Systematicity is thought to be a key inductive bias possessed by humans that is lacking in standard natural language processing systems such as those utilizing transformers. In this work, we investigate the extent to which the failure of transformers on systematic generalization tests can be attributed to a lack of linguistic abstraction in its attention mechanism. We develop a novel modification to the transformer by implementing two separate input streams: a role stream controls the attention distributions (i.e., queries and keys) at each layer, and a filler stream determines the values. Our results show that when abstract role labels are assigned to input sequences and provided to the role stream, systematic generalization is improved.

Building open-domain dialogue systems capable of rich human-like conversational ability is one of the fundamental challenges in language generation. However, even with recent advancements in the field, existing open-domain generative models fail to capture and utilize external knowledge, leading to repetitive or generic responses to unseen utterances. Current work on knowledge-grounded dialogue generation primarily focuses on persona incorporation or searching a fact-based structured knowledge source such as Wikipedia. Our method takes a broader and simpler approach, which aims to improve the raw conversation ability of the system by mimicking the human response behavior through casual interactions found on social media. Utilizing a joint retriever-generator setup, the model queries a large set of filtered comment data from Reddit to act as additional context for the seq2seq generator. Automatic and human evaluations on open-domain dialogue datasets demonstrate the effectiveness of our approach.

Neural models trained with large amount of parallel data have achieved impressive performance in abstractive summarization tasks. However, large-scale parallel corpora are expensive and challenging to construct. In this work, we introduce a low-cost and effective strategy, ExtraPhrase, to augment training data for abstractive summarization tasks. ExtraPhrase constructs pseudo training data in two steps: extractive summarization and paraphrasing. We extract major parts of an input text in the extractive summarization step and obtain its diverse expressions with the paraphrasing step. Through experiments, we show that ExtraPhrase improves the performance of abstractive summarization tasks by more than 0.50 points in ROUGE scores compared to the setting without data augmentation. ExtraPhrase also outperforms existing methods such as back-translation and self-training. We also show that ExtraPhrase is significantly effective when the amount of genuine training data is remarkably small, i.e., a low-resource setting. Moreover, ExtraPhrase is more cost-efficient than the existing approaches

Including memory banks in a natural language processing architecture increases model capacity by equipping it with additional data at inference time. In this paper, we build upon kNN-LM (CITATION), which uses a pre-trained language model together with an exhaustive kNN search through the training data (memory bank) to achieve state-of-the-art results. We investigate whether we can improve the kNN-LM performance by instead training a LM with the knowledge that we will be using a kNN post-hoc. We achieved significant improvement using our method on language modeling tasks on WIKI-2 and WIKI-103. The main phenomenon that we encounter is that adding a simple L2 regularization on the activations (not weights) of the model, a transformer, improves the post-hoc kNN classification performance. We explore some possible reasons for this improvement. In particular, we find that the added L2 regularization seems to improve the performance for high-frequency words without deteriorating the performance for low frequency ones.

Earlier NLP studies on framing in political discourse have focused heavily on shallow classification of issue framing, while framing effect arising from pragmatic cues remains neglected. We put forward this latter type of framing as “pragmatic framing”. To bridge this gap, we take presupposition-triggering adverbs such as ‘again’ as a study case, and quantitatively investigate how different German newspapers use them to covertly evoke different attitudinal subtexts in their report on the event “European Refugee Crisis” (2014-2018). Our study demonstrates the crucial role of presuppositions in framing, and emphasizes the necessity of more attention on pragmatic framing in the research of automated framing detection.

Despite their outstanding performance, large language models (LLMs) suffer notorious flaws related to their preference for shallow textual relations over full semantic complexity of the problem. This proposal investigates a common denominator of this problem in their weak ability to generalise outside of the training domain. We survey diverse research directions providing estimations of model generalisation ability and find that incorporating some of these measures in the training objectives leads to enhanced distributional robustness of neural models. Based on these findings, we present future research directions enhancing the robustness of LLMs.

Retrieval-augmented generation (RAG) methods have been receiving increasing attention from the NLP community and achieved state-of-the-art performance on many NLP downstream tasks. Compared with conventional pre-trained generation models, RAG methods have remarkable advantages such as easy knowledge acquisition, strong scalability, and low training cost. Although existing RAG models have been applied to various knowledge-intensive NLP tasks, such as open-domain QA and dialogue systems, most of the work has focused on retrieving unstructured text documents from Wikipedia. In this paper, I first elaborate on the current obstacles to retrieving knowledge from a single-source homogeneous corpus. Then, I demonstrate evidence from both existing literature and my experiments, and provide multiple solutions on retrieval-augmented generation methods across heterogeneous knowledge.

Information Retriever (IR) aims to find the relevant documents (e.g. snippets, passages, and articles) to a given query at large scale. IR plays an important role in many tasks such as open domain question answering and dialogue systems, where external knowledge is needed. In the past, searching algorithms based on term matching have been widely used. Recently, neural-based algorithms (termed as neural retrievers) have gained more attention which can mitigate the limitations of traditional methods. Regardless of the success achieved by neural retrievers, they still face many challenges, e.g. suffering from a small amount of training data and failing to answer simple entity-centric questions. Furthermore, most of the existing neural retrievers are developed for pure-text query. This prevents them from handling multi-modality queries (i.e. the query is composed of textual description and images). This proposal has two goals. First, we introduce methods to address the abovementioned issues of neural retrievers from three angles, new model architectures, IR-oriented pretraining tasks, and generating large scale training data. Second, we identify the future research direction and propose potential corresponding solution.

Cognitive distortions are counterproductive patterns of thinking that are one of the targets of cognitive behavioral therapy (CBT). These can be challenging for clinicians to detect, especially those without extensive CBT training or supervision. Text classification methods can approximate expert clinician judgment in the detection of frequently occurring cognitive distortions in text-based therapy messages. However, performance with infrequent distortions is relatively poor. In this study, we address this sparsity problem with two approaches: Data Augmentation and Domain-Specific Model. The first approach includes Easy Data Augmentation, back translation, and mixup techniques. The second approach utilizes a domain-specific pretrained language model, MentalBERT. To examine the viability of different data augmentation methods, we utilized a real-world dataset of texts between therapists and clients diagnosed with serious mental illness that was annotated for distorted thinking. We found that with optimized parameter settings, mixup was helpful for rare classes. Performance improvements with an augmented model, MentalBERT, exceed those obtained with data augmentation.

We aim to overcome the lack of diversity in responses of current dialogue systems and to develop a dialogue system that is engaging as a conversational partner. We propose a generator-evaluator model that evaluates multiple responses generated by a response generator and selects the best response by an evaluator. By generating multiple responses, we obtain diverse responses. We conduct human evaluations to compare the output of the proposed system with that of a baseline system. The results of the human evaluations showed that the proposed system’s responses were often judged to be better than the baseline system’s, and indicated the effectiveness of the proposed method.

State of the art performances for entity extraction tasks are achieved by supervised learning, specifically, by fine-tuning pretrained language models such as BERT. As a result, annotating application specific data is the first step in many use cases. However, no practical guidelines are available for annotation requirements. This work supports practitioners by empirically answering the frequently asked questions (1) how many training samples to annotate? (2) which examples to annotate? We found that BERT achieves up to 80% F1 when fine-tuned on only 70 training examples, especially on biomedical domain. The key features for guiding the selection of high performing training instances are identified to be pseudo-perplexity and sentence-length. The best training dataset constructed using our proposed selection strategy shows F1 score that is equivalent to a random selection with twice the sample size. The requirement of only a small number of training data implies cheaper implementations and opens door to wider range of applications.

Multimodal Neural Machine Translation is focusing on using visual information to translate sentences in the source language into the target language. The main idea is to utilise information from visual modalities to promote the output quality of the text-based translation model. Although the recent multimodal strategies extract the most relevant visual information in images, the effectiveness of using visual information on translation quality changes based on the text dataset. Due to this, this work studies the impact of leveraging visual information in multimodal translation models of ambiguous sentences. Our experiments analyse the Multi30k evaluation dataset and calculate ambiguity scores of sentences based on the WordNet hierarchical structure. To calculate the ambiguity of a sentence, we extract the ambiguity scores for all nouns based on the number of senses in WordNet. The main goal is to find in which sentences, visual content can improve the text-based translation model. We report the correlation between the ambiguity scores and translation quality extracted for all sentences in the English-German dataset.

Dialogue systems without consistent responses are not attractive. In this study, we build a dialogue system that can respond based on a given character setting (persona) to bring consistency. Considering the trend of the rapidly increasing scale of language models, we propose an approach that uses prompt-tuning, which has low learning costs, on pre-trained large-scale language models. The results of the automatic and manual evaluations in English and Japanese show that it is possible to build a dialogue system with more natural and personalized responses with less computational resources than fine-tuning.

While there have been advances in Natural Language Processing (NLP), their success is mainly gained by applying a self-attention mechanism into single or multi-modalities. While this approach has brought significant improvements in multiple downstream tasks, it fails to capture the interaction between different entities. Therefore, we propose MM-GATBT, a multimodal graph representation learning model that captures not only the relational semantics within one modality but also the interactions between different modalities. Specifically, the proposed method constructs image-based node embedding which contains relational semantics of entities. Our empirical results show that MM-GATBT achieves state-of-the-art results among all published papers on the MM-IMDb dataset.

Feature structures have been several times considered to enrich categorial grammars in order to build fine-grained grammars. Most attempts to unify both frameworks either model categorial types as feature structures or add feature structures on top of categorial types. We pursue a different approach: using feature structure as categorial atomic types. In this article, we present a procedure to create, from a simplified HPSG grammar, an equivalent abstract categorial grammar (ACG). We represent a feature structure by the enumeration of its totally well-typed upper bounds, so that unification can be simulated as intersection. We implement this idea as a meta-ACG preprocessor.

Character identification is a key element for many narrative-related tasks. To implement it, the baseform of the name of the character (or lemma) needs to be identified, so different appearances of the same character in the narrative could be aligned. In this paper we tackle this problem in translated texts (English–Finnish translation direction), where the challenge regarding lemmatizing foreign names in an agglutinative language appears. To solve this problem, we present and compare several methods. The results show that the method based on a search for the shortest version of the name proves to be the easiest, best performing (83.4% F1), and most resource-independent.

Word Sense Disambiguation (WSD) is a core task in Natural Language Processing (NLP). Ancient Chinese has rarely been used in WSD tasks, however, as no public dataset for ancient Chinese WSD tasks exists. Creation of an ancient Chinese dataset is considered a significant challenge because determining the most appropriate sense in a context is difficult and time-consuming owing to the different usages in ancient and modern Chinese. Actually, no public dataset for ancient Chinese WSD tasks exists. To solve the problem of ancient Chinese WSD, we annotate part of Pre-Qin (221 BC) text Zuo Zhuan using a copyright-free dictionary to create a public sense-tagged dataset. Then, we apply a simple Nearest Neighbors (k-NN) method using a pre-trained language model to the dataset. Our code and dataset will be available on GitHub.

We present ViT5, a pretrained Transformer-based encoder-decoder model for the Vietnamese language. With T5-style self-supervised pretraining, ViT5 is trained on a large corpus of high-quality and diverse Vietnamese texts. We benchmark ViT5 on two downstream text generation tasks, Abstractive Text Summarization and Named Entity Recognition. Although Abstractive Text Summarization has been widely studied for the English language thanks to its rich and large source of data, there has been minimal research into the same task in Vietnamese, a much lower resource language. In this work, we perform exhaustive experiments on both Vietnamese Abstractive Summarization and Named Entity Recognition, validating the performance of ViT5 against many other pretrained Transformer-based encoder-decoder models. Our experiments show that ViT5 significantly outperforms existing models and achieves state-of-the-art results on Vietnamese Text Summarization. On the task of Named Entity Recognition, ViT5 is competitive against previous best results from pretrained encoder-based Transformer models. Further analysis shows the importance of context length during the self-supervised pretraining on downstream performance across different settings.

We provide a study of how induced model sparsity can help achieve compositional generalization and better sample efficiency in grounded language learning problems. We consider simple language-conditioned navigation problems in a grid world environment with disentangled observations. We show that standard neural architectures do not always yield compositional generalization. To address this, we design an agent that contains a goal identification module that encourages sparse correlations between words in the instruction and attributes of objects, composing them together to find the goal. The output of the goal identification module is the input to a value iteration network planner. Our agent maintains a high level of performance on goals containing novel combinations of properties even when learning from a handful of demonstrations. We examine the internal representations of our agent and find the correct correspondences between words in its dictionary and attributes in the environment.

This paper explores how humans conduct conversations with images by investigating an open-domain image conversation dataset, ImageChat. We examined the conversations with images from the perspectives of image relevancy and image information. We found that utterances/conversations are not always related to the given image, and conversation topics diverge within three turns about half of the time. Besides image objects, more comprehensive non-object image information is also indispensable. After inspecting the causes, we suggested that understanding the overall scenario of image and connecting objects based on their high-level attributes might be very helpful to generate more engaging open-domain conversations when an image is presented. We proposed enriching the image information with image caption and object tags based on our analysis. With our proposed image⁺ features, we improved automatic metrics including BLEU and Bert Score, and increased the diversity and image-relevancy of generated responses to the strong baseline. The result verifies that our analysis provides valuable insights and could facilitate future research on open-domain conversations with images.

It is well known that textual data on the internet and other digital platforms contain significant levels of bias and stereotypes. Various research findings have concluded that biased texts have significant effects on target demographic groups. For instance, masculine-worded job advertisements tend to be less appealing to female applicants. In this paper, we present a text-style transfer model that can be trained on non-parallel data and be used to automatically mitigate bias in textual data. Our style transfer model improves on the limitations of many existing text style transfer techniques such as the loss of content information. Our model solves such issues by combining latent content encoding with explicit keyword replacement. We will show that this technique produces better content preservation whilst maintaining good style transfer accuracy.

TextHide was recently proposed to protect the training data via instance encoding in natural language domain. Due to the lack of theoretic privacy guarantee, such instance encoding scheme has been shown to be vulnerable against privacy attacks, e.g., reconstruction attack. To address such limitation, we revise the instance encoding scheme with differential privacy and thus provide a provable guarantee against privacy attacks. The experimental results also show that the proposed scheme can defend against privacy attacks while ensuring learning utility (as a trade-off).

In the open book question answering (OBQA) task, selecting the relevant passages and sentences from distracting information is crucial to reason the answer to a question. HotpotQA dataset is designed to teach and evaluate systems to do both passage ranking and sentence selection. Many existing frameworks use separate models to select relevant passages and sentences respectively. Such systems not only have high complexity in terms of the parameters of models but also fail to take the advantage of training these two tasks together since one task can be beneficial for the other one. In this work, we present a simple yet effective framework to address these limitations by jointly ranking passages and selecting sentences. Furthermore, we propose consistency and similarity constraints to promote the correlation and interaction between passage ranking and sentence selection. The experiments demonstrate that our framework can achieve competitive results with previous systems and outperform the baseline by 28% in terms of exact matching of relevant sentences on the HotpotQA dataset.

In order to build more human-like cognitive agents, systems capable of detecting various human emotions must be designed to respond appropriately. Confusion, the combination of an emotional and cognitive state, is under-explored. In this paper, we build upon prior work to develop models that detect confusion from three modalities: video (facial features), audio (prosodic features), and text (transcribed speech features). Our research improves the data collection process by allowing for continuous (as opposed to discrete) annotation of confusion levels. We also craft models based on recurrent neural networks (RNNs) given their ability to predict sequential data. In our experiments, we find that text and video modalities are the most important in predicting confusion while the explored audio features are relatively unimportant predictors of confusion in our data.

The current study quantitatively (and qualitatively for an illustrative purpose) analyzes BERT’s layer-wise masked word prediction on an English corpus, and finds that (1) the layerwise localization of linguistic knowledge primarily shown in probing studies is replicated in a behavior-based design and (2) that syntactic and semantic information is encoded at different layers for words of different syntactic categories. Hypothesizing that the above results are correlated with the number of likely potential candidates of the masked word prediction, we also investigate how the results differ for tokens within multiword expressions.

This PhD project leverages advancements in multimodal large language models to build an inclusive collaboration feedback loop, in order to facilitate the automated detection, modeling, and feedback for participants developing general collaboration skills. This topic is important given the role of collaboration as an essential 21st century skill, the potential to ground large language models within learning theory and real-world practice, and the expressive potential of transformer models to support equity and inclusion. We address some concerns of integrating advances in natural language processing into downstream tasks such as the learning analytics feedback loop.

Machine learning in hyperbolic spaces has attracted much attention in natural language processing and many other fields. In particular, Hyperbolic Neural Networks (HNNs) have improved a wide variety of tasks, from machine translation to knowledge graph embedding. Although some studies have reported the effectiveness of embedding into the product of multiple hyperbolic spaces, HNNs have mainly been constructed in a single hyperbolic space, and their extension to product spaces has not been sufficiently studied. Therefore, we propose a novel method to extend a given HNN in a single space to a product of hyperbolic spaces. We apply our method to Hyperbolic Graph Convolutional Networks (HGCNs), extending several HNNs. Our model improved the graph node classification accuracy especially on datasets with tree-like structures. The results suggest that neural networks in a product of hyperbolic spaces can be more effective than in a single space in representing structural data.

The current chat dialogue systems implicitly consider the topic given the context, but not explicitly. As a result, these systems often generate inconsistent responses with the topic of the moment. In this study, we propose a dialogue system that responds appropriately following the topic by selecting the entity with the highest “topicality.” In topicality estimation, the model is trained through self-supervised learning that regards entities that appear in both context and response as the topic entities. In response generation, the model is trained to generate topic-relevant responses based on the estimated topicality. Experimental results show that our proposed system can follow the topic more than the existing dialogue system that considers only the context.

Automated metrics to evaluate dialogue systems like BLEU, METEOR, etc., weakly correlate with human judgments. Thus, human evaluation is often used to supplement these metrics for system evaluation. However, human evaluation is time-consuming as well as expensive. This paper provides an alternative approach to human evaluation with respect to three aspects: naturalness, informativeness, and quality in dialogue systems. I propose an approach based on fine-tuning the BERT model with three prediction heads, to predict whether the system-generated output is natural, fluent, and informative. I observe that the proposed model achieves an average accuracy of around 77% over these 3 labels. I also design a baseline approach that uses three different BERT models to make the predictions. Based on experimental analysis, I find that using a shared model to compute the three labels performs better than three separate models.

Named entity linking (NEL) in news is a challenging endeavour due to the frequency of unseen and emerging entities, which necessitates the use of unsupervised or zero-shot methods. However, such methods tend to come with caveats, such as no integration of suitable knowledge bases (like Wikidata) for emerging entities, a lack of scalability, and poor interpretability. Here, we consider person disambiguation in Quotebank, a massive corpus of speaker-attributed quotations from the news, and investigate the suitability of intuitive, lightweight, and scalable heuristics for NEL in web-scale corpora. Our best performing heuristic disambiguates 94% and 63% of the mentions on Quotebank and the AIDA-CoNLL benchmark, respectively. Additionally, the proposed heuristics compare favourably to the state-of-the-art unsupervised and zero-shot methods, Eigenthemes and mGENRE, respectively, thereby serving as strong baselines for unsupervised and zero-shot entity linking.

Each person has a unique personality which affects how they feel and convey emotions. Hence, speaker modeling is important for the task of emotion recognition in conversation (ERC). In this paper, we propose a novel graph-based ERC model which considers both conversational context and speaker personality. We model the internal state of the speaker (personality) as Static and Dynamic speaker state, where the Dynamic speaker state is modeled with a graph neural network based encoder. Experiments on benchmark dataset shows the effectiveness of our model. Our model outperforms baseline and other graph-based methods. Analysis of results also show the importance of explicit speaker modeling.

Abstractive summarization of medical dialogues presents a challenge for standard training approaches, given the paucity of suitable datasets. We explore the performance of state-of-the-art models with zero-shot and few-shot learning strategies and measure the impact of pretraining with general domain and dialogue-specific text on the summarization performance.

We introduce a novel tree-based model that learns its composition function together with its structure. The architecture produces sentence embeddings by composing words according to an induced syntactic tree. The parsing and the composition functions are explicitly connected and, therefore, learned jointly. As a result, the sentence embedding is computed according to an interpretable linguistic pattern and may be used on any downstream task. We evaluate our encoder on downstream tasks, and we observe that it outperforms tree-based models relying on external parsers. In some configurations, it is even competitive with Bert base model. Our model is capable of supporting multiple parser architectures. We exploit this property to conduct an ablation study by comparing different parser initializations. We explore to which extent the trees produced by our model compare with linguistic structures and how this initialization impacts downstream performances. We empirically observe that downstream supervision troubles producing stable parses and preserving linguistically relevant structures.

Transformer-based models have been achieving state-of-the-art results in several fields of Natural Language Processing. However, its direct application to speech tasks is not trivial. The nature of this sequences carries problems such as long sequence lengths and redundancy between adjacent tokens. Therefore, we believe that regular self-attention mechanism might not be well suited for it. Different approaches have been proposed to overcome these problems, such as the use of efficient attention mechanisms. However, the use of these methods usually comes with a cost, which is a performance reduction caused by information loss. In this study, we present the Multiformer, a Transformer-based model which allows the use of different attention mechanisms on each head. By doing this, the model is able to bias the self-attention towards the extraction of more diverse token interactions, and the information loss is reduced. Finally, we perform an analysis of the head contributions, and we observe that those architectures where all heads relevance is uniformly distributed obtain better results. Our results show that mixing attention patterns along the different heads and layers outperforms our baseline by up to 0.7 BLEU.

Compositionality has traditionally been understood as a major factor in productivity of language and, more broadly, human cognition. Yet, recently some research started to question its status showing that artificial neural networks are good at generalization even without noticeable compositional behavior. We argue some of these conclusions are too strong and/or incomplete. In the context of a two-agent communication game, we show that compositionality indeed seems essential for successful generalization when the evaluation is done on a suitable dataset.

Previous research has found that Acoustic Models (AM) of an Automatic Speech Recognition (ASR) system are susceptible to dialect variations within a language, thereby adversely affecting the ASR. To counter this, researchers have proposed to build a dialect-specific AM while keeping the Language Model (LM) constant for all the dialects. This study explores the effect of dialect mismatched LM by considering three different Telugu regional dialects: Telangana, Coastal Andhra, and Rayalaseema. We show that dialect variations that surface in the form of a different lexicon, grammar, and occasionally semantics can significantly degrade the performance of the LM under mismatched conditions. Therefore, this degradation has an adverse effect on the ASR even when dialect-specific AM is used. We show a degradation of up to 13.13 perplexity points when LM is used under mismatched conditions. Furthermore, we show a degradation of over 9% and over 15% in Character Error Rate (CER) and Word Error Rate (WER), respectively, in the ASR systems when using mismatched LMs over matched LMs.