COLING 2020 Papers

We present an open-source online dictionary editing system, Ve′rdd, that offers a chance to re-evaluate and edit grassroots dictionaries that have been exposed to multiple amateur editors. The idea is to incorporate community activities into a state-of-the-art finite-state language description of a seriously endangered minority language, Skolt Sami. Problems involve getting the community to take part in things above the pencil-and-paper level. At times, it seems that the native speakers and the dictionary oriented are lacking technical understanding to utilize the infrastructures which might make their work more meaningful in the future, i.e. multiple reuse of all of their input. Therefore, our system integrates with the existing tools and infrastructures for Uralic language masking the technical complexities behind a user-friendly UI.

Automatic identification of multiword expressions (MWEs), like ‘to cut corners’ (to do an incomplete job), is a pre-requisite for semantically-oriented downstream applications. This task is challenging because MWEs, especially verbal ones (VMWEs), exhibit surface variability. This paper deals with a subproblem of VMWE identification: the identification of occurrences of previously seen VMWEs. A simple language-independent system based on a combination of filters competes with the best systems from a recent shared task: it obtains the best averaged F-score over 11 languages (0.6653) and even the best score for both seen and unseen VMWEs due to the high proportion of seen VMWEs in texts. This highlights the fact that focusing on the identification of seen VMWEs could be a strategy to improve VMWE identification in general.

We introspect black-box sentence embeddings by conditionally generating from them with the objective to retrieve the underlying discrete sentence. We perceive of this as a new unsupervised probing task and show that it correlates well with downstream task performance. We also illustrate how the language generated from different encoders differs. We apply our approach to generate sentence analogies from sentence embeddings.

Traditional computational authorship attribution describes a classification task in a closed-set scenario. Given a finite set of candidate authors and corresponding labeled texts, the objective is to determine which of the authors has written another set of anonymous or disputed texts. In this work, we propose a probabilistic autoencoding framework to deal with this supervised classification task. Variational autoencoders (VAEs) have had tremendous success in learning latent representations. However, existing VAEs are currently still bound by limitations imposed by the assumed Gaussianity of the underlying probability distributions in the latent space. In this work, we are extending a VAE with an embedded Gaussian mixture model to a Student-t mixture model, which allows for an independent control of the “heaviness” of the respective tails of the implied probability densities. Experiments over an Amazon review dataset indicate superior performance of the proposed method.

In response to (i) inconclusive results in the literature as to the properties of coreference chains in written versus spoken language, and (ii) a general lack of work on automatic coreference resolution on both spoken language and social media, we undertake a corpus study involving the various genre sections of Ontonotes, the Switchboard corpus, and a corpus of Twitter conversations. Using a set of measures that previously have been applied individually to different data sets, we find fairly clear patterns of “behavior” for the different genres/media. Besides their role for psycholinguistic investigation (why do we employ different coreference strategies when we write or speak) and for the placement of Twitter in the spoken–written continuum, we see our results as a contribution to approaching genre-/media-specific coreference resolution.

Text-to-image multimodal tasks, generating/retrieving an image from a given text description, are extremely challenging tasks since raw text descriptions cover quite limited information in order to fully describe visually realistic images. We propose a new visual contextual text representation for text-to-image multimodal tasks, VICTR, which captures rich visual semantic information of objects from the text input. First, we use the text description as initial input and conduct dependency parsing to extract the syntactic structure and analyse the semantic aspect, including object quantities, to extract the scene graph. Then, we train the extracted objects, attributes, and relations in the scene graph and the corresponding geometric relation information using Graph Convolutional Networks, and it generates text representation which integrates textual and visual semantic information. The text representation is aggregated with word-level and sentence-level embedding to generate both visual contextual word and sentence representation. For the evaluation, we attached VICTR to the state-of-the-art models in text-to-image generation.VICTR is easily added to existing models and improves across both quantitative and qualitative aspects.

In this paper we present a character-level sequence-to-sequence lemmatization model, utilizing several subword features in multiple configurations. In addition to generic n-gram embeddings (using FastText), we experiment with concatenative (stems) and templatic (roots and patterns) morphological subwords. We present several architectures that embed these features directly at the encoder side, or learn them jointly at the decoder side with a multitask learning architecture. The results indicate that using the generic n-gram embeddings (through FastText) outperform the other linguistically-driven subwords. We use Modern Standard Arabic and Egyptian Arabic as test cases, with up to 22% and 13% relative error reduction, respectively, from a strong baseline. An error analysis shows that our best system is even able to handle word/lemma pairs that are both unseen in the training data.

We tackle implicit discourse relation classification, a task of automatically determining semantic relationships between arguments. The attention-worthy words in arguments are crucial clues for classifying the discourse relations. Attention mechanisms have been proven effective in highlighting the attention-worthy words during encoding. However, our survey shows that some inessential words are unintentionally misjudged as the attention-worthy words and, therefore, assigned heavier attention weights than should be. We propose a penalty-based loss re-estimation method to regulate the attention learning process, integrating penalty coefficients into the computation of loss by means of overstability of attention weight distributions. We conduct experiments on the Penn Discourse TreeBank (PDTB) corpus. The test results show that our loss re-estimation method leads to substantial improvements for a variety of attention mechanisms, and it obtains highly competitive performance compared to the state-of-the-art methods.

Sentence complexity assessment is a relatively new task in Natural Language Processing. One of its aims is to highlight in a text which sentences are more complex to support the simplification of contents for a target audience (e.g., children, cognitively impaired users, non-native speakers and low-literacy readers (Scarton and Specia, 2018)). This task is evaluated using datasets of pairs of aligned sentences including the complex and simple version of the same sentence. For Brazilian Portuguese, the task was addressed by (Leal et al., 2018), who set up the first dataset to evaluate the task in this language, reaching 87.8% of accuracy with linguistic features. The present work advances these results, using models inspired by (Gonzalez-Garduño and Søgaard, 2018), which hold the state-of-the-art for the English language, with multi-task learning and eye-tracking measures. First-Pass Duration, Total Regression Duration and Total Fixation Duration were used in two moments; first to select a subset of linguistic features and then as an auxiliary task in the multi-task and sequential learning models. The best model proposed here reaches the new state-of-the-art for Portuguese with 97.5% accuracy 1 , an increase of almost 10 points compared to the best previous results, in addition to proposing improvements in the public dataset after analysing the errors of our best model.

While bilingual corpora have been instrumental for machine translation, their utility for training translators has been less explored. We investigate the use of bilingual corpora as pedagogical tools for translation in the technical domain. In a user study, novice translators revised Chinese translations of English patents through bilingual concordancing. Results show that concordancing with an in-domain bilingual corpus can yield greater improvement in translation quality of technical terms than a general-domain bilingual corpus.

We study an end-to-end approach for conversational recommendation that dynamically manages and reasons over users’ past (offline) preferences and current (online) requests through a structured and cumulative user memory knowledge graph. This formulation extends existing state tracking beyond the boundary of a single dialog to user state tracking (UST). For this study, we create a new Memory Graph (MG) <-> Conversational Recommendation parallel corpus called MGConvRex with 7K+ human-to-human role-playing dialogs, grounded on a large-scale user memory bootstrapped from real-world user scenarios. MGConvRex captures human-level reasoning over user memory and has disjoint training/testing sets of users for zero-shot (cold-start) reasoning for recommendation. We propose a simple yet expandable formulation for constructing and updating the MG, and an end-to-end graph-based reasoning model that updates MG from unstructured utterances and predicts optimal dialog policies (eg recommendation) based on updated MG. The prediction of our proposed model inherits the graph structure, providing a natural way to explain policies. Experiments are conducted for both offline metrics and online simulation, showing competitive results.

Fine-tuning with pre-trained language models (e.g. BERT) has achieved great success in many language understanding tasks in supervised settings (e.g. text classification). However, relatively little work has been focused on applying pre-trained models in unsupervised settings, such as text clustering. In this paper, we propose a novel method to fine-tune pre-trained models unsupervisedly for text clustering, which simultaneously learns text representations and cluster assignments using a clustering oriented loss. Experiments on three text clustering datasets (namely TREC-6, Yelp, and DBpedia) show that our model outperforms the baseline methods and achieves state-of-the-art results.

Misinformation spreads across media, community, and knowledge graphs in the Web by not only human agents but also information extraction algorithms that extract factual statements from unstructured textual data to populate the existing knowledge graphs. Traditional fact checking by experts or crowds is increasingly difficult to keep pace with the volume of newly created misinformation in the Web. Therefore, it is important and necessary to enhance the computational ability to determine whether a given factual statement is truthful or not. We view this problem as a truth scoring task in a knowledge graph. We present a novel rule-based approach that finds positive and negative evidential paths in a knowledge graph for a given factual statement and calculates a truth score for the given statement by unsupervised ensemble of the found positive and negative evidential paths. For example, we can determine the factual statement “United States is the birth place of Barack Obama” as truthful if there is the positive evidential path (Barack Obama, birthPlace, Hawaii) ∧ (Hawaii, country, United States) in a knowledge graph. For another example, we can determine the factual statement “Canada is the nationality of Barack Obama” as untruthful if there is the negative evidential path (Barack Obama, nationality, United States) ∧ (United States, ≠, Canada) in a knowledge graph. For evaluating on a real-world situation, we constructed an evaluation dataset by labeling truth or untruth label on factual statements that were extracted from Wikipedia texts by using the state-of-the-art BERT-based information extraction system. Our evaluation results show that our approach outperforms the state-of-the-art unsupervised approaches significantly by up to 0.12 AUC-ROC and even outperforms the supervised approach by up to 0.05 AUC-ROC not only in our dataset but also in the two different standard datasets.

Automatic Language Identification (LI) or Dialect Identification (DI) of short texts of closely related languages or dialects, is one of the primary steps in many natural language processing pipelines. Language identification is considered a solved task in many cases; however, in the case of very closely related languages, or in an unsupervised scenario (where the languages are not known in advance), performance is still poor. In this paper, we propose the Unsupervised Deep Language and Dialect Identification (UDLDI) method, which can simultaneously learn sentence embeddings and cluster assignments from short texts. The UDLDI model understands the sentence constructions of languages by applying attention to character relations which helps to optimize the clustering of languages. We have performed our experiments on three short-text datasets for different language families, each consisting of closely related languages or dialects, with very minimal training sets. Our experimental evaluations on these datasets have shown significant improvement over state-of-the-art unsupervised methods and our model has outperformed state-of-the-art LI and DI systems in supervised settings.

RST-based discourse parsing is an important NLP task with numerous downstream applications, such as summarization, machine translation and opinion mining. In this paper, we demonstrate a simple, yet highly accurate discourse parser, incorporating recent contextual language models. Our parser establishes the new state-of-the-art (SOTA) performance for predicting structure and nuclearity on two key RST datasets, RST-DT and Instr-DT. We further demonstrate that pretraining our parser on the recently available large-scale “silver-standard” discourse treebank MEGA-DT provides even larger performance benefits, suggesting a novel and promising research direction in the field of discourse analysis.

Soft contextualized data augmentation is a recent method that replaces one-hot representation of words with soft posterior distributions of an external language model, smoothing the input of neural machine translation systems. Label smoothing is another effective method that penalizes over-confident model outputs by discounting some probability mass from the true target word, smoothing the output of neural machine translation systems. Having the benefit of updating all word vectors in each optimization step and better regularizing the models, the two smoothing methods are shown to bring significant improvements in translation performance. In this work, we study how to best combine the methods and stack the improvements. Specifically, we vary the prior distributions to smooth with, the hyperparameters that control the smoothing strength, and the token selection procedures. We conduct extensive experiments on small datasets, evaluate the recipes on larger datasets, and examine the implications when back-translation is further used. Our results confirm cumulative improvements when input and output smoothing are used in combination, giving up to +1.9 BLEU scores on standard machine translation tasks and reveal reasons why these smoothing methods should be preferred.

We present a new approach for detecting human-like social biases in word embeddings using representational similarity analysis. Specifically, we probe contextualized and non-contextualized embeddings for evidence of intersectional biases against Black women. We show that these embeddings represent Black women as simultaneously less feminine than White women, and less Black than Black men. This finding aligns with intersectionality theory, which argues that multiple identity categories (such as race or sex) layer on top of each other in order to create unique modes of discrimination that are not shared by any individual category.

This paper studies the effects of word-level linguistic annotations in under-resourced neural machine translation, for which there is incomplete evidence in the literature. The study covers eight language pairs, different training corpus sizes, two architectures, and three types of annotation: dummy tags (with no linguistic information at all), part-of-speech tags, and morpho-syntactic description tags, which consist of part of speech and morphological features. These linguistic annotations are interleaved in the input or output streams as a single tag placed before each word. In order to measure the performance under each scenario, we use automatic evaluation metrics and perform automatic error classification. Our experiments show that, in general, source-language annotations are helpful and morpho-syntactic descriptions outperform part of speech for some language pairs. On the contrary, when words are annotated in the target language, part-of-speech tags systematically outperform morpho-syntactic description tags in terms of automatic evaluation metrics, even though the use of morpho-syntactic description tags improves the grammaticality of the output. We provide a detailed analysis of the reasons behind this result.

Complex question answering (CQA) over raw text is a challenging task. A prominent approach to this task is based on the programmer-interpreter framework, where the programmer maps the question into a sequence of reasoning actions and the interpreter then executes these actions on the raw text. Learning an effective CQA model requires large amounts of human-annotated data, consisting of the ground-truth sequence of reasoning actions, which is time-consuming and expensive to collect at scale. In this paper, we address the challenge of learning a high-quality programmer (parser) by projecting natural human-generated questions into unnatural machine-generated questions which are more convenient to parse. We firstly generate synthetic (question, action sequence) pairs by a data generator, and train a semantic parser that associates synthetic questions with their corresponding action sequences. To capture the diversity when applied to natural questions, we learn a projection model to map natural questions into their most similar unnatural questions for which the parser can work well. Without any natural training data, our projection model provides high-quality action sequences for the CQA task. Experimental results show that the QA model trained exclusively with synthetic data outperforms its state-of-the-art counterpart trained on human-labeled data.