COLING 2024 论文列表

Aspect-based sentiment analysis (ABSA) delves into understanding sentiments specific to distinct elements within a user-generated review. It aims to analyze user-generated reviews to determine a) the target entity being reviewed, b) the high-level aspect to which it belongs, c) the sentiment words used to express the opinion, and d) the sentiment expressed toward the targets and the aspects. While various benchmark datasets have fostered advancements in ABSA, they often come with domain limitations and data granularity challenges. Addressing these, we introduce the OATS dataset, which encompasses three fresh domains and consists of 27,470 sentence-level quadruples and 17,092 review-level tuples. Our initiative seeks to bridge specific observed gaps in existing datasets: the recurrent focus on familiar domains like restaurants and laptops, limited data for intricate quadruple extraction tasks, and an occasional oversight of the synergy between sentence and review-level sentiments. Moreover, to elucidate OATS’s potential and shed light on various ABSA subtasks that OATS can solve, we conducted experiments, establishing initial baselines. We hope the OATS dataset augments current resources, paving the way for an encompassing exploration of ABSA (https://github.com/RiTUAL-UH/OATS-ABSA).

Conceptual structure is fundamental to human cognition and natural language understanding. It is significant to explore whether Large Language Models (LLMs) understand such knowledge. Since FrameNet serves as a well-defined conceptual structure knowledge resource, with meaningful frames, fine-grained frame elements, and rich frame relations, we construct a benchmark for coNceptual structure induction based on FrameNet, called NutFrame. It contains three sub-tasks: Frame Induction, Frame Element Induction, and Frame Relation Induction. In addition, we utilize prompts to induce conceptual structure of Framenet with LLMs. Furthermore, we conduct extensive experiments on NutFrame to evaluate various widely-used LLMs. Experimental results demonstrate that FrameNet induction remains a challenge for LLMs.

Headline generation, a key task in abstractive summarization, strives to condense a full-length article into a succinct, single line of text. Notably, while contemporary encoder-decoder models excel based on the ROUGE metric, they often falter when it comes to the precise generation of numerals in headlines. We identify the lack of datasets providing fine-grained annotations for accurate numeral generation as a major roadblock. To address this, we introduce a new dataset, the NumHG, and provide over 27,000 annotated numeral-rich news articles for detailed investigation. Further, we evaluate five well-performing models from previous headline-generation tasks using human evaluation in terms of numerical accuracy, reasonableness, and readability. Our study reveals a need for improvement in numerical accuracy, demonstrating the potential of the NumHG dataset to drive progress in number-focused headline generation and stimulate further discussions in numeral-focused text generation.

In this study we approach the detection of null subjects and impersonal constructions in Spanish using a machine translation methodology. We repurpose the Spanish AnCora corpus, converting it to a parallel set that transforms Spanish sentences into a format that allows us to detect and classify verbs, and train LSTM-based neural machine translation systems to perform this task. Various models differing on output format and hyperparameters were evaluated. Experimental results proved this approach to be highly resource-effective, obtaining results comparable to or surpassing the state of the art found in existing literature, while employing modest computational resources. Additionally, an improved dataset for training and evaluating Spanish null-subject detection tools was elaborated for this project, that could aid in the creation and serve as a benchmark for further developments in the area.

Existing multi-hop question generation (QG) methods treat answer-irrelevant documents as non-essential and remove them as impurities. However, this approach can create a training-inference discrepancy when impurities cannot be completely removed, which can lead to a decrease in model performance. To overcome this problem, we propose an auxiliary task, called order-agnostic, which leverages non-essential data in the training phase to create a robust model and extract the consistent embeddings in real-world inference environments. Additionally, we use a single LM to perform both ranker and generator through a prompt-based approach without applying additional external modules. Furthermore, we discover that appropriate utilization of the non-essential components can achieve a significant performance increase. Finally, experiments conducted on HotpotQA dataset achieve state-of-the-art.

Knowledge Base Question Answering (KBQA) systems play a pivotal role in the domain of natural language processing and information retrieval. Its primary objective is to bridge the gap between natural language questions and structured knowledge representations, especially for complex KBQA. Despite the significant progress in developing effective and interconnected KBQA technologies, the recent emergence of large language models (LLMs) offers an opportunity to address the challenges faced by KBQA systems more efficiently. This study adopts the LLMs, such as Large Language Model Meta AI (LLaMA), as a channel to connect natural language questions with structured knowledge representations and proposes a Three-step Fine-tune Strategy based on large language model to implement the KBQA system (TFS-KBQA). This method achieves direct conversion from natural language questions to structured knowledge representations, thereby overcoming the limitations of existing KBQA methods, such as addressing large search and reasoning spaces and ranking massive candidates. To evaluate the effectiveness of the proposed method, we conduct experiments using three popular complex KBQA datasets. The results achieve state-of-the-art performance across all three datasets, with particularly notable results for the WebQuestionSP dataset, which achieves an F1 value of 79.9%.

Intent classification and slot-filling are essential tasks of Spoken Language Understanding (SLU). In most SLU systems, those tasks are realized by independent modules, but for about fifteen years, models achieving both of them jointly and exploiting their mutual enhancement have been proposed. A multilingual module using a joint model was envisioned to create a touristic dialogue system for a European project, HumanE-AI-Net. A combination of multiple datasets, including the MEDIA dataset, was suggested for training this joint model. The MEDIA SLU dataset is a French dataset distributed since 2005 by ELRA, mainly used by the French research community and free for academic research since 2020. Unfortunately, it is annotated only in slots but not intents. An enhanced version of MEDIA annotated with intents has been built to extend its use to more tasks and use cases. This paper presents the semi-automatic methodology used to obtain this enhanced version. In addition, we present the first results of SLU experiments on this enhanced dataset using joint models for intent classification and slot-filling.

In his Universal 14, Greenberg stated that the normal and dominant order in all world languages was to place the condition before the conclusion in conditional sentences. We take this claim to review it quantitatively and based on occurrences in real texts in more than 50 languages. We can see that Greenberg’s proposal is correct but that it needs a reformulation to be true at all. We propose a quantitatively based and updated Universal 14, which gives a better account of the representation of the different languages analyzed and which is fulfilled in 100% of the cases (as opposed to Greenberg’s 60% in our sample). In addition, we also analyze adverbial sentences. Once we obtain the occurrence data in their direction (before or after the main verb), we plot a new Universal in a typometrical way: 100% of the languages show a higher proportion of preceding conditional clauses than of adverbial clauses, regardless of their type or the direction preference for adverbial clauses. The relationship between the SOV type and a stricter initial conditional location is also proposed.

This paper presents a new phonetic resource for Nigerian Pidgin, a low-resource language of West Africa. Aiming to provide a new tool for research on intonosyntax, we have augmented an existing syntactic treebank of Nigerian Pidgin, associating each orthographically transcribed token with a series of syllable-level alignments and phonetizations. Syllables are further described using a set of continuous and discrete prosodic features. This new approach provides a simple tool for researchers to explore the prosodic characteristics of various syntactic phenomena. In this paper, we present the format of the corpus, the various features added, and several explorations that can be performed using an online interface. We also present a prosodically specified lexicon extracted using this resource. In it, each orthographic form is accompanied by the frequency of its phoneme-level variants, as well as the suprasegmental features that most frequently accompany each syllable. Finally, we present several additional case studies on how this corpus can used in the study of the language’s prosody.

New Intent Discovery (NID) aims to recognize known and infer new intent categories with the help of limited labeled and large-scale unlabeled data. The task is addressed as a feature-clustering problem and recent studies augment instance representation. However, existing methods fail to capture cluster-friendly representations, since they show less capability to effectively control and coordinate within-cluster and between-cluster distances. Tailored to the NID problem, we propose a Robust and Adaptive Prototypical learning (RAP) framework for globally distinct decision boundaries for both known and new intent categories. Specifically, a robust prototypical attracting learning (RPAL) method is designed to compel instances to gravitate toward their corresponding prototype, achieving greater within-cluster compactness. To attain larger between-cluster separation, another adaptive prototypical dispersing learning (APDL) method is devised to maximize the between-cluster distance from the prototype-to-prototype perspective. Experimental results evaluated on three challenging benchmarks (CLINC, BANKING, and StackOverflow) of our method with better cluster-friendly representation demonstrate that RAP brings in substantial improvements over the current state-of-the-art methods (even large language model) by a large margin (average 5.5% improvement).

Text Classification is one of the most common tasks in Natural Language Processing. When proposing new classification models, practitioners select a sample of items the proposed model classified correctly while the baseline did not, and then try to observe patterns across those items to understand the proposed model’s strengths. However, this approach is not comprehensive and requires the effort of observing patterns across text items. In this work, we propose a new evaluation methodology for performing qualitative assessment over multiple classification models. The proposed methodology is driven to discover clusters of text items where each cluster’s items 1) exhibit a linguistic pattern and 2) the proposed model significantly outperforms the baseline when classifying such items. This helps practitioners in learning what their proposed model is powerful at capturing in comparison with the baseline model without having to perform this process manually. We use a fine-tuned BERT and Logistic Regression as the two models to compare with Sentiment Analysis as the downstream task. We show how our proposed evaluation methodology discovers various clusters of text items which BERT classifies significantly more accurately than the Logistic Regression baseline, thus providing insight into what BERT is powerful at capturing.

This paper introduces DACCORD, an original dataset in French for automatic detection of contradictions between sentences. It also presents new, manually translated versions of two datasets, namely the well known dataset RTE3 and the recent dataset GQNLI, from English to French, for the task of natural language inference / recognising textual entailment, which is a sentence-pair classification task. These datasets help increase the admittedly limited number of datasets in French available for these tasks. DACCORD consists of 1034 pairs of sentences and is the first dataset exclusively dedicated to this task and covering among others the topic of the Russian invasion in Ukraine. RTE3-FR contains 800 examples for each of its validation and test subsets, while GQNLI-FR is composed of 300 pairs of sentences and focuses specifically on the use of generalised quantifiers. Our experiments on these datasets show that they are more challenging than the two already existing datasets for the mainstream NLI task in French (XNLI, FraCaS). For languages other than English, most deep learning models for NLI tasks currently have only XNLI available as a training set. Additional datasets, such as ours for French, could permit different training and evaluation strategies, producing more robust results and reducing the inevitable biases present in any single dataset.

Neural topic models can successfully find coherent and diverse topics in textual data. However, they are limited in dealing with multimodal datasets (e.g., images and text). This paper presents the first systematic and comprehensive evaluation of multimodal topic modeling of documents containing both text and images. In the process, we propose two novel topic modeling solutions and two novel evaluation metrics. Overall, our evaluation on an unprecedented rich and diverse collection of datasets indicates that both of our models generate coherent and diverse topics. Nevertheless, the extent to which one method outperforms the other depends on the metrics and dataset combinations, which suggests further exploration of hybrid solutions in the future. Notably, our succinct human evaluation aligns with the outcomes determined by our proposed metrics. This alignment not only reinforces the credibility of our metrics but also highlights the potential for their application in guiding future multimodal topic modeling endeavors.

Training neural models for translating between low-resource languages is challenging due to the scarcity of direct parallel data between such languages. Pivot-based neural machine translation (NMT) systems overcome data scarcity by including a high-resource pivot language in the process of translating between low-resource languages. We propose synthetic pivoting, a novel approach to pivot-based translation in which the pivot sentences are generated synthetically from both the source and target languages. Synthetic pivot sentences are generated through sequence-level knowledge distillation, with the aim of changing the structure of pivot sentences to be closer to that of the source or target languages, thereby reducing pivot translation complexity. We incorporate synthetic pivoting into two paradigms for pivoting: cascading and direct translation using synthetic source and target sentences. We find that the performance of pivot-based systems highly depends on the quality of the NMT model used for sentence regeneration. Furthermore, training back-translation models on these sentences can make the models more robust to input-side noise. The results show that synthetic data generation improves pivot-based systems translating between low-resource Southern African languages by up to 5.6 BLEU points after fine-tuning.

For several NLP tasks, an important substep is the identification of noun phrases in running text. This has typically been done by “chunking” – a way of finding minimal noun phrases by token classification. However, chunking-like methods do not represent the fact that noun phrases can be nested. This paper presents a novel method of finding all noun phrases in a sentence, nested to an arbitrary depth, using the BERT model for token classification. We show that our proposed method achieves very good results for both Swedish and English.

Nested Event Extraction (NEE) aims to extract complex event structures where an event contains other events as its arguments recursively. Nested events involve a kind of Pivot Elements (PEs) that simultaneously act as arguments of outer-nest events and as triggers of inner-nest events, and thus connect them into nested structures. This special characteristic of PEs brings challenges to existing NEE methods, as they cannot well cope with the dual identities of PEs. Therefore, this paper proposes a new model, called PerNee, which extracts nested events mainly based on recognizing PEs. Specifically, PerNee first recognizes the triggers of both inner-nest and outer-nest events and further recognizes the PEs via classifying the relation type between trigger pairs. The model uses prompt learning to incorporate information from both event types and argument roles for better trigger and argument representations to improve NEE performance. Since existing NEE datasets (e.g., Genia11) are limited to specific domains and contain a narrow range of event types with nested structures, we systematically categorize nested events in the generic domain and construct a new NEE dataset, called ACE2005-Nest. Experimental results demonstrate that PerNee consistently achieves state-of-the-art performance on ACE2005-Nest, Genia11, and Genia13. The ACE2005-Nest dataset and the code of the PerNee model are available at https://github.com/waysonren/PerNee.

Previous works of negation understanding mainly focus on negation cue detection and scope resolution, without identifying negation subject which is also significant to the downstream tasks. In this paper, we propose a new negation triplet extraction (NTE) task which aims to extract negation subject along with negation cue and scope. To achieve NTE, we devise a novel Syntax&Semantic-Enhanced Negation Extraction model, namely SSENE, which is built based on a generative pretrained language model (PLM) of Encoder-Decoder architecture with a multi-task learning framework. Specifically, the given sentence’s syntactic dependency tree is incorporated into the PLM’s encoder to discover the correlations between the negation subject, cue and scope. Moreover, the semantic consistency between the sentence and the extracted triplet is ensured by an auxiliary task learning. Furthermore, we have constructed a high-quality Chinese dataset NegComment based on the users’ reviews from the real-world platform of Meituan, upon which our evaluations show that SSENE achieves the best NTE performance compared to the baselines. Our ablation and case studies also demonstrate that incorporating the syntactic information helps the PLM’s recognize the distant dependency between the subject and cue, and the auxiliary task learning is helpful to extract the negation triplets with more semantic consistency. We further demonstrate that SSENE is also competitive on the traditional CDSR task.

Negation scope resolution is the task that identifies the part of a sentence affected by the negation cue. The three major corpora used for this task, the BioScope corpus, the SFU review corpus and the Sherlock dataset, have different annotation schemes for negation scope. Due to the different annotations, the negation scope resolution models based on pre-trained language models (PLMs) perform worse when fine-tuned on the simply combined dataset consisting of the three corpora. To address this issue, we propose a method for automatically converting the scopes of BioScope and SFU to those of Sherlock and merge them into a unified dataset. To verify the effectiveness of the proposed method, we conducted experiments using the unified dataset for fine-tuning PLM-based models. The experimental results demonstrate that the performances of the models increase when fine-tuned on the unified dataset unlike the simply combined one. In the token-level metric, the model fine-tuned on the unified dataset archived the state-of-the-art performance on the Sherlock dataset.

Instruction-tuned Large Language Models (LLMs) have exhibited impressive language understanding and the capacity to generate responses that follow specific prompts. However, due to the computational demands associated with training these models, their applications often adopt a zero-shot setting. In this paper, we evaluate the zero-shot performance of two publicly accessible LLMs, ChatGPT and OpenAssistant, in the context of six Computational Social Science classification tasks, while also investigating the effects of various prompting strategies. Our experiments investigate the impact of prompt complexity, including the effect of incorporating label definitions into the prompt; use of synonyms for label names; and the influence of integrating past memories during foundation model training. The findings indicate that in a zero-shot setting, current LLMs are unable to match the performance of smaller, fine-tuned baseline transformer models (such as BERT-large). Additionally, we find that different prompting strategies can significantly affect classification accuracy, with variations in accuracy and F1 scores exceeding 10%.