Despite the success of Transformer-based language models in a wide variety of natural language processing tasks, our understanding of how these models process a given input in order to represent task-relevant information remains incomplete. In this work, we focus on semantic composition and examine how Transformer-based language models represent semantic information related to the meaning of English noun-noun compounds. We probe Transformer-based language models for their knowledge of the thematic relations that link the head nouns and modifier words of compounds (e.g., KITCHEN CHAIR: a chair located in a kitchen). Firstly, using a dataset featuring groups of compounds with shared lexical or semantic features, we find that token representations of six Transformer-based language models distinguish between pairs of compounds based on whether they use the same thematic relation. Secondly, we utilize fine-grained vector representations of compound semantics derived from human annotations, and find that token vectors from several models elicit a strong signal of the semantic relations used in the compounds. In a novel “compositional probe” setting, where we compare the semantic relation signal in mean-pooled token vectors of compounds to mean-pooled token vectors when the two constituent words appear in separate sentences, we find that the Transformer-based language models that best represent the semantics of noun-noun compounds also do so substantially better than in the control condition where the two constituent works are processed separately. Overall, our results shed light on the ability of Transformer-based language models to support compositional semantic processes in representing the meaning of noun-noun compounds.

In this paper, we perform a systematic analysis of how closely the intermediate layers from LSTM and trans former language models correspond to human semantic knowledge. Furthermore, in order to make more meaningful comparisons with theories of human language comprehension in psycholinguistics, we focus on two key stages where the meaning of a particular target word may arise: immediately before the word’s presentation to the model (comparable to forward inferencing), and immediately after the word token has been input into the network. Our results indicate that the transformer models are better at capturing semantic knowledge relating to lexical concepts, both during word prediction and when retention is required.

Researchers have recently demonstrated that tying the neural weights between the input look-up table and the output classification layer can improve training and lower perplexity on sequence learning tasks such as language modelling. Such a procedure is possible due to the design of the softmax classification layer, which previous work has shown to comprise a viable set of semantic representations for the model vocabulary, and these these output embeddings are known to perform well on word similarity benchmarks. In this paper, we make meaningful comparisons between the input and output embeddings and other SOTA distributional models to gain a better understanding of the types of information they represent. We also construct a new set of word embeddings using the output embeddings to create locally-optimal approximations for the intermediate representations from the language model. These locally-optimal embeddings demonstrate excellent performance across all our evaluations.

Semantic models derived from visual information have helped to overcome some of the limitations of solely text-based distributional semantic models. Researchers have demonstrated that text and image-based representations encode complementary semantic information, which when combined provide a more complete representation of word meaning, in particular when compared with data on human conceptual knowledge. In this work, we reveal that these vision-based representations, whilst quite effective, do not make use of all the semantic information available in the neural network that could be used to inform vector-based models of semantic representation. Instead, we build image-based meta-embeddings from computer vision models, which can incorporate information from all layers of the network, and show that they encode a richer set of semantic attributes and yield a more complete representation of human conceptual knowledge.

While there is a rich literature on the tracking of sentiment and emotion in texts, modelling the emotional trajectory of longer narratives, such as literary texts, poses new challenges. Previous work in the area of sentiment analysis has focused on using information from within a sentence to predict a valence value for that sentence. We propose to explore the influence of previous sentences on the sentiment of a given sentence. In particular, we investigate whether information present in a history of previous sentences can be used to predict a valence value for the following sentence. We explored both linear and non-linear models applied with a range of different feature combinations. We also looked at different context history sizes to determine what range of previous sentence context was the most informative for our models. We establish a linear relationship between sentence context history and the valence value of the current sentence and demonstrate that sentences in closer proximity to the target sentence are more informative. We show that the inclusion of semantic word embeddings further enriches our model predictions.

Despite recent advances in the application of deep neural networks to various kinds of medical data, extracting information from unstructured textual sources remains a challenging task. The challenges of training and interpreting document classification models are amplified when dealing with small and highly technical datasets, as are common in the clinical domain. Using a dataset of de-identified clinical letters gathered at a memory clinic, we construct several recurrent neural network models for letter classification, and evaluate them on their ability to build meaningful representations of the documents and predict patients’ diagnoses. Additionally, we probe sentence embedding models in order to build a human-interpretable representation of the neural network’s features, using a simple and intuitive technique based on perturbative approaches to sentence importance. In addition to showing which sentences in a document are most informative about the patient’s condition, this method reveals the types of sentences that lead the model to make incorrect diagnoses. Furthermore, we identify clusters of sentences in the embedding space that correlate strongly with importance scores for each clinical diagnosis class.

Feature norm datasets of human conceptual knowledge, collected in surveys of human volunteers, yield highly interpretable models of word meaning and play an important role in neurolinguistic research on semantic cognition. However, these datasets are limited in size due to practical obstacles associated with exhaustively listing properties for a large number of words. In contrast, the development of distributional modelling techniques and the availability of vast text corpora have allowed researchers to construct effective vector space models of word meaning over large lexicons. However, this comes at the cost of interpretable, human-like information about word meaning. We propose a method for mapping human property knowledge onto a distributional semantic space, which adapts the word2vec architecture to the task of modelling concept features. Our approach gives a measure of concept and feature affinity in a single semantic space, which makes for easy and efficient ranking of candidate human-derived semantic properties for arbitrary words. We compare our model with a previous approach, and show that it performs better on several evaluation tasks. Finally, we discuss how our method could be used to develop efficient sampling techniques to extend existing feature norm datasets in a reliable way.

Distributional models provide a convenient way to model semantics using dense embedding spaces derived from unsupervised learning algorithms. However, the dimensions of dense embedding spaces are not designed to resemble human semantic knowledge. Moreover, embeddings are often built from a single source of information (typically text data), even though neurocognitive research suggests that semantics is deeply linked to both language and perception. In this paper, we combine multimodal information from both text and image-based representations derived from state-of-the-art distributional models to produce sparse, interpretable vectors using Joint Non-Negative Sparse Embedding. Through in-depth analyses comparing these sparse models to human-derived behavioural and neuroimaging data, we demonstrate their ability to predict interpretable linguistic descriptions of human ground-truth semantic knowledge.

Performance in language modelling has been significantly improved by training recurrent neural networks on large corpora. This progress has come at the cost of interpretability and an understanding of how these architectures function, making principled development of better language models more difficult. We look inside a state-of-the-art neural language model to analyse how this model represents high-level lexico-semantic information. In particular, we investigate how the model represents words by extracting activation patterns where they occur in the text, and compare these representations directly to human semantic knowledge.

Investigating differences in linguistic usage between individuals who have suffered brain injury (hereafter patients) and those who haven’t can yield a number of benefits. It provides a better understanding about the precise way in which impairments affect patients’ language, improves theories of how the brain processes language, and offers heuristics for diagnosing certain types of brain damage based on patients’ speech. One method for investigating usage differences involves the analysis of spontaneous speech. In the work described here we construct a text corpus consisting of transcripts of individuals’ speech produced during two tasks: the Boston-cookie-theft picture description task (Goodglass and Kaplan, 1983) and a spontaneous speech task, which elicits a semi-prompted monologue, and/or free speech. Interviews with patients from 19yrs to 89yrs were transcribed, as were interviews with a comparable number of healthy individuals (20yrs to 89yrs). Structural brain images are available for approximately 30% of participants. This unique data source provides a rich resource for future research in many areas of language impairment and has been constructed to facilitate analysis with natural language processing and corpus linguistics techniques.