Post-hoc explanation methods are an important tool for increasing model transparency for users. Unfortunately, the currently used methods for attributing token importance often yield diverging patterns. In this work, we study potential sources of disagreement across methods from a linguistic perspective. We find that different methods systematically select different classes of words and that methods that agree most with other methods and with humans display similar linguistic preferences. Token-level differences between methods are smoothed out if we compare them on the syntactic span level. We also find higher agreement across methods by estimating the most important spans dynamically instead of relying on a fixed subset of size k. We systematically investigate the interaction between k and spans and propose an improved configuration for selecting important tokens.

Subword tokenization has become the de-facto standard for tokenization although comparative evaluations of their quality across languages are scarce. Existing evaluation studies focus on the effect of a tokenization algorithm on the performance in downstream tasks, or on engineering criteria such as the compression rate. We present a new evaluation paradigm that focuses on the cognitive plausibility of subword tokenization. We analyze the correlation of the tokenizer output with the reading time and accuracy of human responses on a lexical decision task. We compare three tokenization algorithms across several languages and vocabulary sizes. Our results indicate that the Unigram algorithm yields less cognitively plausible tokenization behavior and a worse coverage of derivational morphemes, in contrast with prior work.

Feature attribution scores are used for explaining the prediction of a text classifier to users by highlighting a k number of tokens. In this work, we propose a way to determine the number of optimal k tokens that should be displayed from sequential properties of the attribution scores. Our approach is dynamic across sentences, method-agnostic, and deals with sentence length bias. We compare agreement between multiple methods and humans on an NLI task, using fixed k and dynamic k. We find that perturbation-based methods and Vanilla Gradient exhibit highest agreement on most method–method and method–human agreement metrics with a static k. Their advantage over other methods disappears with dynamic ks which mainly improve Integrated Gradient and GradientXInput. To our knowledge, this is the first evidence that sequential properties of attribution scores are informative for consolidating attribution signals for human interpretation.

An inclusive society needs to facilitate access to information for all of its members, including citizens with low literacy and with non-native language skills. We present an approach to assess Dutch text complexity on the sentence level and conduct an interpretability analysis to explore the link between neural models and linguistic complexity features. Building on these findings, we develop the first contextual lexical simplification model for Dutch and publish a pilot dataset for evaluation. We go beyondprevious work which primarily targeted lexical substitution and propose strategies for adjusting the model’s linguistic register to generate simpler candidates. Our results indicate that continual pre-training and multi-task learning with conceptually related tasks are promising directions for ensuring the simplicity of the generated substitutions.

When humans read a text, their eye movements are influenced by the structural complexity of the input sentences. This cognitive phenomenon holds across languages and recent studies indicate that multilingual language models utilize structural similarities between languages to facilitate cross-lingual transfer. We use sentence-level eye-tracking patterns as a cognitive indicator for structural complexity and show that the multilingual model XLM-RoBERTa can successfully predict varied patterns for 13 typologically diverse languages, despite being fine-tuned only on English data. We quantify the sensitivity of the model to structural complexity and distinguish a range of complexity characteristics. Our results indicate that the model develops a meaningful bias towards sentence length but also integrates cross-lingual differences. We conduct a control experiment with randomized word order and find that the model seems to additionally capture more complex structural information.

WordNet is a database that represents relations between words and concepts as an abstraction of the contexts in which words are used. Contextualized language models represent words in contexts but leave the underlying concepts implicit. In this paper, we investigate how different layers of a pre-trained language model shape the abstract lexical relationship toward the actual contextual concept. Can we define the amount of contextualized concept forming needed given the abstracted representation of a word? Specifically, we consider samples of words with different polysemy profiles shared across three languages, assuming that words with a different polysemy profile require a different degree of concept shaping by context. We conduct probing experiments to investigate the impact of prior polysemy profiles on the representation in different layers. We analyze how contextualized models can approximate meaning through context and examine crosslingual interference effects.

Argument Unit Recognition and Classification aims at identifying argument units from text and classifying them as pro or against. One of the design choices that need to be made when developing systems for this task is what the unit of classification should be: segments of tokens or full sentences. Previous research suggests that fine-tuning language models on the token-level yields more robust results for classifying sentences compared to training on sentences directly. We reproduce the study that originally made this claim and further investigate what exactly token-based systems learned better compared to sentence-based ones. We develop systematic tests for analysing the behavioural differences between the token-based and the sentence-based system. Our results show that token-based models are generally more robust than sentence-based models both on manually perturbed examples and on specific subpopulations of the data.

It has been shown that multilingual transformer models are able to predict human reading behavior when fine-tuned on small amounts of eye tracking data. As the cumulated prediction results do not provide insights into the linguistic cues that the model acquires to predict reading behavior, we conduct a deeper analysis of the predictions from the perspective of readability. We try to disentangle the three-fold relationship between human eye movements, the capability of language models to predict these eye movement patterns, and sentence-level readability measures for English. We compare a range of model configurations to multiple baselines. We show that the models exhibit difficulties with function words and that pre-training only provides limited advantages for linguistic generalization.

This paper presents our system used to participate in task 11 (MultiCONER) of the SemEval 2022 competition. Our system ranked fourth place in track 12 (Multilingual) and fifth place in track 13 (Code-Mixed). The goal of track 12 is to detect complex named entities in a multilingual setting, while track 13 is dedicated to detecting complex named entities in a code-mixed setting. Both systems were developed using transformer-based language models. We used an ensemble of XLM-RoBERTa-large and Microsoft/infoxlm-large with a Conditional Random Field (CRF) layer. In addition, we describe the algorithms employed to train our models and our hyper-parameter selection. We furthermore study the impact of different methods to aggregate the outputs of the individual models that compose our ensemble. Finally, we present an extensive analysis of the results and errors.

Relative word importance is a key metric for natural language processing. In this work, we compare human and model relative word importance to investigate if pretrained neural language models focus on the same words as humans cross-lingually. We perform an extensive study using several importance metrics (gradient-based saliency and attention-based) in monolingual and multilingual models, including eye-tracking corpora from four languages (German, Dutch, English, and Russian). We find that gradient-based saliency, first-layer attention, and attention flow correlate strongly with human eye-tracking data across all four languages. We further analyze the role of word length and word frequency in determining relative importance and find that it strongly correlates with length and frequency, however, the mechanisms behind these non-linear relations remain elusive. We obtain a cross-lingual approximation of the similarity between human and computational language processing and insights into the usability of several importance metrics.

Determining the relative importance of the elements in a sentence is a key factor for effortless natural language understanding. For human language processing, we can approximate patterns of relative importance by measuring reading fixations using eye-tracking technology. In neural language models, gradient-based saliency methods indicate the relative importance of a token for the target objective. In this work, we compare patterns of relative importance in English language processing by humans and models and analyze the underlying linguistic patterns. We find that human processing patterns in English correlate strongly with saliency-based importance in language models and not with attention-based importance. Our results indicate that saliency could be a cognitively more plausible metric for interpreting neural language models. The code is available on github: https://github.com/beinborn/relative_importance.

We analyze if large language models are able to predict patterns of human reading behavior. We compare the performance of language-specific and multilingual pretrained transformer models to predict reading time measures reflecting natural human sentence processing on Dutch, English, German, and Russian texts. This results in accurate models of human reading behavior, which indicates that transformer models implicitly encode relative importance in language in a way that is comparable to human processing mechanisms. We find that BERT and XLM models successfully predict a range of eye tracking features. In a series of experiments, we analyze the cross-domain and cross-language abilities of these models and show how they reflect human sentence processing.

Multilingual representations have mostly been evaluated based on their performance on specific tasks. In this article, we look beyond engineering goals and analyze the relations between languages in computational representations. We introduce a methodology for comparing languages based on their organization of semantic concepts. We propose to conduct an adapted version of representational similarity analysis of a selected set of concepts in computational multilingual representations. Using this analysis method, we can reconstruct a phylogenetic tree that closely resembles those assumed by linguistic experts. These results indicate that multilingual distributional representations that are only trained on monolingual text and bilingual dictionaries preserve relations between languages without the need for any etymological information. In addition, we propose a measure to identify semantic drift between language families. We perform experiments on word-based and sentence-based multilingual models and provide both quantitative results and qualitative examples. Analyses of semantic drift in multilingual representations can serve two purposes: They can indicate unwanted characteristics of the computational models and they provide a quantitative means to study linguistic phenomena across languages.

NLP models are imperfect and lack intricate capabilities that humans access automatically when processing speech or reading a text. Human language processing data can be leveraged to increase the performance of models and to pursue explanatory research for a better understanding of the differences between human and machine language processing. We review recent studies leveraging different types of cognitive processing signals, namely eye-tracking, M/EEG and fMRI data recorded during language understanding. We discuss the role of cognitive data for machine learning-based NLP methods and identify fundamental challenges for processing pipelines. Finally, we propose practical strategies for using these types of cognitive signals to enhance NLP models.

When speakers describe an image, they tend to look at objects before mentioning them. In this paper, we investigate such sequential cross-modal alignment by modelling the image description generation process computationally. We take as our starting point a state-of-the-art image captioning system and develop several model variants that exploit information from human gaze patterns recorded during language production. In particular, we propose the first approach to image description generation where visual processing is modelled sequentially. Our experiments and analyses confirm that better descriptions can be obtained by exploiting gaze-driven attention and shed light on human cognitive processes by comparing different ways of aligning the gaze modality with language production. We find that processing gaze data sequentially leads to descriptions that are better aligned to those produced by speakers, more diverse, and more natural—particularly when gaze is encoded with a dedicated recurrent component.

We probe the layers in multilingual BERT (mBERT) for phylogenetic and geographic language signals across 100 languages and compute language distances based on the mBERT representations. We 1) employ the language distances to infer and evaluate language trees, finding that they are close to the reference family tree in terms of quartet tree distance, 2) perform distance matrix regression analysis, finding that the language distances can be best explained by phylogenetic and worst by structural factors and 3) present a novel measure for measuring diachronic meaning stability (based on cross-lingual representation variability) which correlates significantly with published ranked lists based on linguistic approaches. Our results contribute to the nascent field of typological interpretability of cross-lingual text representations.

In this paper, we define and apply representational stability analysis (ReStA), an intuitive way of analyzing neural language models. ReStA is a variant of the popular representational similarity analysis (RSA) in cognitive neuroscience. While RSA can be used to compare representations in models, model components, and human brains, ReStA compares instances of the same model, while systematically varying single model parameter. Using ReStA, we study four recent and successful neural language models, and evaluate how sensitive their internal representations are to the amount of prior context. Using RSA, we perform a systematic study of how similar the representational spaces in the first and second (or higher) layers of these models are to each other and to patterns of activation in the human brain. Our results reveal surprisingly strong differences between language models, and give insights into where the deep linguistic processing, that integrates information over multiple sentences, is happening in these models. The combination of ReStA and RSA on models and brains allows us to start addressing the important question of what kind of linguistic processes we can hope to observe in fMRI brain imaging data. In particular, our results suggest that the data on story reading from Wehbe et al./ (2014) contains a signal of shallow linguistic processing, but show no evidence on the more interesting deep linguistic processing.

This survey discusses how recent developments in multimodal processing facilitate conceptual grounding of language. We categorize the information flow in multimodal processing with respect to cognitive models of human information processing and analyze different methods for combining multimodal representations. Based on this methodological inventory, we discuss the benefit of multimodal grounding for a variety of language processing tasks and the challenges that arise. We particularly focus on multimodal grounding of verbs which play a crucial role for the compositional power of language.

We explore a domain-agnostic approach for analyzing speech with the goal of opinion prediction. We represent the speech signal by mel-frequency cepstral coefficients and apply long short-term memory neural networks to automatically learn temporal regularities in speech. In contrast to previous work, our approach does not require complex feature engineering and works without textual transcripts. As a consequence, it can easily be applied on various speech analysis tasks for different languages and the results show that it can nevertheless be competitive to the state-of-the-art in opinion prediction. In a detailed error analysis for opinion mining we find that our approach performs well in identifying speaker-specific characteristics, but should be combined with additional information if subtle differences in the linguistic content need to be identified.

Language proficiency tests are used to evaluate and compare the progress of language learners. We present an approach for automatic difficulty prediction of C-tests that performs on par with human experts. On the basis of detailed analysis of newly collected data, we develop a model for C-test difficulty introducing four dimensions: solution difficulty, candidate ambiguity, inter-gap dependency, and paragraph difficulty. We show that cues from all four dimensions contribute to C-test difficulty.