EMNLP 2022 Papers

Sentence representations are essential in many NLP tasks operating at the sentence level.Recently, research attention has shifted towards learning how to represent sentences without any annotations, i.e., unsupervised representation learning. Despite the benefit of training without supervised data, there is still a performance penalty compared to supervised methods.Furthermore, the supervised-unsupervised performance gap widens as we reduce the model size. In this paper, we propose an unsupervised sentence representation method to reduce the supervised-unsupervised performance gap, especially for smaller models. Utilizing the concept for knowledge distillation, we derive a distillation framework comprising two training objectives, control and generalize, called ConGen. Experiments on semantic textual similarity (STS), text classification (transfer), and natural language inference (NLI) tasks show that ConGen is on par with supervised training even on smaller models.Furthermore, our method consistently outperformed competitors on multilingual STS.The code and models are available at https://github.com/KornWtp/ConGen.

Inspired by the success of contrastive learning in natural language processing, we incorporate contrastive learning into the conditional masked language model which is extensively used in non-autoregressive neural machine translation (NAT). Accordingly, we propose a Contrastive Non-autoregressive Neural Machine Translation (Con-NAT) model. Con-NAT optimizes the similarity of several different representations of the same token in the same sentence. We propose two methods to obtain various representations: Contrastive Common Mask and Contrastive Dropout. Positive pairs are various different representations of the same token, while negative pairs are representations of different tokens. In the feature space, the model with contrastive loss pulls positive pairs together and pushes negative pairs away. We conduct extensive experiments on six translation directions with different data sizes. The results demonstrate that Con-NAT showed a consistent and significant improvement in fully and iterative NAT. Con-NAT is state-of-the-art on WMT’16 Ro-En (34.18 BLEU).

In this paper, we propose a novel architecture called Composition Attention Grammars (CAGs) that recursively compose subtrees into a single vector representation with a composition function, and selectively attend to previous structural information with a self-attention mechanism. We investigate whether these components—the composition function and the self-attention mechanism—can both induce human-like syntactic generalization. Specifically, we train language models (LMs) with and without these two components with the model sizes carefully controlled, and evaluate their syntactic generalization performance against six test circuits on the SyntaxGym benchmark. The results demonstrated that the composition function and the self-attention mechanism both play an important role to make LMs more human-like, and closer inspection of linguistic phenomenon implied that the composition function allowed syntactic features, but not semantic features, to percolate into subtree representations.

Composing Ci (also widely known as Song Ci), a special type of classical Chinese poetry, requires to follow particular format once their tune patterns are given. To automatically generate a well-formed Ci, text generation systems should strictly take into account pre-defined rigid formats (e.g., length and rhyme). Yet, most existing approaches regard Ci generation as a conventional sequence-to-sequence task and use autoregressive models, while it is challenging for such models to properly handle the constraints (according to tune patterns) of Ci during the generation process. Moreover, consider that with the format prepared, Ci generation can be operated by an efficient synchronous process, where autoregressive models are limited in doing so since they follow the character-by-character generation protocol. Therefore, in this paper, we propose to compose Ci through a non-autoregressive approach, which not only ensure that the generation process accommodates tune patterns by controlling the rhythm and essential meaning of each sentence, but also allow the model to perform synchronous generation. In addition, we further improve our approach by applying reinforcement learning to the generation process with the rigid constraints of Ci as well as the diversity in content serving as rewards, so as to further maintain the format and content requirement. Experiments on a collected Ci dataset confirm that our proposed approach outperforms strong baselines and previous studies in terms of both automatic evaluation metrics and human judgements.

The choice of geometric space for knowledge graph (KG) embeddings can have significant effects on the performance of KG completion tasks. The hyperbolic geometry has been shown to capture the hierarchical patterns due to its tree-like metrics, which addressed the limitations of the Euclidean embedding models. Recent explorations of the complex hyperbolic geometry further improved the hyperbolic embeddings for capturing a variety of hierarchical structures. However, the performance of the hyperbolic KG embedding models for non-transitive relations is still unpromising, while the complex hyperbolic embeddings do not deal with multi-relations. This paper aims to utilize the representation capacity of the complex hyperbolic geometry in multi-relational KG embeddings. To apply the geometric transformations which account for different relations and the attention mechanism in the complex hyperbolic space, we propose to use the fast Fourier transform (FFT) as the conversion between the real and complex hyperbolic space. Constructing the attention-based transformations in the complex space is very challenging, while the proposed Fourier transform-based complex hyperbolic approaches provide a simple and effective solution. Experimental results show that our methods outperform the baselines, including the Euclidean and the real hyperbolic embedding models.

Neural machine translation (NMT) is often criticized for failures that happenwithout awareness. The lack of competency awareness makes NMT untrustworthy. This is in sharp contrast to human translators who give feedback or conduct further investigations whenever they are in doubt about predictions. To fill this gap, we propose a novel competency-aware NMT by extending conventional NMT with a self-estimator, offering abilities to translate a source sentence and estimate its competency.The self-estimator encodes the information of the decoding procedure and then examines whether it can reconstruct the original semantics of the source sentence. Experimental results on four translation tasks demonstrate that the proposed method not only carries out translation tasks intact but also delivers outstanding performance on quality estimation.Without depending on any reference or annotated data typically required by state-of-the-art metric and quality estimation methods, our model yields an even higher correlation with human quality judgments than a variety of aforementioned methods, such as BLEURT, COMET, and BERTScore. Quantitative and qualitative analyses show better robustness of competency awareness in our model.

We compare the 0-shot performance of a neural caption-based image retriever when given as input either human-produced captions or captions generated by a neural captioner. We conduct this comparison on the recently introduced ImageCoDe data-set (Krojer et al. 2022), which contains hard distractors nearly identical to the images to be retrieved. We find that the neural retriever has much higher performance when fed neural rather than human captions, despite the fact that the former, unlike the latter, were generated without awareness of the distractors that make the task hard. Even more remarkably, when the same neural captions are given to human subjects, their retrieval performance is almost at chance level. Our results thus add to the growing body of evidence that, even when the “language” of neural models resembles English, this superficial resemblance might be deeply misleading.

In e-commerce, the salience of commonsense knowledge (CSK) is beneficial for widespread applications such as product search and recommendation. For example, when users search for “running” in e-commerce, they would like to find products highly related to running, such as “running shoes” rather than “shoes”. Nevertheless, many existing CSK collections rank statements solely by confidence scores, and there is no information about which ones are salient from a human perspective. In this work, we define the task of supervised salience evaluation, where given a CSK triple, the model is required to learn whether the triple is salient or not. In addition to formulating the new task, we also release a new Benchmark dataset of Salience Evaluation in E-commerce (BSEE) and hope to promote related research on commonsense knowledge salience evaluation. We conduct experiments in the dataset with several representative baseline models. The experimental results show that salience evaluation is a hard task where models perform poorly on our evaluation set. We further propose a simple but effective approach, PMI-tuning, which shows promise for solving this novel problem. Code is available in https://github.com/OpenBGBenchmark/OpenBG-CSK.

Relation extraction (RE) is an important task in natural language processing which aims to annotate the relation between two given entities, which requires a deep understanding of the running text. To import model performance, existing approaches leverage syntactic information to facilitate the relation extraction process, where they mainly focus on dependencies among words while paying limited attention to other types of syntactic structure. Considering that combinatory categorial grammar (CCG) is a lexicalized grammatical formalism that carries the syntactic and semantic knowledge for text understanding, we propose an alternative solution for RE that takes advantage of CCG to detect the relation between entities. In doing so, we perform a multi-task learning process to learn from RE and auto-annotated CCG supertags, where an attention mechanism is performed over all input words to distinguish the important ones for RE with the attention weights guided by the supertag decoding process. We evaluate our model on two widely used English benchmark datasets (i.e., ACE2005EN and SemEval 2010 Task 8 datasets) for RE, where the effectiveness of our approach is demonstrated by the experimental results with our approach achieving state-of-the-art performance on both datasets.

Understanding rich narratives, such as dialogues and stories, often requires natural language processing systems to access relevant knowledge from commonsense knowledge graphs. However, these systems typically retrieve facts from KGs using simple heuristics that disregard the complex challenges of identifying situationally-relevant commonsense knowledge (e.g., contextualization, implicitness, ambiguity).In this work, we propose the new task of commonsense fact linking, where models are given contexts and trained to identify situationally-relevant commonsense knowledge from KGs. Our novel benchmark, ComFact, contains ~293k in-context relevance annotations for commonsense triplets across four stylistically diverse dialogue and storytelling datasets. Experimental results confirm that heuristic fact linking approaches are imprecise knowledge extractors. Learned fact linking models demonstrate across-the-board performance improvements (~34.6% F1) over these heuristics. Furthermore, improved knowledge retrieval yielded average downstream improvements of 9.8% for a dialogue response generation task. However, fact linking models still significantly underperform humans, suggesting our benchmark is a promising testbed for research in commonsense augmentation of NLP systems.

We study video-grounded dialogue generation, where a response is generated based on the dialogue context and the associated video. The primary challenges of this task lie in (1) the difficulty of integrating video data into pre-trained language models (PLMs) which presents obstacles to exploiting the power of large-scale pre-training; and (2) the necessity of taking into account the complementarity of various modalities throughout the reasoning process. Although having made remarkable progress in video-grounded dialogue generation, existing methods still fall short when it comes to integrating with PLMs in a way that allows information from different modalities to complement each other. To alleviate these issues, we first propose extracting pertinent information from videos and turning it into reasoning paths that are acceptable to PLMs. Additionally, we propose a multi-agent reinforcement learning method to collaboratively perform reasoning on different modalities (i.e., video and dialogue context). Empirical experiment results on two public datasets indicate that the proposed model can significantly outperform state-of-the-art models by large margins on both automatic and human evaluations.

In this paper, we propose the CodeRetriever model, which learns the function-level code semantic representations through large-scale code-text contrastive pre-training. We adopt two contrastive learning schemes in CodeRetriever: unimodal contrastive learning and bimodal contrastive learning. For unimodal contrastive learning, we design an unsupervised learning approach to build semantic-related code pairs based on the documentation and function name. For bimodal contrastive learning, we leverage the documentation and in-line comments of code to build code-text pairs. Both contrastive objectives can fully leverage large-scale code corpus for pre-training. Extensive experimental results show that CodeRetriever achieves new state-of-the-art with significant improvement over existing code pre-trained models, on eleven domain/language-specific code search tasks with six programming languages in different code granularity (function-level, snippet-level and statement-level).These results demonstrate the effectiveness and robustness of CodeRetriever.The codes and resources are available at https://github.com/microsoft/AR2/tree/main/CodeRetriever.

Developing models that can automatically generate detailed code explanation can greatly benefit software maintenance and programming education. However, existing code-to-text generation models often produce only high-level summaries of code that do not capture implementation-level choices essential for these scenarios. To fill in this gap, we propose the code explanation generation task. We first conducted a human study to identify the criteria for high-quality explanatory docstring for code. Based on that, we collected and refined a large-scale code docstring corpus and formulated automatic evaluation metrics that best match human assessments. Finally, we present a multi-stage fine-tuning strategy and baseline models for the task. Our experiments show that (1) our refined training dataset lets models achieve better performance in the explanation generation tasks compared to larger-scale unrefined data (15x larger), and (2) fine-tuned models can generate well-structured long docstrings comparable to human-written ones. We envision our training dataset, human-evaluation protocol, recommended metrics, and fine-tuning strategy can boost future code explanation research. The code and annotated data are available at https://github.com/subercui/CodeExp.

Code vulnerability detection is a fundamental and challenging task in the software security field. Existing research works aim to learn semantic information from the source code by utilizing NLP technologies. However, in vulnerability detection tasks, some vulnerable samples are very similar to non-vulnerable samples, which are difficult to identify. To address this issue and improve detection performance, we introduce the k-nearest neighbor mechanism which retrieves multiple neighbor samples and utilizes label information of retrieved neighbor samples to provide help for model predictions. Besides, we use supervised contrastive learning to make the model learn the discriminative representation and ensure that label information of retrieved neighbor samples is as consistent as possible with the label information of testing samples. Extensive experiments show that our method can achieve obvious performance improvements compared to baseline models.

Currently, researchers focus on generating codes from the requirement documents. However, current approaches still perform poorly on some requirements needing complex problem-solving skills. In reality, to tackle such complex requirements, instead of directly translating requirement documents into codes, software engineers write codes via unified modeling language diagrams, such as flowcharts, an intermediate tool to analyze and visualize the system. Therefore, we propose a new source code generation task, that is, to generate source code from flowcharts with texts. We manually construct a benchmark dataset containing 320 flowcharts with their corresponding source codes. Obviously, it is not straightforward to employ the current approaches for the new source code generation task since (1) the flowchart is a graph that contains various structures, including loop, selection, and others which is different from texts; (2) the connections between nodes in the flowchart are abundant and diverse which need to be carefully handled. To solve the above problems, we propose a two-stage code generation model. In the first stage, a structure recognition algorithm is employed to transform the flowchart into pseudo-code containing the structural conventions of a typical programming language such as while, if. In the second stage, a code generation model is employed to convert the pseudo-code into code. Experimental results show that the proposed approach can achieve some improvement over the baselines.

Code-switching has seen growing interest in recent years as an important multilingual NLP phenomenon. Generating code-switched text for data augmentation has been sufficiently well-explored. However, there is no prior work on generating code-switched text with fine-grained control on the degree of code-switching and the lexical choices used to convey formality. We present CoCoa, an encoder-decoder translation model that converts monolingual Hindi text to Hindi-English code-switched text with both encoder-side and decoder-side interventions to achieve fine-grained controllable generation. CoCoa can be invoked at test-time to synthesize code-switched text that is simultaneously faithful to syntactic and lexical attributes relevant to code-switching. CoCoa outputs were subjected to rigorous subjective and objective evaluations. Human evaluations establish that our outputs are of superior quality while being faithful to desired attributes. We show significantly improved BLEU scores when compared with human-generated code-switched references. Compared to competitive baselines, we show 10% reduction in perplexity on a language modeling task and also demonstrate clear improvements on a downstream code-switched sentiment analysis task.

Intent discovery is to mine new intents from user utterances, which are not present in the set of manually predefined intents. Previous approaches to intent discovery usually automatically cluster novel intents with prior knowledge from intent-labeled data in a semi-supervised way. In this paper, we focus on the discriminative user utterance representation learning and the compactness of the learned intent clusters. We propose a novel semi-supervised intent discovery framework CoCoID with two essential components: contrastive user utterance representation learning and intra-cluster knowledge distillation. The former attempts to detect similar and dissimilar intents from a minibatch-wise perspective. The latter regularizes the predictive distribution of the model over samples in a cluster-wise way. We conduct experiments on both real-life challenging datasets (i.e., CLINC and BANKING) that are curated to emulate the true environment of commercial/production systems and traditional datasets (i.e., StackOverflow and DBPedia) to evaluate the proposed CoCoID. Experiment results demonstrate that our model substantially outperforms state-of-the-art intent discovery models (12 baselines) by over 1.4 ACC and ARI points and 1.1 NMI points across the four datasets. Further analyses suggest that CoCoID is able to learn contrastive representations and compact clusters for intent discovery.

Recent graph-based models for joint multiple intent detection and slot filling have obtained promising results through modeling the guidance from the prediction of intents to the decoding of slot filling.However, existing methods (1) only model the unidirectional guidance from intent to slot; (2) adopt homogeneous graphs to model the interactions between the slot semantics nodes and intent label nodes, which limit the performance.In this paper, we propose a novel model termed Co-guiding Net, which implements a two-stage framework achieving the mutual guidances between the two tasks.In the first stage, the initial estimated labels of both tasks are produced, and then they are leveraged in the second stage to model the mutual guidances.Specifically, we propose two heterogeneous graph attention networks working on the proposed two heterogeneous semantics-label graphs, which effectively represent the relations among the semantics nodes and label nodes.Experiment results show that our model outperforms existing models by a large margin, obtaining a relative improvement of 19.3% over the previous best model on MixATIS dataset in overall accuracy.

Derivative-free prompt learning has emerged as a lightweight alternative to prompt tuning, which only requires model inference to optimize the prompts. However, existing work did not take full advantage of the over-parameterized characteristics of large pre-trained language models (PLMs). In this paper, we propose Clip-Tuning, a simple yet effective method that adopts diverse frozen “thinned” networks of PLMs to obtain *a mixture of rewards* and thus advance the derivative-free prompt learning. The thinned networks consist of all the hidden units that survive a stationary dropout strategy, whose inference predictions reflect an ensemble of partial views over prompted training samples. Our method outperforms previous gradient-free prompt learning methods and achieves parity with gradient-based counterparts on seven language understanding benchmarks under few-shot settings.