overall performance. To address these two issues, we propose a joint model named MISCA. Our MISCA introduces an intent-slot co-attention mechanism and an underlying layer of label attention mechanism. These mechanisms enable MISCA to effectively capture correlations between intents and slot labels, eliminating the need for graph construction. They also facilitate the transfer of correlation information in both directions: from intents to slots and from slots to intents, through multiple levels of label-specific representations, without relying on token-level intent information. Experimental results show that MISCA outperforms previous models, achieving new state-of-the-art overall accuracy performances on two benchmark datasets MixATIS and MixSNIPS. This highlights the effectiveness of our attention mechanisms.

Entity linking aims to link ambiguous mentions to their corresponding entities in a knowledge base. One of the key challenges comes from insufficient labeled data for specific domains. Although dense retrievers have achieved excellent performance on several benchmarks, their performance decreases significantly when only a limited amount of in-domain labeled data is available. In such few-shot setting, we revisit the sparse retrieval method, and propose an ELECTRA-based keyword extractor to denoise the mention context and construct a better query expression. For training the extractor, we propose a distant supervision method to automatically generate training data based on overlapping tokens between mention contexts and entity descriptions. Experimental results on the ZESHEL dataset demonstrate that the proposed method outperforms state-of-the-art models by a significant margin across all test domains, showing the effectiveness of keyword-enhanced sparse retrieval.

Emotion recognition in conversation (ERC) has attracted increasing attention in natural language processing community. Previous work commonly first extract semantic-view features via fine-tuning PLMs, then models context-view features based on the obtained semantic-view features by various graph neural networks. However, it is difficult to fully model interaction between utterances simply through a graph neural network and the features at semantic-view and context-view are not well aligned. Moreover, the previous parametric learning paradigm struggle to learn the patterns of tail class given fewer instances. To this end, we treat the pre-trained conversation model as a prior knowledge base and from which we elicit correlations between utterances by a probing procedure. And we adopt supervised contrastive learning to align semantic-view and context-view features, these two views of features work together in a complementary manner, contributing to ERC from distinct perspectives. Meanwhile, we propose a new semi-parametric paradigm of inferencing through memorization to solve the recognition problem of tail class samples. We consistently achieve state-of-the-art results on four widely used benchmarks. Extensive experiments demonstrate the effectiveness of our proposed multi-view feature alignment and memorization.

Previous work on noun classification implies that gender systems are inherently optimized to accommodate communicative pressures on human language learning and processing (Dye. et al 2017, 2018). They state that languages make use of either grammatical (e.g., gender) or probabilistic (pre-nominal modifiers) to smoothe the entropy of nouns in context. We show that even languages that are considered genderless, like Mandarin Chinese, possess a noun classification device that plays the same functional role as gender markers. Based on close to 1M Mandarin noun phrases extracted from the Leipzig Corpora Collection (Goldhahn et al. 2012) and their corresponding fastText embeddings (Bojanowski et al. 2016), we show that noun-classifier combinations are sensitive to same frequency, similarity, and co-occurrence interactions that structure gender systems. We also present the first study of the effects of the interaction between grammatical and probabilisitic noun classification.

Text simplification systems rewrite text to make it more readable while preserving its content. However, what makes a text easy to read depends on the intended readers. Recent work has shown that pre-trained language models can simplify text using a wealth of techniques to control output simplicity, ranging from specifying only the desired reading grade level, to directly specifying low-level edit operations. Yet it remains unclear how to set these control parameters in practice. Existing approaches set them at the corpus level, disregarding the complexity of individual inputs and considering only one level of output complexity. In this work, we conduct an empirical study to understand how different control mechanisms impact the adequacy and simplicity of text simplification systems. Based on these insights, we introduce a simple method that predicts the edit operations required for simplifying a text for a specific grade level on an instance-per-instance basis. This approach improves the quality of the simplified outputs over corpus-level search-based heuristics.

The probing classifiers framework has been employed for interpreting deep neural network models for a variety of natural language processing (NLP) applications. Studies, however, have largely focused on sentencelevel NLP tasks. This work is the first to apply the probing paradigm to representations learned for document-level information extraction (IE). We designed eight embedding probes to analyze surface, semantic, and event-understanding capabilities relevant to document-level event extraction. We apply them to the representations acquired by learning models from three different LLM-based document-level IE approaches on a standard dataset. We found that trained encoders from these models yield embeddings that can modestly improve argument detections and labeling but only slightly enhance event-level tasks, albeit trade-offs in information helpful for coherence and event-type prediction. We further found that encoder models struggle with document length and cross-sentence discourse.

The increasing ubiquity of language technology necessitates a shift towards considering cultural diversity in the machine learning realm, particularly for subjective tasks that rely heavily on cultural nuances, such as Offensive Language Detection (OLD). Current understanding underscores that these tasks are substantially influenced by cultural values, however, a notable gap exists in determining if cultural features can accurately predict the success of cross-cultural transfer learning for such subjective tasks. Addressing this, our study delves into the intersection of cultural features and transfer learning effectiveness. The findings reveal that cultural value surveys indeed possess a predictive power for cross-cultural transfer learning success in OLD tasks, and that it can be further improved using offensive word distance. Based on these results, we advocate for the integration of cultural information into datasets. Additionally, we recommend leveraging data sources rich in cultural information, such as surveys, to enhance cultural adaptability. Our research signifies a step forward in the quest for more inclusive, culturally sensitive language technologies.