Although the curse of multilinguality significantly restricts the language abilities of multilingual models in monolingual settings, researchers now still have to rely on multilingual models to develop state-of-the-art systems in Vietnamese Machine Reading Comprehension. This difficulty in researching is because of the limited number of high-quality works in developing Vietnamese language models. In order to encourage more work in this research field, we present a comprehensive analysis of language weaknesses and strengths of current Vietnamese monolingual models using the downstream task of Machine Reading Comprehension. From the analysis results, we suggest new directions for developing Vietnamese language models. Besides this main contribution, we also successfully reveal the existence of artifacts in Vietnamese Machine Reading Comprehension benchmarks and suggest an urgent need for new high-quality benchmarks to track the progress of Vietnamese Machine Reading Comprehension. Moreover, we also introduced a minor but valuable modification to the process of annotating unanswerable questions for Machine Reading Comprehension from previous work. Our proposed modification helps improve the quality of unanswerable questions to a higher level of difficulty for Machine Reading Comprehension systems to solve.

Representation of coreferential relations is a challenging and actively studied topic for pro-drop and morphologically rich languages (PD-MRLs) due to dropped pronouns (e.g., null subjects and omitted possessive pronouns). These phenomena require a representation scheme at the morphology level and enhanced evaluation methods. In this paper, we propose a representation & evaluation scheme to incorporate dropped pronouns into coreference resolution and validate it on the Turkish language. Using the scheme, we extend the annotations on the only existing Turkish coreference dataset, which originally did not contain annotations for dropped pronouns. We provide publicly available pre and post processors to enhance the prominent CoNLL coreference scorer also to cover coreferential relations arising from dropped pronouns. As a final step, the paper reports the first neural Turkish coreference resolution results in the literature. Although validated on Turkish, the proposed scheme is language-independent and may be used for other PD-MRLs.

Dealing with humor is an important step to develop Natural Language Processing tools capable of handling sophisticated semantic and pragmatic knowledge. In this context, this PhD thesis focuses on the automatic generation and recognition of verbal punning humor in Portuguese, which is still an underdeveloped language when compared to English. One of the main goals of this research is to conciliate Natural Language Generation computational models with existing theories of humor from the Humanities while avoiding mere generation by including contextual information into the generation process. Another point that is of utmost importance is the inclusion of the listener as an active part in the process of understanding and creating humor; we hope to achieve this by using concepts from Recommender Systems in our methods. Ultimately, we want to not only advance the current state-of-the-art in humor generation and recognition, but also to help the general Portuguese-speaking research community with methods, tools and resources that may aid in the development of further techniques for this language. We also expect our systems to provide insightful ideas about how humor is created and perceived by both humans and machines.

Automatic code generation from natural language descriptions can be highly beneficial during the process of software development. In this work, we propose GAP-Gen, a Guided Automatic Python Code Generation method based on Python syntactic constraints and semantic constraints. We first introduce Python syntactic constraints in the form of Syntax-Flow, which is a simplified version of Abstract Syntax Tree (AST) reducing the size and high complexity of Abstract Syntax Tree but maintaining crucial syntactic information of Python code. In addition to Syntax-Flow, we introduce Variable-Flow which abstracts variable and function names consistently through out the code. In our work, rather than pretraining, we focus on modifying the finetuning process which reduces computational requirements but retains high generation performance on automatic Python code generation task. GAP-Gen fine-tunes the transformer based language models T5 and CodeT5 using the Code-to-Docstring datasets CodeSearchNet, CodeSearchNet AdvTest and Code-Docstring Corpus from EdinburghNLP. Our experiments show that GAP-Gen achieves better results on automatic Python code generation task than previous works

Clinical natural language processing aims to tackle language and prediction tasks using text from medical practice, such as clinical notes, prescriptions, and discharge summaries. Several approaches have been tried to deal with these tasks. Since 2017, pre-trained language models (PLMs) have achieved state-of-the-art performance in many tasks. However, most works have been developed in English. This PhD research proposal addresses the development of PLMs for clinical NLP in Spanish. To carry out this study, we will build a clinical corpus big enough to implement a functional PLM. We will test several PLM architectures and evaluate them with language and prediction tasks. The novelty of this work lies in the use of only clinical text, while previous clinical PLMs have used a mix of general, biomedical, and clinical text.

The demand for multimodal dialogue systems has been rising in various domains, emphasizing the importance of interpreting multimodal inputs from conversational and situational contexts. One main challenge in multimodal dialogue understanding is multimodal object identification, which constitutes the ability to identify objects relevant to a multimodal user-system conversation. We explore three methods to tackle this problem and evaluate them on the largest situated dialogue dataset, SIMMC 2.1. Our best method, scene-dialogue alignment, improves the performance by ~20% F1-score compared to the SIMMC 2.1 baselines. We provide analysis and discussion regarding the limitation of our methods and the potential directions for future works.

Social chatbots have gained immense popularity, and their appeal lies not just in their capacity to respond to the diverse requests from users, but also in the ability to develop an emotional connection with users. To further develop and promote social chatbots, we need to concentrate on increasing user interaction and take into account both the intellectual and emotional quotient in the conversational agents. In this paper, we propose a multi-task framework that jointly identifies the emotion of a given dialogue and generates response in accordance to the identified emotion. We employ a BERT based network for creating an empathetic system and use a mixed objective function that trains the end-to-end network with both the classification and generation loss. Experimental results show that our proposed framework outperforms current state-of-the-art models.

With the rise in larger language models, researchers started exploiting them by pivoting the downstream tasks as language modeling tasks using prompts. In this work, we convert the Named Entity Recognition task into a seq2seq task by generating the synthetic sentences using templates. Our main contribution is the conversion framework which provides faster inference. In addition, we test our method’s performance in resource-rich, low resource and domain transfer settings. Results show that our method achieves comparable results in the resource-rich setting and outperforms the current seq2seq paradigm state-of-the-art approach in few-shot settings. Through the experiments, we observed that the negative examples play an important role in model’s performance. We applied our approach over BART and T5-base models, and we notice that the T5 architecture aligns better with our task. The work is performed on the datasets in English language.

Generating polite responses is essential to build intelligent and engaging dialogue systems. However, this task is far from well-explored due to the difficulties of rendering a particular style in coherent responses, especially when parallel datasets for regular-to-polite pairs are usually unavailable. This paper proposes a polite chatbot that can produce responses that are polite and coherent to the given context. In this study, a politeness transfer model is first used to generate polite synthetic dialogue pairs of contexts and polite utterances. Then, these synthetic pairs are employed to train a dialogue model. Automatic and human evaluations demonstrate that our method outperforms baselines in producing polite dialogue responses while staying competitive in terms of coherent to the given context.

Writing is an important element of language learning, and an increasing amount of learner writing is taking place in online environments. Teachers can provide valuable feedback by commenting on learner text. However, providing relevant feedback for every issue for every student can be time-consuming. To address this, we turn to the NLP subfield of feedback comment generation, the task of automatically generating explanatory notes for learner text with the goal of enhancing learning outcomes. However, freely-generated comments may mix multiple topics seen in the training data or even give misleading advice. In this thesis proposal, we seek to address these issues by categorizing comments and constraining the outputs of noisy classes. We describe an annotation scheme for feedback comment corpora using comment topics with a broader scope than existing typologies focused on error correction. We outline plans for experiments in grouping and clustering, replacing particularly diverse categories with modular templates, and comparing the generation results of using different linguistic features and model architectures with the original dataset versus the newly annotated one. This paper presents the first two years (the master’s component) of a research project for a five-year combined master’s and Ph.D program.

While text-based medical applications have become increasingly prominent, access to clinicaldata remains a major concern. To resolve this issue, further de-identification and anonymization of the data are required. This might, however, alter the contextual information within the clinical texts and therefore influence the learning and performance of possible language models. This paper systematically analyses the potential effects of various anonymization techniques on the performance of state-of-the-art machine learning models based on several datasets corresponding to five different NLP tasks. On this basis, we derive insightful findings and recommendations concerning text anonymization with regard to the performance of machine learning models. In addition, we present a simple re-identification attack applied to the anonymized text data, which can break the anonymization.

Today, human assistants are often replacedby chatbots, designed to communicate via natural language, however, some disadvantages are notorious with this replacement. This PhD thesis project consists of researching, implementing, and testing a solution for guiding the action of a human in a contact center. It will start with the discovery and creation of datasets in Portuguese.Next, it will go through three main components: Extraction for processing dialogs and using the information todescribe interactions; Representation for discovering the most frequent dialog flowsrepresented by graphs; Guidance for helping the agent during a new dialog. These will be integrated in a single framework. In order to avoid service degradation resulting from the adoption of chatbots, this work aims to explore technologies in order to increase the efficiency of the human’s job without losing human contact.

Question Generation (QG) systems have shown promising results in reducing the time and effort required to create questions for students. Typically, a first step in QG is to select the content to design a question for. In an educational setting, it is crucial that the resulting questions cover the most relevant/important pieces of knowledge the student should have acquired. Yet, current QG systems either consider just a single sentence or paragraph (thus do not include a selection step), or do not consider this educational viewpoint of content selection. Aiming to fill this research gap with a solution for educational document level QG, we thus propose to select contents for QG based on relevance and topic diversity. We demonstrate the effectiveness of our proposed content selection strategy for QG on 2 educational datasets. In our performance assessment, we also highlight limitations of existing QG evaluation metrics in light of the content selection problem.

Automatic error type classification is an important process in both learner corpora creation and evaluation of large-scale grammatical error correction systems. Rule-based classifier approaches such as ERRANT have been widely used to classify edits between correct-erroneous sentence pairs into predefined error categories. However, the used error categories are far from being universal yielding many language specific variants of ERRANT.In this paper, we discuss the applicability of the previously introduced grammatical error types to an agglutinative language, Turkish. We suggest changes on current error categories and discuss a hierarchical structure to better suit the inflectional and derivational properties of this morphologically highly rich language. We also introduce ERRANT-TR, the first automatic error type classification toolkit for Turkish. ERRANT-TR currently uses a rule-based error type classification pipeline which relies on word level morphological information. Due to unavailability of learner corpora in Turkish, the proposed system is evaluated on a small set of 106 annotated sentences and its performance is measured as 77.04% F0.5 score. The next step is to use ERRANT-TR for the development of a Turkish learner corpus.

Previous work has established that RNNs with an unbounded activation function have the capacity to count exactly. However, it has also been shown that RNNs are challenging to train effectively and generally do not learn exact counting behaviour. In this paper, we focus on this problem by studying the simplest possible RNN, a linear single-cell network. We conduct a theoretical analysis of linear RNNs and identify conditions for the models to exhibit exact counting behaviour. We provide a formal proof that these conditions are necessary and sufficient. We also conduct an empirical analysis using tasks involving a Dyck-1-like Balanced Bracket language under two different settings. We observe that linear RNNs generally do not meet the necessary and sufficient conditions for counting behaviour when trained with the standard approach. We investigate how varying the length of training sequences and utilising different target classes impacts model behaviour during training and the ability of linear RNN models to effectively approximate the indicator conditions.

Recent task-oriented dialogue systems are trained on annotated dialogues, which, in turn, reflect certain domain information (e.g., restaurants or hotels in a given region). However, when such domain knowledge changes (e.g., new restaurants open), the initial dialogue model may become obsolete, decreasing the overall performance of the system. Through a number of experiments, we show, for instance, that adding 50% of new slot-values reduces of about 55% the dialogue state-tracker performance. In light of such evidence, we suggest that automatic adaptation of training dialogues is a valuable option for re-training obsolete models. We experimented with a dialogue adaptation approach based on fine-tuning a generative language model on domain changes, showing that a significant reduction of performance decrease can be obtained.