Word embeddings are widely used on a variety of tasks and can substantially improve the performance. However, their quality is not consistent throughout the vocabulary due to the long-tail distribution of word frequency. Without sufficient contexts, rare word embeddings are usually less reliable than those of common words. However, current models typically trust all word embeddings equally regardless of their reliability and thus may introduce noise and hurt the performance. Since names often contain rare and uncommon words, this problem is particularly critical for name tagging. In this paper, we propose a novel reliability-aware name tagging model to tackle this issue. We design a set of word frequency-based reliability signals to indicate the quality of each word embedding. Guided by the reliability signals, the model is able to dynamically select and compose features such as word embedding and character-level representation using gating mechanisms. For example, if an input word is rare, the model relies less on its word embedding and assigns higher weights to its character and contextual features. Experiments on OntoNotes 5.0 show that our model outperforms the baseline model by 2.7% absolute gain in F-score. In cross-genre experiments on five genres in OntoNotes, our model improves the performance for most genre pairs and obtains up to 5% absolute F-score gain.

Word embedding algorithms produce very reliable feature representations of words that are used by neural network models across a constantly growing multitude of NLP tasks. As such, it is imperative for NLP practitioners to understand how their word representations are produced, and why they are so impactful. The present work presents the Simple Embedder framework, generalizing the state-of-the-art existing word embedding algorithms (including Word2vec (SGNS) and GloVe) under the umbrella of generalized low rank models. We derive that both of these algorithms attempt to produce embedding inner products that approximate pointwise mutual information (PMI) statistics in the corpus. Once cast as Simple Embedders, comparison of these models reveals that these successful embedders all resemble a straightforward maximum likelihood estimate (MLE) of the PMI parametrized by the inner product (between embeddings). This MLE induces our proposed novel word embedding model, Hilbert-MLE, as the canonical representative of the Simple Embedder framework. We empirically compare these algorithms with evaluations on 17 different datasets. Hilbert-MLE consistently observes second-best performance on every extrinsic evaluation (news classification, sentiment analysis, POS-tagging, and supersense tagging), while the first-best model depends varying on the task. Moreover, Hilbert-MLE consistently observes the least variance in results with respect to the random initialization of the weights in bidirectional LSTMs. Our empirical results demonstrate that Hilbert-MLE is a very consistent word embedding algorithm that can be reliably integrated into existing NLP systems to obtain high-quality results.

Word Sense Disambiguation models exist in many flavors. Even though supervised ones tend to perform best in terms of accuracy, they often lose ground to more flexible knowledge-based solutions, which do not require training by a word expert for every disambiguation target. To bridge this gap we adopt a different perspective and rely on sequence learning to frame the disambiguation problem: we propose and study in depth a series of end-to-end neural architectures directly tailored to the task, from bidirectional Long Short-Term Memory to encoder-decoder models. Our extensive evaluation over standard benchmarks and in multiple languages shows that sequence learning enables more versatile all-words models that consistently lead to state-of-the-art results, even against word experts with engineered features.

With the requirements of natural language applications, multi-task sequence labeling methods have some immediate benefits over the single-task sequence labeling methods. Recently, many state-of-the-art multi-task sequence labeling methods were proposed, while still many issues to be resolved including (C1) exploring a more general relationship between tasks, (C2) extracting the task-shared knowledge purely and (C3) merging the task-shared knowledge for each task appropriately. To address the above challenges, we propose MTAA, a symmetric multi-task sequence labeling model, which performs an arbitrary number of tasks simultaneously. Furthermore, MTAA extracts the shared knowledge among tasks by adversarial learning and integrates the proposed multi-representation fusion attention mechanism for merging feature representations. We evaluate MTAA on two widely used data sets: CoNLL2003 and OntoNotes5.0. Experimental results show that our proposed model outperforms the latest methods on the named entity recognition and the syntactic chunking task by a large margin, and achieves state-of-the-art results on the part-of-speech tagging task.

With the recent rise of toxicity in online conversations on social media platforms, using modern machine learning algorithms for toxic comment detection has become a central focus of many online applications. Researchers and companies have developed a variety of models to identify toxicity in online conversations, reviews, or comments with mixed successes. However, many existing approaches have learned to incorrectly associate non-toxic comments that have certain trigger-words (e.g. gay, lesbian, black, muslim) as a potential source of toxicity. In this paper, we evaluate several state-of-the-art models with the specific focus of reducing model bias towards these commonly-attacked identity groups. We propose a multi-task learning model with an attention layer that jointly learns to predict the toxicity of a comment as well as the identities present in the comments in order to reduce this bias. We then compare our model to an array of shallow and deep-learning models using metrics designed especially to test for unintended model bias within these identity groups.

With the rapid development of knowledge bases (KBs), question answering (QA) based on KBs has become a hot research issue. The KB-QA technology can be divided into two technical routes: (1) symbol based representations, such as traditional semantic parsing, and (2) distribution based embedding. With the emergence of deep learning, the development of NLP has greatly promoted. The effect of KB-QA can be improved by combining deep learning with the above two technical routes respectively. In this paper, the impact of the second route (i.e., Distribution Embedding) combined with deep learning is mainly discussed. This route can be divided into pipeline frameworks and end-to-end frameworks. For comprehensive analysis, two frameworks (i.e., a pipeline framework, an end-to-end framework) are proposed to focus on answering single-relation factoid questions. In both of two frameworks, the effect of context information on the quality of QA is studied, such as the entity’s notable type, out-degree. In the pipeline framework, it includes two cascaded steps: entity detection and relation detection. In this framework, multiple modules need to be built, and corresponding training data sets must be constructed for them respectively. The entire process of the pipleine framework is complex, costly and has the problem of error propagation. In the end-to-end framework, the two subtasks of entity detection and relation detection are merged together, and then combined into one framework. Questions, entities and relations are mapped into the same semantic space through the encoding of the recurrent neural network. Moreover, the question-entity similarity and the question-relation similarity are calculated, so that the candidate answers can be sorted and selected. Moreover, character-level(char-level) encoding and self-attention mechanisms are combined using weight sharing and multi-task strategies to enhance the accuracy of QA. Experimental results show that context information can get better results of simple QA whether it is the pipeline framework or the end-to-end framework. In addition, the end-to-end framework achieves results competitive with state-of-the-art approaches in terms of accuracy.

With the advent of the World Wide Web (WWW) and the rise of digital media consumption, abundant information is available nowadays for any topic. But these days users often suffer from information overload posing a great challenge for finding relevant and important information. To alleviate this information overload and provide significant value to the users, there is a need for automatic information preparation methods. Such methods need to support users by discovering and recommending important information while filtering redundant and irrelevant information. They need to ensure that the users do not drown in, but rather benefit from the prepared information. However, the definition of what is relevant and important is subjective and highly specific to the user’s information need and the task at hand. Therefore, a method must continually learn from the feedback of its users. In this thesis, we propose new approaches to put the human in the loop in order to interactively prepare information along the three major lines of research: information aggregation, condensation, and recommendation. For multiple well-studied tasks in natural language processing, we point out the limitation of existing methods and discuss how our approach can successfully close the gap to the human upper bound by considering user feedback and adapting to the user’s information need. We put a particular focus on applications in digital journalism and introduce the new task of live blog summarization. We show that the corpora we create for this task are highly heterogeneous as compared to the standard summarization datasets which pose new challenges to previously proposed non-interactive methods. One way to alleviate information overload is information aggregation. We focus on the corresponding task of multi-document summarization and argue that previously proposed methods are of limited usefulness in the real-world application as they do not take the users’ goal into account. To address these drawbacks, we propose an interactive summarization loop to iteratively create and refine multi-document summaries based on the users’ feedback. We investigate sampling strategies based on active machine learning and joint optimization to reduce the number of iterations and the amount of user feedback required. Our approach significantly improves the quality of the summaries and reaches a performance near the human upper bound. We present a system demonstration implementing the interactive summarization loop, study its scalability, and highlight its use cases in exploring document collections and creating focused summaries in journalism. For information condensation, we investigate a text compression setup. We address the problem of neural models requiring huge amounts of training data and propose a new interactive text compression method to reduce the need for large-scale annotated data. We employ state-of-the-art Seq2Seq text compression methods as our base models and propose an active learning setup with multiple sampling strategies to efficiently use minimal training data. We find that our method significantly reduces the amount of data needed to train and that it adapts well to new datasets and domains. We finally focus on information recommendation and discuss the need for explainable models in machine learning. We propose a new joint recommendation system of rating prediction and review summarization, which shows major improvements over state-of-the-art systems in both the rating prediction and the review summarization task. By solving this task jointly based on multi-task learning techniques, we furthermore obtain explanations for a rating by showing the generated review summary marked based on the model’s attention and a histogram of user preferences learned from the reviews of the users. We conclude the thesis with a summary of how human-in-the-loop approaches improve information preparation systems and envision the use of interactive machine learning methods also for other areas of natural language processing.

With the advent of FrameNet and PropBank, many semantic role labeling (SRL) systems have been proposed in English. Although research on Japanese predicate argument structure analysis (PASA) has been conducted, most studies focused on surface cases. There are only few previous works on Japanese SRL for deep cases, and their models’ accuracies are low. Therefore, we propose a hierarchical multitask learning method with dependency parsing (DP) and show that our model achieves state-of-the-art results in Japanese SRL. Also, we conduct experiments with a joint model that performs both argument identification and argument classification simultaneously. The result suggests that multitasking with DP is mainly effective for argument identification.

While neural dependency parsers provide state-of-the-art accuracy for several languages, they still rely on large amounts of costly labeled training data. We demonstrate that in the small data regime, where uncertainty around parameter estimation and model prediction matters the most, Bayesian neural modeling is very effective. In order to overcome the computational and statistical costs of the approximate inference step in this framework, we utilize an efficient sampling procedure via stochastic gradient Langevin dynamics to generate samples from the approximated posterior. Moreover, we show that our Bayesian neural parser can be further improved when integrated into a multi-task parsing and POS tagging framework, designed to minimize task interference via an adversarial procedure. When trained and tested on 6 languages with less than 5k training instances, our parser consistently outperforms the strong bilstm baseline (Kiperwasser and Goldberg, 2016). Compared with the biaffine parser (Dozat et al., 2017) our model achieves an improvement of up to 3% for Vietnames and Irish, while our multi-task model achieves an improvement of up to 9% across five languages: Farsi, Russian, Turkish, Vietnamese, and Irish.

While keyphrase extraction has received considerable attention in recent years, relatively few studies exist on extracting keyphrases from social media platforms such as Twitter, and even fewer for extracting disaster-related keyphrases from such sources. During a disaster, keyphrases can be extremely useful for filtering relevant tweets that can enhance situational awareness. Previously, joint training of two different layers of a stacked Recurrent Neural Network for keyword discovery and keyphrase extraction had been shown to be effective in extracting keyphrases from general Twitter data. We improve the model's performance on both general Twitter data and disaster-related Twitter data by incorporating contextual word embeddings, POS-tags, phonetics, and phonological features. Moreover, we discuss the shortcomings of the often used F1-measure for evaluating the quality of predicted keyphrases with respect to the ground truth annotations. Instead of the F1-measure, we propose the use of embedding-based metrics to better capture the correctness of the predicted keyphrases. In addition, we also present a novel extension of an embedding-based metric. The extension allows one to better control the penalty for the difference in the number of ground-truth and predicted keyphrases.

When parsing morphologically-rich languages with neural models, it is beneficial to model input at the character level, and it has been claimed that this is because character-level models learn morphology. We test these claims by comparing character-level models to an oracle with access to explicit morphological analysis on twelve languages with varying morphological typologies. Our results highlight many strengths of character-level models, but also show that they are poor at disambiguating some words, particularly in the face of case syncretism. We then demonstrate that explicitly modeling morphological case improves our best model, showing that character-level models can benefit from targeted forms of explicit morphological modeling.

We utilize multi-task learning to improve argument mining in persuasive online discussions, in which both micro-level and macro-level argumentation must be taken into consideration. Our models learn to identify argument components and the relations between them at the same time. We also tackle the low-precision which arises from imbalanced relation data by experimenting with SMOTE and XGBoost. Our approaches improve over baselines that use the same pre-trained language model but process the argument component task and two relation tasks separately. Furthermore, our results suggest that the tasks to be incorporated into multi-task learning should be taken into consideration as using all relevant tasks does not always lead to the best performance.

We take a multi-task learning approach to the shared Task 1 at SemEval-2018. The general idea concerning the model structure is to use as little external data as possible in order to preserve the task relatedness and reduce complexity. We employ multi-task learning with hard parameter sharing to exploit the relatedness between sub-tasks. As a base model, we use a standard recurrent neural network for both the classification and regression subtasks. Our system ranks 32nd out of 48 participants with a Pearson score of 0.557 in the first subtask, and 20th out of 35 in the fifth subtask with an accuracy score of 0.464.

We study three general multi-task learning (MTL) approaches on 11 sequence tagging tasks. Our extensive empirical results show that in about 50% of the cases, jointly learning all 11 tasks improves upon either independent or pairwise learning of the tasks. We also show that pairwise MTL can inform us what tasks can benefit others or what tasks can be benefited if they are learned jointly. In particular, we identify tasks that can always benefit others as well as tasks that can always be harmed by others. Interestingly, one of our MTL approaches yields embeddings of the tasks that reveal the natural clustering of semantic and syntactic tasks. Our inquiries have opened the doors to further utilization of MTL in NLP.

We study the impact of a new multi-task learning approach in deep neural network for polarity and irony detection in Italian Twitter posts. We compare this approach with traditional single-task learning models. The different behavior of the two approaches shows the effectiveness of the proposed method that is able to combine the information from the two tasks improving the accuracy in both tasks. This is particularly true on edge cases in which knowledge about the two tasks is needed to classify a tweet, this is the case, for example, when the literal polarity of a tweet is inverted by irony.

We study approaches to improve fine-grained short answer Question Answering models by integrating coarse-grained data annotated for paragraph-level relevance and show that coarsely annotated data can bring significant performance gains. Experiments demonstrate that the standard multi-task learning approach of sharing representations is not the most effective way to leverage coarse-grained annotations. Instead, we can explicitly model the latent fine-grained short answer variables and optimize the marginal log-likelihood directly or use a newly proposed \emph{posterior distillation} learning objective. Since these latent-variable methods have explicit access to the relationship between the fine and coarse tasks, they result in significantly larger improvements from coarse supervision.

We study a variant of domain adaptation for named-entity recognition where multiple, heterogeneously tagged training sets are available. Furthermore, the test tag-set is not identical to any individual training tag-set. Yet, the relations between all tags are provided in a tag hierarchy, covering the test tags as a combination of training tags. This setting occurs when various datasets are created using different annotation schemes. This is also the case of extending a tag-set with a new tag by annotating only the new tag in a new dataset. We propose to use the given tag hierarchy to jointly learn a neural network that shares its tagging layer among all tag-sets. We compare this model to combining independent models and to a model based on the multitasking approach. Our experiments show the benefit of the tag-hierarchy model, especially when facing non-trivial consolidation of tag-sets.

We reduce the task of (span-based) PropBank-style semantic role labeling (SRL) to syntactic dependency parsing. Our approach is motivated by our empirical analysis that shows three common syntactic patterns account for over 98% of the SRL annotations for both English and Chinese data. Based on this observation, we present a conversion scheme that packs SRL annotations into dependency tree representations through joint labels that permit highly accurate recovery back to the original format. This representation allows us to train statistical dependency parsers to tackle SRL and achieve competitive performance with the current state of the art. Our findings show the promise of syntactic dependency trees in encoding semantic role relations within their syntactic domain of locality, and point to potential further integration of syntactic methods into semantic role labeling in the future.

We propose the technique of quasi-multitask learning (Q-MTL), a simple and easy to implement modification of standard multitask learning, in which the tasks to be modeled are identical. With this easy modification of a standard neural classifier we can get benefits similar to an ensemble of classifiers with a fraction of the resources required.We illustrate it through a series of sequence labeling experiments over a diverse set of languages, that applying Q-MTL consistently increases the generalization ability of the applied models. The proposed architecture can be regarded as a new regularization technique that encourages the model to develop an internal representation of the problem at hand which is beneficial to multiple output units of the classifier at the same time. Our experiments corroborate that by relying on the proposed algorithm, we can approximate the quality of an ensemble of classifiers at a fraction of computational resources required. Additionally, our results suggest that Q-MTL handles the presence of noisy training labels better than ensembles.

We propose dynamic task prioritization for multitask learning. This allows a model to dynamically prioritize difficult tasks during training, where difficulty is inversely proportional to performance, and where difficulty changes over time. In contrast to curriculum learning, where easy tasks are prioritized above difficult tasks, we present several studies showing the importance of prioritizing difficult tasks first. We observe that imbalances in task difficulty can lead to unnecessary emphasis on easier tasks, thus neglecting and slowing progress on difficult tasks. Motivated by this finding, we introduce a notion of dynamic task prioritization to automatically prioritize more difficult tasks by adaptively adjusting the mixing weight of each task’s loss objective. Additional ablation studies show the impact of the task hierarchy, or the task ordering, when explicitly encoded in the network architecture. Our method outperforms existing multitask methods and demonstrates competitive results with modern single-task models on the COCO and MPII datasets.