COLING 2020 Papers

We release ManyNames v2 (MN v2), a verified version of an object naming dataset that contains dozens of valid names per object for 25K images. We analyze issues in the data collection method originally employed, standard in Language & Vision (L&V), and find that the main source of noise in the data comes from simulating a naming context solely from an image with a target object marked with a bounding box, which causes subjects to sometimes disagree regarding which object is the target. We also find that both the degree of this uncertainty in the original data and the amount of true naming variation in MN v2 differs substantially across object domains. We use MN v2 to analyze a popular L&V model and demonstrate its effectiveness on the task of object naming. However, our fine-grained analysis reveals that what appears to be human-like model behavior is not stable across domains, e.g., the model confuses people and clothing objects much more frequently than humans do. We also find that standard evaluations underestimate the actual effectiveness of the naming model: on the single-label names of the original dataset (Visual Genome), it obtains −27% accuracy points than on MN v2, that includes all valid object names.

Deep neural models tremendously improved machine translation. In this context, we investigate whether distinguishing machine from human translations is still feasible. We trained and applied 18 classifiers under two settings: a monolingual task, in which the classifier only looks at the translation; and a bilingual task, in which the source text is also taken into consideration. We report on extensive experiments involving 4 neural MT systems (Google Translate, DeepL, as well as two systems we trained) and varying the domain of texts. We show that the bilingual task is the easiest one and that transfer-based deep-learning classifiers perform best, with mean accuracies around 85% in-domain and 75% out-of-domain .

Despite the recent advances in coherence modelling, most such models including state-of-the-art neural ones, are evaluated on either contrived proxy tasks such as the standard order discrimination benchmark, or tasks that require special expert annotation. Moreover, most evaluations are conducted on small newswire corpora. To address these shortcomings, in this paper we propose four generic evaluation tasks that draw on different aspects of coherence at both the lexical and document levels, and can be applied to any corpora. In designing these tasks, we aim at capturing coherence-specific properties, such as the correct use of discourse connectives, lexical cohesion, as well as the overall temporal and causal consistency among events and participants in a story. Importantly, our proposed tasks either rely on automatically-generated data, or data annotated for other purposes, hence alleviating the need for annotation specifically targeted to the task of coherence modelling. We perform experiments with several existing state-of-the-art neural models of coherence on these tasks, across large corpora from different domains, including newswire, dialogue, as well as narrative and instructional text. Our findings point to a strong need for revisiting the common practices in the development and evaluation of coherence models.

Selectional preference is defined as the tendency of a predicate to favor particular arguments within a certain linguistic context, and likewise, reject others that result in conflicting or implausible meanings. The stellar success of contextual word embedding models such as BERT in NLP tasks has led many to question whether these models have learned linguistic information, but up till now, most research has focused on syntactic information. We investigate whether Bert contains information on the selectional preferences of words, by examining the probability it assigns to the dependent word given the presence of a head word in a sentence. We are using word pairs of head-dependent words in five different syntactic relations from the SP-10K corpus of selectional preference (Zhang et al., 2019b), in sentences from the ukWaC corpus, and we are calculating the correlation of the plausibility score (from SP-10K) and the model probabilities. Our results show that overall, there is no strong positive or negative correlation in any syntactic relation, but we do find that certain head words have a strong correlation and that masking all words but the head word yields the most positive correlations in most scenarios –which indicates that the semantics of the predicate is indeed an integral and influential factor for the selection of the argument.

The prevalent approach for unsupervised text style transfer is disentanglement between content and style. However, it is difficult to completely separate style information from the content. Other approaches allow the latent text representation to contain style and the target style to affect the generated output more than the latent representation does. In both approaches, however, it is impossible to adjust the strength of the style in the generated output. Moreover, those previous approaches typically perform both the sentence reconstruction and style control tasks in a single model, which complicates the overall architecture. In this paper, we address these issues by separating the model into a sentence reconstruction module and a style module. We use the Transformer-based autoencoder model for sentence reconstruction and the adaptive style embedding is learned directly in the style module. Because of this separation, each module can better focus on its own task. Moreover, we can vary the style strength of the generated sentence by changing the style of the embedding expression. Therefore, our approach not only controls the strength of the style, but also simplifies the model architecture. Experimental results show that our approach achieves better style transfer performance and content preservation than previous approaches.

Long Short-Term Memory recurrent neural network (LSTM) is widely used and known to capture informative long-term syntactic dependencies. However, how such information are reflected in its internal vectors for natural text has not yet been sufficiently investigated. We analyze them by learning a language model where syntactic structures are implicitly given. We empirically show that the context update vectors, i.e. outputs of internal gates, are approximately quantized to binary or ternary values to help the language model to count the depth of nesting accurately, as Suzgun et al. (2019) recently show for synthetic Dyck languages. For some dimensions in the context vector, we show that their activations are highly correlated with the depth of phrase structures, such as VP and NP. Moreover, with an L1 regularization, we also found that it can accurately predict whether a word is inside a phrase structure or not from a small number of components of the context vector. Even for the case of learning from raw text, context vectors are shown to still correlate well with the phrase structures. Finally, we show that natural clusters of the functional words and the part of speeches that trigger phrases are represented in a small but principal subspace of the context-update vector of LSTM.

Recent research on the multi-head attention mechanism, especially that in pre-trained models such as BERT, has shown us heuristics and clues in analyzing various aspects of the mechanism. As most of the research focus on probing tasks or hidden states, previous works have found some primitive patterns of attention head behavior by heuristic analytical methods, but a more systematic analysis specific on the attention patterns still remains primitive. In this work, we clearly cluster the attention heatmaps into significantly different patterns through unsupervised clustering on top of a set of proposed features, which corroborates with previous observations. We further study their corresponding functions through analytical study. In addition, our proposed features can be used to explain and calibrate different attention heads in Transformer models.

Meetings are essential to modern organizations. Numerous meetings are held and recorded daily, more than can ever be comprehended. A meeting summarization system that identifies salient utterances from the transcripts to automatically generate meeting minutes can help. It empowers users to rapidly search and sift through large meeting collections. To date, the impact of domain terminology on the performance of meeting summarization remains understudied, despite that meetings are rich with domain knowledge. In this paper, we create gold-standard annotations for domain terminology on a sizable meeting corpus; they are known as jargon terms. We then analyze the performance of a meeting summarization system with and without jargon terms. Our findings reveal that domain terminology can have a substantial impact on summarization performance. We publicly release all domain terminology to advance research in meeting summarization.

This work presents a method of word sense clustering that differentiates homonyms and merge homophones, taking Japanese as an example, where orthographical variation causes problem for language processing. It uses contextualised embeddings (BERT) to cluster tokens into distinct sense groups, and we use these groups to normalise synonymous instances to a single representative form. We see the benefit of this normalisation in language model, as well as in transliteration.

Recently, automatic trivia fact extraction has attracted much research interest. Modern search engines have begun to provide trivia facts as the information for entities because they can motivate more user engagement. In this paper, we propose a new unsupervised algorithm that automatically mines trivia facts for a given entity. Unlike previous studies, the proposed algorithm targets at a single Wikipedia article and leverages its hierarchical structure via top-down processing. Thus, the proposed algorithm offers two distinctive advantages: it does not incur high computation time, and it provides a domain-independent approach for extracting trivia facts. Experimental results demonstrate that the proposed algorithm is over 100 times faster than the existing method which considers Wikipedia categories. Human evaluation demonstrates that the proposed algorithm can mine better trivia facts regardless of the target entity domain and outperforms the existing methods.

In this paper, we address the segmentation of books of hours, Latin devotional manuscripts of the late Middle Ages, that exhibit challenging issues: a complex hierarchical entangled structure, variable content, noisy transcriptions with no sentence markers, and strong correlations between sections for which topical information is no longer sufficient to draw segmentation boundaries. We show that the main state-of-the-art segmentation methods are either inefficient or inapplicable for books of hours and propose a bottom-up greedy approach that considerably enhances the segmentation results. We stress the importance of such hierarchical segmentation of books of hours for historians to explore their overarching differences underlying conception about Church.

Event extraction plays an important role in legal applications, including case push and auxiliary judgment. However, traditional event structure cannot express the connections between arguments, which are extremely important in legal events. Therefore, this paper defines a dynamic event structure for Chinese legal events. To distinguish between similar events, we design hierarchical event features for event detection. Moreover, to address the problem of long-distance semantic dependence and anaphora resolution in argument classification, we propose a novel pedal attention mechanism to extract the semantic relation between two words through their dependent adjacent words. We label a Chinese legal event dataset and evaluate our model on it. Experimental results demonstrate that our model can surpass other state-of-the-art models.

Review rating prediction of text reviews is a rapidly growing technology with a wide range of applications in natural language processing. However, most existing methods either use hand-crafted features or learn features using deep learning with simple text corpus as input for review rating prediction, ignoring the hierarchies among data. In this paper, we propose a Hierarchical bi-directional self-attention Network framework (HabNet) for paper review rating prediction and recommendation, which can serve as an effective decision-making tool for the academic paper review process. Specifically, we leverage the hierarchical structure of the paper reviews with three levels of encoders: sentence encoder (level one), intra-review encoder (level two) and inter-review encoder (level three). Each encoder first derives contextual representation of each level, then generates a higher-level representation, and after the learning process, we are able to identify useful predictors to make the final acceptance decision, as well as to help discover the inconsistency between numerical review ratings and text sentiment conveyed by reviewers. Furthermore, we introduce two new metrics to evaluate models in data imbalance situations. Extensive experiments on a publicly available dataset (PeerRead) and our own collected dataset (OpenReview) demonstrate the superiority of the proposed approach compared with state-of-the-art methods.

Emotion detection in conversations (EDC) is to detect the emotion for each utterance in conversations that have multiple speakers. Different from the traditional non-conversational emotion detection, the model for EDC should be context-sensitive (e.g., understanding the whole conversation rather than one utterance) and speaker-sensitive (e.g., understanding which utterance belongs to which speaker). In this paper, we propose a transformer-based context- and speaker-sensitive model for EDC, namely HiTrans, which consists of two hierarchical transformers. We utilize BERT as the low-level transformer to generate local utterance representations, and feed them into another high-level transformer so that utterance representations could be sensitive to the global context of the conversation. Moreover, we exploit an auxiliary task to make our model speaker-sensitive, called pairwise utterance speaker verification (PUSV), which aims to classify whether two utterances belong to the same speaker. We evaluate our model on three benchmark datasets, namely EmoryNLP, MELD and IEMOCAP. Results show that our model outperforms previous state-of-the-art models.

Most recent works in the field of grammatical error correction (GEC) rely on neural machine translation-based models. Although these models boast impressive performance, they require a massive amount of data to properly train. Furthermore, NMT-based systems treat GEC purely as a translation task and overlook the editing aspect of it. In this work we propose a heterogeneous approach to Chinese GEC, composed of a NMT-based model, a sequence editing model, and a spell checker. Our methodology not only achieves a new state-of-the-art performance for Chinese GEC, but also does so without relying on data augmentation or GEC-specific architecture changes. We further experiment with all possible configurations of our system with respect to model composition order and number of rounds of correction. A detailed analysis of each model and their contributions to the correction process is performed by adapting the ERRANT scorer to be able to score Chinese sentences.

Given an incomplete event chain, script learning aims to predict the missing event, which can support a series of NLP applications. Existing work cannot well represent the heterogeneous relations and capture the discontinuous event segments that are common in the event chain. To address these issues, we introduce a heterogeneous-event (HeterEvent) graph network. In particular, we employ each unique word and individual event as nodes in the graph, and explore three kinds of edges based on realistic relations (e.g., the relations of word-and-word, word-and-event, event-and-event). We also design a message passing process to realize information interactions among homo or heterogeneous nodes. And the discontinuous event segments could be explicitly modeled by finding the specific path between corresponding nodes in the graph. The experimental results on one-step and multi-step inference tasks demonstrate that our ensemble model HeterEvent[W+E] can outperform existing baselines.

End-to-end neural data-to-text (D2T) generation has recently emerged as an alternative to pipeline-based architectures. However, it has faced challenges generalizing to new domains and generating semantically consistent text. In this work, we present DataTuner, a neural, end-to-end data-to-text generation system that makes minimal assumptions about the data representation and target domain. We take a two-stage generation-reranking approach, combining a fine-tuned language model with a semantic fidelity classifier. Each component is learnt end-toe-nd without needing dataset-specific heuristics, entity delexicalization, or post-processing. We show that DataTuner achieves state of the art results on automated metrics across four major D2T datasets (LDC2017T10, WebNLG, ViGGO, and Cleaned E2E), with fluency assessed by human annotators as nearing or exceeding the human-written reference texts. Our generated text has better semantic fidelity than the state of the art on these datasets. We further demonstrate that our model-based semantic fidelity scorer is a better assessment tool compared to traditional heuristic-based measures of semantic accuracy.

Academia and industry have developed machine learning and natural language processing models to detect online hate speech automatically. However, most of these existing methods adopt a supervised approach that heavily depends on labeled datasets for training. This results in the methods’ poor detection performance of the hate speech class as the training datasets are highly imbalanced. In this paper, we propose HateGAN, a deep generative reinforcement learning model, which addresses the challenge of imbalance class by augmenting the dataset with hateful tweets. We conduct extensive experiments to augment two commonly-used hate speech detection datasets with the HateGAN generated tweets. Our experiment results show that HateGAN improves the detection performance of the hate speech class regardless of the classifiers and datasets used in the detection task. Specifically, we observe an average 5% improvement for the hate class F1 scores across all state-of-the-art hate speech classifiers. We also conduct case studies to empirically examine the HateGAN generated hate speeches and show that the generated tweets are diverse, coherent, and relevant to hate speech detection.

Cognates are variants of the same lexical form across different languages; for example “fonema” in Spanish and “phoneme” in English are cognates, both of which mean “a unit of sound”. The task of automatic detection of cognates among any two languages can help downstream NLP tasks such as Cross-lingual Information Retrieval, Computational Phylogenetics, and Machine Translation. In this paper, we demonstrate the use of cross-lingual word embeddings for detecting cognates among fourteen Indian Languages. Our approach introduces the use of context from a knowledge graph to generate improved feature representations for cognate detection. We, then, evaluate the impact of our cognate detection mechanism on neural machine translation (NMT), as a downstream task. We evaluate our methods to detect cognates on a challenging dataset of twelve Indian languages, namely, Sanskrit, Hindi, Assamese, Oriya, Kannada, Gujarati, Tamil, Telugu, Punjabi, Bengali, Marathi, and Malayalam. Additionally, we create evaluation datasets for two more Indian languages, Konkani and Nepali. We observe an improvement of up to 18% points, in terms of F-score, for cognate detection. Furthermore, we observe that cognates extracted using our method help improve NMT quality by up to 2.76 BLEU. We also release our code, newly constructed datasets and cross-lingual models publicly.