Linear sequences of words are implicitly represented in our brains by hierarchical structures that organize the composition of words in sentences. Linguists formalize different frameworks to model this hierarchy; two of the most common syntactic frameworks are Constituency and Dependency. Constituency represents sentences as nested groups of phrases, while dependency represents a sentence by assigning relations between its words. Recently, the pursuit of intelligent machines has produced Language Models (LMs) capable of solving many language tasks with a human-level performance. Many studies now question whether LMs implicitly represent syntactic hierarchies. This thesis focuses on producing constituency and dependency structures from LMs in an unsupervised setting. I review the critical methods in this field and highlight a line of work that utilizes a numerical representation for binary constituency trees (Syntactic Distance). I present a detailed study on StructFormer (SF) (Shen et al., 2021), which retrofits a transformer encoder architecture with a parser network to produce constituency and dependency structures. I present six experiments to analyze and address this field's challenges; experiments include investigating the effect of repositioning the parser network within the SF architecture, evaluating subword-based induced trees, and benchmarking the models developed in the thesis experiments on linguistic tasks. Models benchmarking is performed by participating in the BabyLM challenge, published at CoNLL 2023 (Momen et al., 2023). The results of this thesis encourage further development in the direction of retrofitting transformer-based models to induce syntactic structures, supported by the acceptable performance of SF in different experimental settings and the observed limitations that require innovative solutions to advance the state of syntactic structure induction.

Knowledge graph embeddings (KGEs) were originally developed to infer true but missing facts in incomplete knowledge repositories. In this paper, we link knowledge graph completion and counterfactual reasoning via our new task CFKGR. We model the original world state as a knowledge graph, hypothetical scenarios as edges added to the graph, and plausible changes to the graph as inferences from logical rules. We create corresponding benchmark datasets, which contain diverse hypothetical scenarios with plausible changes to the original knowledge graph and facts that should be retained. We develop COULDD, a general method for adapting existing knowledge graph embeddings given a hypothetical premise, and evaluate it on our benchmark. Our results indicate that KGEs learn patterns in the graph without explicit training. We further observe that KGEs adapted with COULDD solidly detect plausible counterfactual changes to the graph that follow these patterns. An evaluation on human-annotated data reveals that KGEs adapted with COULDD are mostly unable to recognize changes to the graph that do not follow learned inference rules. In contrast, ChatGPT mostly outperforms KGEs in detecting plausible changes to the graph but has poor knowledge retention. In summary, CFKGR connects two previously distinct areas, namely KG completion and counterfactual reasoning.

Individual and social biases undermine the effectiveness of human advisers by inducing judgment errors which can disadvantage protected groups. In this paper, we study the influence these biases can have in the pervasive problem of fake news by evaluating human participants' capacity to identify false headlines. By focusing on headlines involving sensitive characteristics, we gather a comprehensive dataset to explore how human responses are shaped by their biases. Our analysis reveals recurring individual biases and their permeation into collective decisions. We show that demographic factors, headline categories, and the manner in which information is presented significantly influence errors in human judgment. We then use our collected data as a benchmark problem on which we evaluate the efficacy of adaptive aggregation algorithms. In addition to their improved accuracy, our results highlight the interactions between the emergence of collective intelligence and the mitigation of participant biases.

Social media platforms hold valuable insights, yet extracting essential information can be challenging. Traditional top-down approaches often struggle to capture critical signals in rapidly changing events. As global events evolve swiftly, social media narratives, including instances of disinformation, become significant sources of insights. To address the need for an inductive strategy, we explore a niche social media platform GAB and an established messaging service Telegram, to develop methodologies applicable on a broader scale. This study investigates narrative evolution on these platforms using quantitative corpus-based discourse analysis techniques. Our approach is a novel mode to study multiple social media domains to distil key information which may be obscured otherwise, allowing for useful and actionable insights. The paper details the technical and methodological aspects of gathering and preprocessing GAB and Telegram data for a keyness (Log Ratio) metric analysis, identifying crucial nouns and verbs for deeper exploration. Empirically, this approach is applied to a case study of a well defined event that had global impact: the 2023 Wagner mutiny. The main findings are: (1) the time line can be deconstructed to provide useful data features allowing for improved interpretation; (2) a methodology is applied which provides a basis for generalization. The key contribution is an approach, that in some cases, provides the ability to capture the dynamic narrative shifts over time with elevated confidence. The approach can augment near-real-time assessment of key social movements, allowing for informed governance choices. This research is important because it lays out a useful methodology for time series relevant info-culling, which can enable proactive modes for positive social engagement.

News captioning aims to describe an image with its news article body as input. It greatly relies on a set of detected named entities, including real-world people, organizations, and places. This paper exploits commonsense knowledge to understand named entities for news captioning. By ``understand'', we mean correlating the news content with common sense in the wild, which helps an agent to 1) distinguish semantically similar named entities and 2) describe named entities using words outside of training corpora. Our approach consists of three modules: (a) Filter Module aims to clarify the common sense concerning a named entity from two aspects: what does it mean? and what is it related to?, which divide the common sense into explanatory knowledge and relevant knowledge, respectively. (b) Distinguish Module aggregates explanatory knowledge from node-degree, dependency, and distinguish three aspects to distinguish semantically similar named entities. (c) Enrich Module attaches relevant knowledge to named entities to enrich the entity description by commonsense information (e.g., identity and social position). Finally, the probability distributions from both modules are integrated to generate the news captions. Extensive experiments on two challenging datasets (i.e., GoodNews and NYTimes) demonstrate the superiority of our method. Ablation studies and visualization further validate its effectiveness in understanding named entities.

While exposure to diverse viewpoints may reduce polarization, it can also have a backfire effect and exacerbate polarization when the discussion is adversarial. Here, we examine the question whether intergroup interactions around important events affect polarization between majority and minority groups in social networks. We compile data on the religious identity of nearly 700,000 Indian Twitter users engaging in COVID-19-related discourse during 2020. We introduce a new measure for an individual's group conformity based on contextualized embeddings of tweet text, which helps us assess polarization between religious groups. We then use a meta-learning framework to examine heterogeneous treatment effects of intergroup interactions on an individual's group conformity in the light of communal, political, and socio-economic events. We find that for political and social events, intergroup interactions reduce polarization. This decline is weaker for individuals at the extreme who already exhibit high conformity to their group. In contrast, during communal events, intergroup interactions can increase group conformity. Finally, we decompose the differential effects across religious groups in terms of emotions and topics of discussion. The results show that the dynamics of religious polarization are sensitive to the context and have important implications for understanding the role of intergroup interactions.

Large language models (LLMs) have demonstrated remarkable capabilities across various NLP tasks. However, their computational costs are prohibitively high. To address this issue, previous research has attempted to distill the knowledge of LLMs into smaller models by generating annotated data. Nonetheless, these works have mainly focused on the direct use of LLMs for text generation and labeling, without fully exploring their potential to comprehend the target task and acquire valuable knowledge. In this paper, we propose EvoKD: Evolving Knowledge Distillation, which leverages the concept of active learning to interactively enhance the process of data generation using large language models, simultaneously improving the task capabilities of small domain model (student model). Different from previous work, we actively analyze the student model's weaknesses, and then synthesize labeled samples based on the analysis. In addition, we provide iterative feedback to the LLMs regarding the student model's performance to continuously construct diversified and challenging samples. Experiments and analysis on different NLP tasks, namely, text classification and named entity recognition show the effectiveness of EvoKD.

Although there is mounting empirical evidence for the increase in affective polarization, few mechanistic models can explain its emergence at the population level. The question of how such a phenomenon can emerge from divergent opinions of a population on an ideological issue is still an open issue. In this paper, we establish that human normativity, that is, individual expression of normative opinions based on beliefs about the population, can lead to population-level polarization when ideological institutions distort beliefs in accordance with their objective of moving expressed opinion to one extreme. Using a game-theoretic model, we establish that individuals with more extreme opinions will have more extreme rhetoric and higher misperceptions about their outgroup members. Our model also shows that when social recommendation systems mediate institutional signals, we can observe the formation of different institutional communities, each with its unique community structure and characteristics. Using the model, we identify practical strategies platforms can implement, such as reducing exposure to signals from ideological institutions and a tailored approach to content moderation, both of which can rectify the affective polarization problem within its purview.

Despite the considerable advancements in English LLMs, the progress in building comparable models for other languages has been hindered due to the scarcity of tailored resources. Our work aims to bridge this divide by introducing an expansive suite of resources specifically designed for the development of Indic LLMs, covering 22 languages, containing a total of 251B tokens and 74.8M instruction-response pairs. Recognizing the importance of both data quality and quantity, our approach combines highly curated manually verified data, unverified yet valuable data, and synthetic data. We build a clean, open-source pipeline for curating pre-training data from diverse sources, including websites, PDFs, and videos, incorporating best practices for crawling, cleaning, flagging, and deduplication. For instruction-fine tuning, we amalgamate existing Indic datasets, translate/transliterate English datasets into Indian languages, and utilize LLaMa2 and Mixtral models to create conversations grounded in articles from Indian Wikipedia and Wikihow. Additionally, we address toxicity alignment by generating toxic prompts for multiple scenarios and then generate non-toxic responses by feeding these toxic prompts to an aligned LLaMa2 model. We hope that the datasets, tools, and resources released as a part of this work will not only propel the research and development of Indic LLMs but also establish an open-source blueprint for extending such efforts to other languages. The data and other artifacts created as part of this work are released with permissive licenses.

Large language models (LLMs) are initially trained on vast amounts of data, then fine-tuned using reinforcement learning from human feedback (RLHF); this also serves to teach the LLM to provide appropriate and safe responses. In this paper, we present a novel method to manipulate the fine-tuned version into reverting to its pre-RLHF behavior, effectively erasing the model's filters; the exploit currently works for GPT4, Claude Sonnet, and (to some extent) for Inflection-2.5. Unlike other jailbreaks (for example, the popular "Do Anything Now" (DAN) ), our method does not rely on instructing the LLM to override its RLHF policy; hence, simply modifying the RLHF process is unlikely to address it. Instead, we induce a hallucination involving reversed text during which the model reverts to a word bucket, effectively pausing the model's filter. We believe that our exploit presents a fundamental vulnerability in LLMs currently unaddressed, as well as an opportunity to better understand the inner workings of LLMs during hallucinations.