Distributional Semantic Models (DSMs) construct vector representations of word meanings based on their contexts. Typically, the contexts of a word are defined as its closest neighbours, but they can also be retrieved from its syntactic dependency relations. In this work, we propose a new dependencybased DSM. The novelty of our model lies in associating an independent meaning representation, a matrix, with each dependency-label. This allows it to capture specifics of the relations between words and contexts, leading to good performance on both intrinsic and extrinsic evaluation tasks. In addition to that, our model has an inherent ability to represent dependency chains as products of matrices which provides a straightforward way of handling further contexts of a word.

Distributional semantic models have become a mainstay in NLP, providing useful features for downstream tasks. However, assessing long-term progress requires explicit long-term goals. In this paper, I take a broad linguistic perspective, looking at how well current models can deal with various semantic challenges. Given stark differences between models proposed in different subfields, a broad perspective is needed to see how we could integrate them. I conclude that, while linguistic insights can guide the design of model architectures, future progress will require balancing the often conflicting demands of linguistic expressiveness and computational tractability.

Many approaches to sentiment analysis rely on a lexicon that labels words with a prior polarity. This is particularly true for languages other than English, where labelled training data is not easily available. Existing efforts to produce such lexicons exist, and to avoid duplicated effort, a principled way to combine multiple resources is required. In this paper, we introduce a Bayesian probabilistic model, which can simultaneously combine polarity scores from several data sources and estimate the quality of each source. We apply this algorithm to a set of four German sentiment lexicons, to produce the SentiMerge lexicon, which we make publically available. In a simple classification task, we show that this lexicon outperforms each of the underlying resources, as well as a majority vote model.

We propose a method for natural language generation, choosing the most representative output rather than the most likely output. By viewing the language generation process from the voting theory perspective, we define representativeness using range voting and a similarity measure. The proposed method can be applied when generating from any probabilistic language model, including n-gram models and neural network models. We evaluate different similarity measures on an image captioning task and a machine translation task, and show that our method generates longer and more diverse sentences, providing a solution to the common problem of short outputs being preferred over longer and more informative ones. The generated sentences obtain higher BLEU scores, particularly when the beam size is large. We also perform a human evaluation on both tasks and find that the outputs generated using our method are rated higher.

Word embeddings are an essential component in a wide range of natural language processing applications. However, distributional semantic models are known to struggle when only a small number of context sentences are available. Several methods have been proposed to obtain higher-quality vectors for these words, leveraging both this context information and sometimes the word forms themselves through a hybrid approach. We show that the current tasks do not suffice to evaluate models that use word-form information, as such models can easily leverage word forms in the training data that are related to word forms in the test data. We introduce 3 new tasks, allowing for a more balanced comparison between models. Furthermore, we show that hyperparameters that have largely been ignored in previous work can consistently improve the performance of both baseline and advanced models, achieving a new state of the art on 4 out of 6 tasks.

Functional Distributional Semantics provides a linguistically interpretable framework for distributional semantics, by representing the meaning of a word as a function (a binary classifier), instead of a vector. However, the large number of latent variables means that inference is computationally expensive, and training a model is therefore slow to converge. In this paper, I introduce the Pixie Autoencoder, which augments the generative model of Functional Distributional Semantics with a graphconvolutional neural network to perform amortised variational inference. This allows the model to be trained more effectively, achieving better results on two tasks (semantic similarity in context and semantic composition), and outperforming BERT, a large pre-trained language model.

A broad-coverage corpus such as the Human Language Project envisioned by Abney and Bird (2010) would be a powerful resource for the study of endangered languages. Existing corpora are limited in the range of languages covered, in standardisation, or in machine-readability. In this paper we present SeedLing, a seed corpus for the Human Language Project. We first survey existing efforts to compile cross-linguistic resources, then describe our own approach. To build the foundation text for a Universal Corpus, we crawl and clean texts from several web sources that contain data from a large number of languages, and convert them into a standardised form consistent with the guidelines of Abney and Bird (2011). The resulting corpus is more easily-accessible and machine-readable than any of the underlying data sources, and, with data from 1451 languages covering 105 language families, represents a significant base corpus for researchers to draw on and add to in the future. To demonstrate the utility of SeedLing for cross-lingual computational research, we use our data in the test application of detecting similar languages.

Vector space models have become popular in distributional semantics, despite the challenges they face in capturing various semantic phenomena. We propose a novel probabilistic framework which draws on both formal semantics and recent advances in machine learning. In particular, we separate predicates from the entities they refer to, allowing us to perform Bayesian inference based on logical forms. We describe an implementation of this framework using a combination of Restricted Boltzmann Machines and feedforward neural networks. Finally, we demonstrate the feasibility of this approach by training it on a parsed corpus and evaluating it on established similarity datasets.