Links are a fundamental part of information networks, turning isolated pieces of knowledge into a network of information that is much richer than the sum of its parts. However, adding a new link to the network is not trivial: it requires not only the identification of a suitable pair of source and target entities but also the understanding of the content of the source to locate a suitable position for the link in the text. The latter problem has not been addressed effectively, particularly in the absence of text spans in the source that could serve as anchors to insert a link to the target entity. To bridge this gap, we introduce and operationalize the task of entity insertion in information networks. Focusing on the case of Wikipedia, we empirically show that this problem is, both, relevant and challenging for editors. We compile a benchmark dataset in 105 languages and develop a framework for entity insertion called LocEI (Localized Entity Insertion) and its multilingual variant XLocEI. We show that XLocEI outperforms all baseline models (including state-of-the-art prompt-based ranking with LLMs such as GPT-4) and that it can be applied in a zero-shot manner on languages not seen during training with minimal performance drop. These findings are important for applying entity insertion models in practice, e.g., to support editors in adding links across the more than 300 language versions of Wikipedia.

Given the prompt “Rome is in”, can we steer a language model to flip its prediction of an incorrect token “France” to a correct token “Italy” by only multiplying a few relevant activation vectors with scalars? We argue that successfully intervening on a model is a prerequisite for interpreting its internal workings. Concretely, we establish a three-term objective: a successful intervention should flip the correct with the wrong token and vice versa (effectiveness), and leave other tokens unaffected (faithfulness), all while being sparse (minimality). Using gradient-based optimization, this objective lets us learn (and later evaluate) a specific kind of efficient and interpretable intervention: activation scaling only modifies the signed magnitude of activation vectors to strengthen, weaken, or reverse the steering directions already encoded in the model. On synthetic tasks, this intervention performs comparably with steering vectors in terms of effectiveness and faithfulness, but is much more minimal allowing us to pinpoint interpretable model components. We evaluate activation scaling from different angles, compare performance on different datasets, and make activation scalars a learnable function of the activation vectors themselves to generalize to varying-length prompts.

The capabilities of large language models (LLMs) have raised concerns about their potential to create and propagate convincing narratives. Here, we study their performance in detecting convincing arguments to gain insights into LLMs’ persuasive capabilities without directly engaging in experimentation with humans. We extend a dataset by Durmus and Cardie (2018) with debates, votes, and user traits and propose tasks measuring LLMs’ ability to (1) distinguish between strong and weak arguments, (2) predict stances based on beliefs and demographic characteristics, and (3) determine the appeal of an argument to an individual based on their traits. We show that LLMs perform on par with humans in these tasks and that combining predictions from different LLMs yields significant performance gains, surpassing human performance. The data and code released with this paper contribute to the crucial effort of continuously evaluating and monitoring LLMs’ capabilities and potential impact. (https://go.epfl.ch/persuasion-llm)

A rapidly growing number of applications rely on a small set of closed-source language models (LMs). This dependency might introduce novel security risks if LMs develop self-recognition capabilities. Inspired by human identity verification methods, we propose a novel approach for assessing self-recognition in LMs using model-generated “security questions”. Our test can be externally administered to keep track of frontier models as it does not require access to internal model parameters or output probabilities. We use our test to examine self-recognition in ten of the most capable open- and closed-source LMs currently publicly available. Our extensive experiments found no empirical evidence of general or consistent self-recognition in any examined LM. Instead, our results suggest that given a set of alternatives, LMs seek to pick the “best” answer, regardless of its origin. Moreover, we find indications that preferences about which models produce the best answers are consistent across LMs. We additionally uncover novel insights on position bias considerations for LMs in multiple-choice settings.

Large language models (LLMs) have been shown to perform better when asked to reason step-by-step before answering a question. However, it is unclear to what degree the model’s final answer is faithful to the stated reasoning steps. In this paper, we perform a causal mediation analysis on twelve LLMs to examine how intermediate reasoning steps generated by the LLM influence the final outcome and find that LLMs do not reliably use their intermediate reasoning steps when generating an answer. To address this issue, we introduce FRODO, a framework to tailor small-sized LMs to generate correct reasoning steps and robustly reason over these steps. FRODO consists of an inference module that learns to generate correct reasoning steps using an implicit causal reward function and a reasoning module that learns to faithfully reason over these intermediate inferences using a counterfactual and causal preference objective. Our experiments show that FRODO significantly outperforms four competitive baselines. Furthermore, FRODO improves the robustness and generalization ability of the reasoning LM, yielding higher performance on out-of-distribution test sets. Finally, we find that FRODO’s rationales are more faithful to its final answer predictions than standard supervised fine-tuning.

Language models (LMs) have recently shown remarkable performance on reasoning tasks by explicitly generating intermediate inferences,e.g., chain-of-thought prompting. However, these intermediate inference steps may be inappropriate deductions from the initial contextand lead to incorrect final predictions. Here we introduce REFINER, a framework for finetuning LMs to explicitly generate intermediate reasoning steps while interacting with a critic model that provides automated feedback on the reasoning. Specifically, the critic provides structured feedback that the reasoning LM uses to iteratively improve its intermediate arguments. Empirical evaluations of REFINER on three diverse reasoning tasks show significant improvements over baseline LMs of comparable scale. Furthermore, when using GPT-3.5 or ChatGPT as the reasoner, the trained critic significantly improves reasoning without finetuning the reasoner. Finally, our critic model is trained without expensive human-in-the-loop data but can be substituted with humans at inference time.

Large language models (LLMs) have an impressive ability to draw on novel information supplied in their context. Yet the mechanisms underlying this contextual grounding remain unknown, especially in situations where contextual information contradicts factual knowledge stored in the parameters, which LLMs also excel at recalling. Favoring the contextual information is critical for retrieval-augmented generation methods, which enrich the context with up-to-date information, hoping that grounding can rectify outdated or noisy stored knowledge. We present a novel method to study grounding abilities using Fakepedia, a novel dataset of counterfactual texts constructed to clash with a model’s internal parametric knowledge. In this study, we introduce Fakepedia, a counterfactual dataset designed to evaluate grounding abilities when the internal parametric knowledge clashes with the contextual information. We benchmark various LLMs with Fakepedia and conduct a causal mediation analysis of LLM components when answering Fakepedia queries, based on our Masked Grouped Causal Tracing (MGCT) method. Through this analysis, we identify distinct computational patterns between grounded and ungrounded responses. We finally demonstrate that distinguishing grounded from ungrounded responses is achievable through computational analysis alone. Our results, together with existing findings about factual recall mechanisms, provide a coherent narrative of how grounding and factual recall mechanisms interact within LLMs.

We ask whether multilingual language models trained on unbalanced, English-dominated corpora use English as an internal pivot language—-a question of key importance for understanding how language models function and the origins of linguistic bias. Focusing on the Llama-2 family of transformer models, our study is based on carefully constructed non-English prompts with a unique correct single-token continuation. From layer to layer, transformers gradually map an input embedding of the final prompt token to an output embedding from which next-token probabilities are computed. Tracking intermediate embeddings through their high-dimensional space reveals three distinct phases, whereby intermediate embeddings (1) start far away from output token embeddings; (2) already in middle layers allow for decoding a semantically correct next token, but giving higher probability to its version in English than in the input language; (3) move into an input-language-specific region of the embedding space. We cast these results into a conceptual model where the three phases operate in ”input space”, ”concept space”, and ”output space”, respectively. Crucially, our evidence suggests that the abstract ”concept space” lies closer to English than to other input languages, which may have important consequences regarding the biases embodied by multilingual language models.

Constrained decoding, a technique for enforcing constraints on language model outputs, offers a way to control text generation without retraining or architectural modifications. Its application is, however, typically restricted to models that give users access to next-token distributions (usually via softmax logits), which poses a limitation with blackbox large language models (LLMs). This paper introduces sketch-guided constrained decoding (SketchGCD), a novel approach to constrained decoding for blackbox LLMs, which operates without access to the logits of the blackbox LLM. SketchGCD utilizes a locally hosted auxiliary model to refine the output of an unconstrained blackbox LLM, effectively treating this initial output as a “sketch” for further elaboration. This approach is complementary to traditional logit-based techniques and enables the application of constrained decoding in settings where full model transparency is unavailable. We demonstrate the efficacy of SketchGCD through experiments in closed information extraction and constituency parsing, showing how it enhances the utility and flexibility of blackbox LLMs for complex NLP tasks.

Humor is a fundamental facet of human cognition and interaction. Yet, despite recent advances in natural language processing, humor detection remains a challenging task that is complicated by the scarcity of datasets that pair humorous texts with similar non-humorous counterparts. We investigate whether large language models (LLMs) can generate synthetic data for humor detection via editing texts. We benchmark LLMs on an existing human dataset and show that current LLMs display an impressive ability to “unfun” jokes, as judged by humans and as measured on the downstream task of humor detection. We extend our approach to a code-mixed English-Hindi humor dataset where we find that GPT-4’s synthetic data is highly rated by bilingual annotators and provides challenging adversarial examples for humor classifiers.

Measuring the intensity of events is crucial for monitoring and tracking armed conflict. Advances in automated event extraction have yielded massive data sets of “who did what to whom” micro-records that enable data-driven approaches to monitoring conflict. The Goldstein scale is a widely-used expert-based measure that scores events on a conflictual–cooperative scale. It is based only on the action category (“what”) and disregards the subject (“who”) and object (“to whom”) of an event, as well as contextual information, like associated casualty count, that should contribute to the perception of an event’s “intensity”. This paper takes a latent variable-based approach to measuring conflict intensity. We introduce a probabilistic generative model that assumes each observed event is associated with a latent intensity class. A novel aspect of this model is that it imposes an ordering on the classes, such that higher-valued classes denote higher levels of intensity. The ordinal nature of the latent variable is induced from naturally ordered aspects of the data (e.g., casualty counts) where higher values naturally indicate higher intensity. We evaluate the proposed model both intrinsically and extrinsically, showing that it obtains comparatively good held-out predictive performance.

Large language models (LLMs) have great potential for synthetic data generation. This work shows that useful data can be synthetically generated even for tasks that cannot be solved directly by LLMs: for problems with structured outputs, it is possible to prompt an LLM to perform the task in the reverse direction, by generating plausible input text for a target output structure. Leveraging this asymmetry in task difficulty makes it possible to produce large-scale, high-quality data for complex tasks. We demonstrate the effectiveness of this approach on closed information extraction, where collecting ground-truth data is challenging, and no satisfactory dataset exists to date. We synthetically generate a dataset of 1.8M data points, establish its superior quality compared to existing datasets in a human evaluation, and use it to finetune small models (220M and 770M parameters), termed SynthIE, that outperform the prior state of the art (with equal model size) by a substantial margin of 57 absolute points in micro-F1 and 79 points in macro-F1. Code, data, and models are available at anonymous.

Despite their impressive performance, large language models (LMs) still struggle with reliably generating complex output structures when not finetuned to follow the required output format exactly. To address this issue, grammar-constrained decoding (GCD) can be used to control the generation of LMs, guaranteeing that the output follows a given structure. Most existing GCD methods are, however, limited to specific tasks, such as parsing or code generation. In this work, we demonstrate that formal grammars can describe the output space for a much wider range of tasks and argue that GCD can serve as a unified framework for structured NLP tasks in general. For increased flexibility, we introduce input-dependent grammars, which allow the grammar to depend on the input and thus enable the generation of different output structures for different inputs. We then empirically demonstrate the power and flexibility of GCD-enhanced LMs on (1) information extraction, (2) entity disambiguation, and (3) constituency parsing. Our results indicate that grammar-constrained LMs substantially outperform unconstrained LMs or even beat task-specific finetuned models. Grammar constraints thus hold great promise for harnessing off-the-shelf LMs for a wide range of structured NLP tasks, especially where training data is scarce or finetuning is expensive. Code and data: https://github.com/epfl-dlab/GCD.

Structured and grounded representation of text is typically formalized by closed information extraction, the problem of extracting an exhaustive set of (subject, relation, object) triplets that are consistent with a predefined set of entities and relations from a knowledge base schema. Most existing works are pipelines prone to error accumulation, and all approaches are only applicable to unrealistically small numbers of entities and relations. We introduce GenIE (generative information extraction), the first end-to-end autoregressive formulation of closed information extraction. GenIE naturally exploits the language knowledge from the pre-trained transformer by autoregressively generating relations and entities in textual form. Thanks to a new bi-level constrained generation strategy, only triplets consistent with the predefined knowledge base schema are produced. Our experiments show that GenIE is state-of-the-art on closed information extraction, generalizes from fewer training data points than baselines, and scales to a previously unmanageable number of entities and relations. With this work, closed information extraction becomes practical in realistic scenarios, providing new opportunities for downstream tasks. Finally, this work paves the way towards a unified end-to-end approach to the core tasks of information extraction.

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.

Modern pretrained language models are critical components of NLP pipelines. Yet, they suffer from spurious correlations, poor out-of-domain generalization, and biases.Inspired by recent progress in causal machine learning, in particular the invariant risk minimization (IRM) paradigm, we propose invariant language modeling, a framework for learning invariant representations that generalize better across multiple environments. In particular, we adapt a game-theoretic implementation of IRM (IRM-games) to language models, where the invariance emerges from a specific training schedule in which all the environments compete to optimize their own environment-specific loss by updating subsets of the model in a round-robin fashion.We focused on controlled experiments to precisely demonstrate the ability of our method to (i) remove structured noise, (ii) ignore specific spurious correlations without affecting global performance, and (iii) achieve better out-of-domain generalization.These benefits come with a negligible computational overhead compared to standard training, do not require changing the local loss, and can be applied to any language model. We believe this framework is promising to help mitigate spurious correlations and biases in language models.

Candidate generation is a crucial module in entity linking. It also plays a key role in multiple NLP tasks that have been proven to beneficially leverage knowledge bases. Nevertheless, it has often been overlooked in the monolingual English entity linking literature, as naïve approaches obtain very good performance. Unfortunately, the existing approaches for English cannot be successfully transferred to poorly resourced languages. This paper constitutes an in-depth analysis of the candidate generation problem in the context of cross-lingual entity linking with a focus on low-resource languages. Among other contributions, we point out limitations in the evaluation conducted in previous works. We introduce a characterization of queries into types based on their difficulty, which improves the interpretability of the performance of different methods. We also propose a light-weight and simple solution based on the construction of indexes whose design is motivated by more complex transfer learning based neural approaches. A thorough empirical analysis on 9 real-world datasets under 2 evaluation settings shows that our simple solution outperforms the state-of-the-art approach in terms of both quality and efficiency for almost all datasets and query types.

Evaluation in NLP is usually done by comparing the scores of competing systems independently averaged over a common set of test instances. In this work, we question the use of averages for aggregating evaluation scores into a final number used to decide which system is best, since the average, as well as alternatives such as the median, ignores the pairing arising from the fact that systems are evaluated on the same test instances. We illustrate the importance of taking the instancelevel pairing of evaluation scores into account and demonstrate, both theoretically and empirically, the advantages of aggregation methods based on pairwise comparisons, such as the Bradley–Terry (BT) model, a mechanism based on the estimated probability that a given system scores better than another on the test set. By re-evaluating 296 real NLP evaluation setups across four tasks and 18 evaluation metrics, we show that the choice of aggregation mechanism matters and yields different conclusions as to which systems are state of the art in about 30% of the setups. To facilitate the adoption of pairwise evaluation, we release a practical tool for performing the full analysis of evaluation scores with the mean, median, BT, and two variants of BT (Elo and TrueSkill), alongside functionality for appropriate statistical testing.

Understanding the origins of militarized conflict is a complex, yet important undertaking. Existing research seeks to build this understanding by considering bi-lateral relationships between entity pairs (dyadic causes) and multi-lateral relationships among multiple entities (systemic causes). The aim of this work is to compare these two causes in terms of how they correlate with conflict between two entities. We do this by devising a set of textual and graph-based features which represent each of the causes. The features are extracted from Wikipedia and modeled as a large graph. Nodes in this graph represent entities connected by labeled edges representing ally or enemy-relationships. This allows casting the problem as an edge classification task, which we term dyad classification. We propose and evaluate classifiers to determine if a particular pair of entities are allies or enemies. Our results suggest that our systemic features might be slightly better correlates of conflict. Further, we find that Wikipedia articles of allies are semantically more similar than enemies.

Entity linking is an important problem with many applications. Most previous solutions were designed for settings where annotated training data is available, which is, however, not the case in numerous domains. We propose a light-weight and scalable entity linking method, Eigenthemes, that relies solely on the availability of entity names and a referent knowledge base. Eigenthemes exploits the fact that the entities that are truly mentioned in a document (the “gold entities”) tend to form a semantically dense subset of the set of all candidate entities in the document. Geometrically speaking, when representing entities as vectors via some given embedding, the gold entities tend to lie in a low-rank subspace of the full embedding space. Eigenthemes identifies this subspace using the singular value decomposition and scores candidate entities according to their proximity to the subspace. On the empirical front, we introduce multiple strong baselines that compare favorably to (and sometimes even outperform) the existing state of the art. Extensive experiments on benchmark datasets from a variety of real-world domains showcase the effectiveness of our approach.

Evaluation of cross-lingual encoders is usually performed either via zero-shot cross-lingual transfer in supervised downstream tasks or via unsupervised cross-lingual textual similarity. In this paper, we concern ourselves with reference-free machine translation (MT) evaluation where we directly compare source texts to (sometimes low-quality) system translations, which represents a natural adversarial setup for multilingual encoders. Reference-free evaluation holds the promise of web-scale comparison of MT systems. We systematically investigate a range of metrics based on state-of-the-art cross-lingual semantic representations obtained with pretrained M-BERT and LASER. We find that they perform poorly as semantic encoders for reference-free MT evaluation and identify their two key limitations, namely, (a) a semantic mismatch between representations of mutual translations and, more prominently, (b) the inability to punish “translationese”, i.e., low-quality literal translations. We propose two partial remedies: (1) post-hoc re-alignment of the vector spaces and (2) coupling of semantic-similarity based metrics with target-side language modeling. In segment-level MT evaluation, our best metric surpasses reference-based BLEU by 5.7 correlation points.

The goal of text summarization is to compress documents to the relevant information while excluding background information already known to the receiver. So far, summarization researchers have given considerably more attention to relevance than to background knowledge. In contrast, this work puts background knowledge in the foreground. Building on the realization that the choices made by human summarizers and annotators contain implicit information about their background knowledge, we develop and compare techniques for inferring background knowledge from summarization data. Based on this framework, we define summary scoring functions that explicitly model background knowledge, and show that these scoring functions fit human judgments significantly better than baselines. We illustrate some of the many potential applications of our framework. First, we provide insights into human information importance priors. Second, we demonstrate that averaging the background knowledge of multiple, potentially biased annotators or corpora greatly improves summaryscoring performance. Finally, we discuss potential applications of our framework beyond summarization.

We propose a new method to detect when users express the intent to leave a service, also known as churn. While previous work focuses solely on social media, we show that this intent can be detected in chatbot conversations. As companies increasingly rely on chatbots they need an overview of potentially churny users. To this end, we crowdsource and publish a dataset of churn intent expressions in chatbot interactions in German and English. We show that classifiers trained on social media data can detect the same intent in the context of chatbots. We introduce a classification architecture that outperforms existing work on churn intent detection in social media. Moreover, we show that, using bilingual word embeddings, a system trained on combined English and German data outperforms monolingual approaches. As the only existing dataset is in English, we crowdsource and publish a novel dataset of German tweets. We thus underline the universal aspect of the problem, as examples of churn intent in English help us identify churn in German tweets and chatbot conversations.

Person-to-person evaluations are prevalent in all kinds of discourse and important for establishing reputations, building social bonds, and shaping public opinion. Such evaluations can be analyzed separately using signed social networks and textual sentiment analysis, but this misses the rich interactions between language and social context. To capture such interactions, we develop a model that predicts individual A’s opinion of individual B by synthesizing information from the signed social network in which A and B are embedded with sentiment analysis of the evaluative texts relating A to B. We prove that this problem is NP-hard but can be relaxed to an efficiently solvable hinge-loss Markov random field, and we show that this implementation outperforms text-only and network-only versions in two very different datasets involving community-level decision-making: the Wikipedia Requests for Adminship corpus and the Convote U.S. Congressional speech corpus.