Entity typing is the task of assigning semantic types to the entities that are mentioned in a text. In the case of fine-grained entity typing (FET), a large set of candidate type labels is considered. Since obtaining sufficient amounts of manual annotations is then prohibitively expensive, FET models are typically trained using distant supervision. In this paper, we propose to improve on this process by pre-training an entity encoder such that embeddings of coreferring entities are more similar to each other than to the embeddings of other entities. The main problem with this strategy, which helps to explain why it has not previously been considered, is that predicted coreference links are often too noisy. We show that this problem can be addressed by using a simple trick: we only consider coreference links that are predicted by two different off-the-shelf systems. With this prudent use of coreference links, our pre-training strategy allows us to improve the state-of-the-art in benchmarks on fine-grained entity typing, as well as traditional entity extraction.
Relations such as “is influenced by”, “is known for” or “is a competitor of” are inherently graded: we can rank entity pairs based on how well they satisfy these relations, but it is hard to draw a line between those pairs that satisfy them and those that do not. Such graded relations play a central role in many applications, yet they are typically not covered by existing Knowledge Graphs. In this paper, we consider the possibility of using Large Language Models (LLMs) to fill this gap. To this end, we introduce a new benchmark, in which entity pairs have to be ranked according to how much they satisfy a given graded relation. The task is formulated as a few-shot ranking problem, where models only have access to a description of the relation and five prototypical instances. We use the proposed benchmark to evaluate state-of-the-art relation embedding strategies as well as several publicly available LLMs and closed conversational models such as GPT-4. We find that smaller language models struggle to outperform a naive baseline. Overall, the best results are obtained with the 11B parameter Flan-T5 model and the 13B parameter OPT model, where further increasing the model size does not seem to be beneficial. For all models, a clear gap with human performance remains.
We introduce RAGAs (Retrieval Augmented Generation Assessment), a framework for reference-free evaluation of Retrieval Augmented Generation (RAG) pipelines. RAGAs is available at [https://github.com/explodinggradients/ragas]. RAG systems are composed of a retrieval and an LLM based generation module. They provide LLMs with knowledge from a reference textual database, enabling them to act as a natural language layer between a user and textual databases, thus reducing the risk of hallucinations. Evaluating RAG architectures is challenging due to several dimensions to consider: the ability of the retrieval system to identify relevant and focused context passages, the ability of the LLM to exploit such passages faithfully, and the quality of the generation itself. With RAGAs, we introduce a suite of metrics that can evaluate these different dimensions without relying on ground truth human annotations. We posit that such a framework can contribute crucially to faster evaluation cycles of RAG architectures, which is especially important given the fast adoption of LLMs.
Concept embeddings offer a practical and efficient mechanism for injecting commonsense knowledge into downstream tasks. Their core purpose is often not to predict the commonsense properties of concepts themselves, but rather to identify commonalities, i.e. sets of concepts which share some property of interest. Such commonalities are the basis for inductive generalisation, hence high-quality concept embeddings can make learning easier and more robust. Unfortunately, standard embeddings primarily reflect basic taxonomic categories, making them unsuitable for finding commonalities that refer to more specific aspects (e.g. the colour of objects or the materials they are made of). In this paper, we address this limitation by explicitly modelling the different facets of interest when learning concept embeddings. We show that this leads to embeddings which capture a more diverse range of commonsense properties, and consistently improves results in downstream tasks such as ultra-fine entity typing and ontology completion.
Conceptual spaces represent entities in terms of their primitive semantic features. Such representations are highly valuable but they are notoriously difficult to learn, especially when it comes to modelling perceptual and subjective features. Distilling conceptual spaces from Large Language Models (LLMs) has recently emerged as a promising strategy, but existing work has been limited to probing pre-trained LLMs using relatively simple zero-shot strategies. We focus in particular on the task of ranking entities according to a given conceptual space dimension. Unfortunately, we cannot directly fine-tune LLMs on this task, because ground truth rankings for conceptual space dimensions are rare. We therefore use more readily available features as training data and analyse whether the ranking capabilities of the resulting models transfer to perceptual and subjective features. We find that this is indeed the case, to some extent, but having at least some perceptual and subjective features in the training data seems essential for achieving the best results.
Contextualised Language Models (LM) improve on traditional word embeddings by encoding the meaning of words in context. However, such models have also made it possible to learn high-quality decontextualised concept embeddings. Three main strategies for learning such embeddings have thus far been considered: (i) fine-tuning the LM to directly predict concept embeddings from the name of the concept itself, (ii) averaging contextualised representations of mentions of the concept in a corpus, and (iii) encoding definitions of the concept. As these strategies have complementary strengths and weaknesses, we propose to learn a unified embedding space in which all three types of representations can be integrated. We show that this allows us to outperform existing approaches in tasks such as ontology completion, which heavily depends on access to high-quality concept embeddings. We furthermore find that mentions and definitions are well-aligned in the resulting space, enabling tasks such as target sense verification, even without the need for any fine-tuning.
Natural language offers an appealing alternative to formal logics as a vehicle for representing knowledge. However, using natural language means that standard methods for automated reasoning can no longer be used. A popular solution is to use transformer-based language models (LMs) to directly reason about knowledge expressed in natural language, but this has two important limitations. First, the set of premises is often too large to be directly processed by the LM. This means that we need a retrieval strategy which can select the most relevant premises when trying to infer some conclusion. Second, LMs have been found to learn shortcuts and thus lack robustness, putting in doubt to what extent they actually understand the knowledge that is expressed. Given these limitations, we explore the following alternative: rather than using LMs to perform reasoning directly, we use them to learn embeddings of individual assertions. Reasoning is then carried out by manipulating the learned embeddings. We show that this strategy is feasible to some extent, while at the same time also highlighting the limitations of directly fine-tuning LMs to learn the required embeddings.
The task of inductive knowledge graph completion requires models to learn inference patterns from a training graph, which can then be used to make predictions on a disjoint test graph. Rule-based methods seem like a natural fit for this task, but in practice they significantly underperform state-of-the-art methods based on Graph Neural Networks (GNNs), such as NBFNet. We hypothesise that the underperformance of rule-based methods is due to two factors: (i) implausible entities are not ranked at all and (ii) only the most informative path is taken into account when determining the confidence in a given link prediction answer. To analyse the impact of these factors, we study a number of variants of a rule-based approach, which are specifically aimed at addressing the aforementioned issues. We find that the resulting models can achieve a performance which is close to that of NBFNet. Crucially, the considered variants only use a small fraction of the evidence that NBFNet relies on, which means that they largely keep the interpretability advantage of rule-based methods. Moreover, we show that a further variant, which does look at the full KG, consistently outperforms NBFNet.
Dictionary definitions play a prominent role in a wide range of NLP tasks, for instance by providing additional context about the meaning of rare and emerging terms. Many dictionaries also provide examples to illustrate the prototypical usage of words, which brings further opportunities for training or enriching NLP models. The intrinsic qualities of dictionaries, and related lexical resources such as glossaries and encyclopedias, are however still not well-understood. While there has been significant work on developing best practices, such guidance has been aimed at traditional usages of dictionaries (e.g. supporting language learners), and it is currently unclear how different quality aspects affect the NLP systems that rely on them. To address this issue, we compare WordNet, the most commonly used lexical resource in NLP, with a variety of dictionaries, as well as with examples that were generated by ChatGPT. Our analysis involves human judgments as well as automatic metrics. We furthermore study the quality of word embeddings derived from dictionary examples, as a proxy for downstream performance. We find that WordNet’s examples lead to lower-quality embeddings than those from the Oxford dictionary. Surprisingly, however, the ChatGPT generated examples were found to be most effective overall.
Ultra-fine entity typing (UFET) is the task of inferring the semantic types from a large set of fine-grained candidates that apply to a given entity mention. This task is especially challenging because we only have a small number of training examples for many types, even with distant supervision strategies. State-of-the-art models, therefore, have to rely on prior knowledge about the type labels in some way. In this paper, we show that the performance of existing methods can be improved using a simple technique: we use pre-trained label embeddings to cluster the labels into semantic domains and then treat these domains as additional types. We show that this strategy consistently leads to improved results as long as high-quality label embeddings are used. Furthermore, we use the label clusters as part of a simple post-processing technique, which results in further performance gains. Both strategies treat the UFET model as a black box and can thus straightforwardly be used to improve a wide range of existing models.
Modelling how concepts are related is a central topic in Lexical Semantics. A common strategy is to rely on knowledge graphs (KGs) such as ConceptNet, and to model the relation between two concepts as a set of paths. However, KGs are limited to a fixed set of relation types, and they are incomplete and often noisy. Another strategy is to distill relation embeddings from a fine-tuned language model. However, this is less suitable for words that are only indirectly related and it does not readily allow us to incorporate structured domain knowledge. In this paper, we aim to combine the best of both worlds. We model relations as paths but associate their edges with relation embeddings. The paths are obtained by first identifying suitable intermediate words and then selecting those words for which informative relation embeddings can be obtained. We empirically show that our proposed representations are useful for solving hard analogy questions.
Concepts play a central role in many applications. This includes settings where concepts have to be modelled in the absence of sentence context. Previous work has therefore focused on distilling decontextualised concept embeddings from language models. But concepts can be modelled from different perspectives, whereas concept embeddings typically mostly capture taxonomic structure. To address this issue, we propose a strategy for identifying what different concepts, from a potentially large concept vocabulary, have in common with others. We then represent concepts in terms of the properties they share with the other concepts. To demonstrate the practical usefulness of this way of modelling concepts, we consider the task of ultra-fine entity typing, which is a challenging multi-label classification problem. We show that by augmenting the label set with shared properties, we can improve the performance of the state-of-the-art models for this task.
The theory of Conceptual Spaces is an influential cognitive-linguistic framework for representing the meaning of concepts. Conceptual spaces are constructed from a set of quality dimensions, which essentially correspond to primitive perceptual features (e.g. hue or size). These quality dimensions are usually learned from human judgements, which means that applications of conceptual spaces tend to be limited to narrow domains (e.g. modelling colour or taste). Encouraged by recent findings about the ability of Large Language Models (LLMs) to learn perceptually grounded representations, we explore the potential of such models for learning conceptual spaces. Our experiments show that LLMs can indeed be used for learning meaningful representations to some extent. However, we also find that fine-tuned models of the BERT family are able to match or even outperform the largest GPT-3 model, despite being 2 to 3 orders of magnitude smaller.
In the last five years, there has been a significant focus in Natural Language Processing (NLP) on developing larger Pretrained Language Models (PLMs) and introducing benchmarks such as SuperGLUE and SQuAD to measure their abilities in language understanding, reasoning, and reading comprehension. These PLMs have achieved impressive results on these benchmarks, even surpassing human performance in some cases. This has led to claims of superhuman capabilities and the provocative idea that certain tasks have been solved. In this position paper, we take a critical look at these claims and ask whether PLMs truly have superhuman abilities and what the current benchmarks are really evaluating. We show that these benchmarks have serious limitations affecting the comparison between humans and PLMs and provide recommendations for fairer and more transparent benchmarks.
Many applications crucially rely on the availability of high-quality word vectors. To learn such representations, several strategies based on language models have been proposed in recent years. While effective, these methods typically rely on a large number of contextualised vectors for each word, which makes them impractical. In this paper, we investigate whether similar results can be obtained when only a few contextualised representations of each word can be used. To this end, we analyse a range of strategies for selecting the most informative sentences. Our results show that with a careful selection strategy, high-quality word vectors can be learned from as few as 5 to 10 sentences.
Patronizing and Condescending Language (PCL) is a subtle but harmful type of discourse, yet the task of recognizing PCL remains under-studied by the NLP community. Recognizing PCL is challenging because of its subtle nature, because available datasets are limited in size, and because this task often relies on some form of commonsense knowledge. In this paper, we study to what extent PCL detection models can be improved by pre-training them on other, more established NLP tasks. We find that performance gains are indeed possible in this way, in particular when pre-training on tasks focusing on sentiment, harmful language and commonsense morality. In contrast, for tasks focusing on political speech and social justice, no or only very small improvements were witnessed. These findings improve our understanding of the nature of PCL.
This paper presents an overview of Task 4 at SemEval-2022, which was focused on detecting Patronizing and Condescending Language (PCL) towards vulnerable communities. Two sub-tasks were considered: a binary classification task, where participants needed to classify a given paragraph as containing PCL or not, and a multi-label classification task, where participants needed to identify which types of PCL are present (if any). The task attracted more than 300 participants, 77 teams and 229 valid submissions. We provide an overview of how the task was organized, discuss the techniques that were employed by the different participants, and summarize the main resulting insights about PCL detection and categorization.
Interpreting patient case descriptions has emerged as a challenging problem for biomedical NLP, where the aim is typically to predict diagnoses, to recommended treatments, or to answer questions about cases more generally. Previous work has found that biomedical language models often lack the knowledge that is needed for such tasks. In this paper, we aim to improve their performance through a self-supervised intermediate fine-tuning strategy based on PubMed abstracts. Our solution builds on the observation that many of these abstracts are case reports, and thus essentially patient case descriptions. As a general strategy, we propose to fine-tune biomedical language models on the task of predicting masked medical concepts from such abstracts. We find that the success of this strategy crucially depends on the selection of the medical concepts to be masked. By ensuring that these concepts are sufficiently salient, we can substantially boost the performance of biomedical language models, achieving state-of-the-art results on two benchmarks.
Grasping the commonsense properties of everyday concepts is an important prerequisite to language understanding. While contextualised language models are reportedly capable of predicting such commonsense properties with human-level accuracy, we argue that such results have been inflated because of the high similarity between training and test concepts. This means that models which capture concept similarity can perform well, even if they do not capture any knowledge of the commonsense properties themselves. In settings where there is no overlap between the properties that are considered during training and testing, we find that the empirical performance of standard language models drops dramatically. To address this, we study the possibility of fine-tuning language models to explicitly model concepts and their properties. In particular, we train separate concept and property encoders on two types of readily available data: extracted hyponym-hypernym pairs and generic sentences. Our experimental results show that the resulting encoders allow us to predict commonsense properties with much higher accuracy than is possible by directly fine-tuning language models. We also present experimental results for the related task of unsupervised hypernym discovery.
Patronizing and condescending language is characterized, among others, by its subtle nature. It thus seems reasonable to assume that detecting condescending language in text would be harder than detecting more explicitly harmful language, such as hate speech. However, the results of a SemEval-2022 Task devoted to this topic paint a different picture, with the top-performing systems achieving remarkably strong results. In this paper, we analyse the surprising effectiveness of standard text classification methods in more detail. In particular, we highlight the presence of two rather different types of condescending language in the dataset from the SemEval task. Some inputs are condescending because of the way they talk about a particular subject, i.e. condescending language in this case is a linguistic phenomenon, which can, in principle, be learned from training examples. However, other inputs are condescending because of the nature of what is said, rather than the way in which it is expressed, e.g. by emphasizing stereotypes about a given community. In such cases, our ability to detect condescending language, with current methods, largely depends on the presence of similar examples in the training data.
One of the long-standing challenges in lexical semantics consists in learning representations of words which reflect their semantic properties. The remarkable success of word embeddings for this purpose suggests that high-quality representations can be obtained by summarizing the sentence contexts of word mentions. In this paper, we propose a method for learning word representations that follows this basic strategy, but differs from standard word embeddings in two important ways. First, we take advantage of contextualized language models (CLMs) rather than bags of word vectors to encode contexts. Second, rather than learning a word vector directly, we use a topic model to partition the contexts in which words appear, and then learn different topic-specific vectors for each word. Finally, we use a task-specific supervision signal to make a soft selection of the resulting vectors. We show that this simple strategy leads to high-quality word vectors, which are more predictive of semantic properties than word embeddings and existing CLM-based strategies.
Pre-trained language models have been found to capture a surprisingly rich amount of lexical knowledge, ranging from commonsense properties of everyday concepts to detailed factual knowledge about named entities. Among others, this makes it possible to distill high-quality word vectors from pre-trained language models. However, it is currently unclear to what extent it is possible to distill relation embeddings, i.e. vectors that characterize the relationship between two words. Such relation embeddings are appealing because they can, in principle, encode relational knowledge in a more fine-grained way than is possible with knowledge graphs. To obtain relation embeddings from a pre-trained language model, we encode word pairs using a (manually or automatically generated) prompt, and we fine-tune the language model such that relationally similar word pairs yield similar output vectors. We find that the resulting relation embeddings are highly competitive on analogy (unsupervised) and relation classification (supervised) benchmarks, even without any task-specific fine-tuning. Source code to reproduce our experimental results and the model checkpoints are available in the following repository: https://github.com/asahi417/relbert
Analogies play a central role in human commonsense reasoning. The ability to recognize analogies such as “eye is to seeing what ear is to hearing”, sometimes referred to as analogical proportions, shape how we structure knowledge and understand language. Surprisingly, however, the task of identifying such analogies has not yet received much attention in the language model era. In this paper, we analyze the capabilities of transformer-based language models on this unsupervised task, using benchmarks obtained from educational settings, as well as more commonly used datasets. We find that off-the-shelf language models can identify analogies to a certain extent, but struggle with abstract and complex relations, and results are highly sensitive to model architecture and hyperparameters. Overall the best results were obtained with GPT-2 and RoBERTa, while configurations using BERT were not able to outperform word embedding models. Our results raise important questions for future work about how, and to what extent, pre-trained language models capture knowledge about abstract semantic relations.
Pre-trained neural language models (LMs) have achieved impressive results in various natural language processing tasks, across different languages. Surprisingly, this extends to the social media genre, despite the fact that social media often has very different characteristics from the language that LMs have seen during training. A particularly striking example is the performance of AraBERT, an LM for the Arabic language, which is successful in categorizing social media posts in Arabic dialects, despite only having been trained on Modern Standard Arabic. Our hypothesis in this paper is that the performance of LMs for social media can nonetheless be improved by incorporating static word vectors that have been specifically trained on social media. We show that a simple method for incorporating such word vectors is indeed successful in several Arabic and English benchmarks. Curiously, however, we also find that similar improvements are possible with word vectors that have been trained on traditional text sources (e.g. Wikipedia).
Cross-lingual word embeddings are vector representations of words in different languages where words with similar meaning are represented by similar vectors, regardless of the language. Recent developments which construct these embeddings by aligning monolingual spaces have shown that accurate alignments can be obtained with little or no supervision, which usually comes in the form of bilingual dictionaries. However, the focus has been on a particular controlled scenario for evaluation, and there is no strong evidence on how current state-of-the-art systems would fare with noisy text or for language pairs with major linguistic differences. In this paper we present an extensive evaluation over multiple cross-lingual embedding models, analyzing their strengths and limitations with respect to different variables such as target language, training corpora and amount of supervision. Our conclusions put in doubt the view that high-quality cross-lingual embeddings can always be learned without much supervision.
Various methods have already been proposed for learning entity embeddings from text descriptions. Such embeddings are commonly used for inferring properties of entities, for recommendation and entity-oriented search, and for injecting background knowledge into neural architectures, among others. Entity embeddings essentially serve as a compact encoding of a similarity relation, but similarity is an inherently multi-faceted notion. By representing entities as single vectors, existing methods leave it to downstream applications to identify these different facets, and to select the most relevant ones. In this paper, we propose a model that instead learns several vectors for each entity, each of which intuitively captures a different aspect of the considered domain. We use a mixture-of-experts formulation to jointly learn these facet-specific embeddings. The individual entity embeddings are learned using a variant of the GloVe model, which has the advantage that we can easily identify which properties are modelled well in which of the learned embeddings. This is exploited by an associated gating network, which uses pre-trained word vectors to encourage the properties that are modelled by a given embedding to be semantically coherent, i.e. to encourage each of the individual embeddings to capture a meaningful facet.
In this paper, we introduce a new annotated dataset which is aimed at supporting the development of NLP models to identify and categorize language that is patronizing or condescending towards vulnerable communities (e.g. refugees, homeless people, poor families). While the prevalence of such language in the general media has long been shown to have harmful effects, it differs from other types of harmful language, in that it is generally used unconsciously and with good intentions. We furthermore believe that the often subtle nature of patronizing and condescending language (PCL) presents an interesting technical challenge for the NLP community. Our analysis of the proposed dataset shows that identifying PCL is hard for standard NLP models, with language models such as BERT achieving the best results.
We describe the system submitted to SemEval-2020 Task 6, Subtask 1. The aim of this subtask is to predict whether a given sentence contains a definition or not. Unsurprisingly, we found that strong results can be achieved by fine-tuning a pre-trained BERT language model. In this paper, we analyze the performance of this strategy. Among others, we show that results can be improved by using a two-step fine-tuning process, in which the BERT model is first fine-tuned on the full training set, and then further specialized towards a target domain.
This paper summarizes our contribution to the Hyperpartisan News Detection task in SemEval 2019. We experiment with two different approaches: 1) an SVM classifier based on word vector averages and hand-crafted linguistic features, and 2) a BiLSTM-based neural text classifier trained on a filtered training set. Surprisingly, despite their different nature, both approaches achieve an accuracy of 0.74. The main focus of this paper is to further analyze the remarkable fact that a simple feature-based approach can perform on par with modern neural classifiers. We also highlight the effectiveness of our filtering strategy for training the neural network on a large but noisy training set.
While word embeddings have been shown to implicitly encode various forms of attributional knowledge, the extent to which they capture relational information is far more limited. In previous work, this limitation has been addressed by incorporating relational knowledge from external knowledge bases when learning the word embedding. Such strategies may not be optimal, however, as they are limited by the coverage of available resources and conflate similarity with other forms of relatedness. As an alternative, in this paper we propose to encode relational knowledge in a separate word embedding, which is aimed to be complementary to a given standard word embedding. This relational word embedding is still learned from co-occurrence statistics, and can thus be used even when no external knowledge base is available. Our analysis shows that relational word vectors do indeed capture information that is complementary to what is encoded in standard word embeddings.
Word embedding models typically learn two types of vectors: target word vectors and context word vectors. These vectors are normally learned such that they are predictive of some word co-occurrence statistic, but they are otherwise unconstrained. However, the words from a given language can be organized in various natural groupings, such as syntactic word classes (e.g. nouns, adjectives, verbs) and semantic themes (e.g. sports, politics, sentiment). Our hypothesis in this paper is that embedding models can be improved by explicitly imposing a cluster structure on the set of context word vectors. To this end, our model relies on the assumption that context word vectors are drawn from a mixture of von Mises-Fisher (vMF) distributions, where the parameters of this mixture distribution are jointly optimized with the word vectors. We show that this results in word vectors which are qualitatively different from those obtained with existing word embedding models. We furthermore show that our embedding model can also be used to learn high-quality document representations.
Lexical relation classification is the task of predicting whether a certain relation holds between a given pair of words. In this paper, we explore to which extent the current distributional landscape based on word embeddings provides a suitable basis for classification of collocations, i.e., pairs of words between which idiosyncratic lexical relations hold. First, we introduce a novel dataset with collocations categorized according to lexical functions. Second, we conduct experiments on a subset of this benchmark, comparing it in particular to the well known DiffVec dataset. In these experiments, in addition to simple word vector arithmetic operations, we also investigate the role of unsupervised relation vectors as a complementary input. While these relation vectors indeed help, we also show that lexical function classification poses a greater challenge than the syntactic and semantic relations that are typically used for benchmarks in the literature.
Conceptual spaces are geometric representations of meaning that were proposed by G ̈ardenfors (2000). They share many similarities with the vector space embeddings that are commonly used in natural language processing. However, rather than representing entities in a single vector space, conceptual spaces are usually decomposed into several facets, each of which is then modelled as a relatively low dimensional vector space. Unfortunately, the problem of learning such conceptual spaces has thus far only received limited attention. To address this gap, we analyze how, and to what extent, a given vector space embedding can be decomposed into meaningful facets in an unsupervised fashion. While this problem is highly challenging, we show that useful facets can be discovered by relying on word embeddings to group semantically related features.
Entity embeddings are vector space representations of a given domain of interest. They are typically learned from text corpora (possibly in combination with any available structured knowledge), based on the intuition that similar entities should be represented by similar vectors. The usefulness of such entity embeddings largely stems from the fact that they implicitly encode a rich amount of knowledge about the considered domain, beyond mere similarity. In an embedding of movies, for instance, we may expect all movies from a given genre to be located in some low-dimensional manifold. This is particularly useful in supervised learning settings, where it may e.g. allow neural movie recommenders to base predictions on the genre of a movie, without that genre having to be specified explicitly for each movie, or without even the need to specify that the genre of a movie is a property that may have predictive value for the considered task. In unsupervised settings, however, such implicitly encoded knowledge cannot be leveraged. Conceptual spaces, as proposed by Grdenfors, are similar to entity embeddings, but provide more structure. In conceptual spaces, among others, dimensions are interpretable and grouped into facets, and properties and concepts are explicitly modelled as (vague) regions. Thanks to this additional structure, conceptual spaces can be used as a knowledge representation framework, which can also be effectively exploited in unsupervised settings. Given a conceptual space of movies, for instance, we are able to answer queries that ask about similarity w.r.t. a particular facet (e.g. movies which are cinematographically similar to Jurassic Park), that refer to a given feature (e.g. movies which are scarier than Jurassic Park but otherwise similar), or that refer to particular properties or concepts (e.g. thriller from the 1990s with a dinosaur theme). Compared to standard entity embeddings, however, conceptual spaces are more challenging to learn in a purely data-driven fashion. In this talk, I will give an overview of some approaches for learning such representations that have recently been developed within the context of the FLEXILOG project.
Given a set of instances of some relation, the relation induction task is to predict which other word pairs are likely to be related in the same way. While it is natural to use word embeddings for this task, standard approaches based on vector translations turn out to perform poorly. To address this issue, we propose two probabilistic relation induction models. The first model is based on translations, but uses Gaussians to explicitly model the variability of these translations and to encode soft constraints on the source and target words that may be chosen. In the second model, we use Bayesian linear regression to encode the assumption that there is a linear relationship between the vector representations of related words, which is considerably weaker than the assumption underlying translation based models.
We present SeVeN (Semantic Vector Networks), a hybrid resource that encodes relationships between words in the form of a graph. Different from traditional semantic networks, these relations are represented as vectors in a continuous vector space. We propose a simple pipeline for learning such relation vectors, which is based on word vector averaging in combination with an ad hoc autoencoder. We show that by explicitly encoding relational information in a dedicated vector space we can capture aspects of word meaning that are complementary to what is captured by word embeddings. For example, by examining clusters of relation vectors, we observe that relational similarities can be identified at a more abstract level than with traditional word vector differences. Finally, we test the effectiveness of semantic vector networks in two tasks: measuring word similarity and neural text categorization. SeVeN is available at bitbucket.org/luisespinosa/seven.
Word embedding models such as GloVe rely on co-occurrence statistics to learn vector representations of word meaning. While we may similarly expect that co-occurrence statistics can be used to capture rich information about the relationships between different words, existing approaches for modeling such relationships are based on manipulating pre-trained word vectors. In this paper, we introduce a novel method which directly learns relation vectors from co-occurrence statistics. To this end, we first introduce a variant of GloVe, in which there is an explicit connection between word vectors and PMI weighted co-occurrence vectors. We then show how relation vectors can be naturally embedded into the resulting vector space.
Automatically identifying definitional knowledge in text corpora (Definition Extraction or DE) is an important task with direct applications in, among others, Automatic Glossary Generation, Taxonomy Learning, Question Answering and Semantic Search. It is generally cast as a binary classification problem between definitional and non-definitional sentences. In this paper we present a set of neural architectures combining Convolutional and Recurrent Neural Networks, which are further enriched by incorporating linguistic information via syntactic dependencies. Our experimental results in the task of sentence classification, on two benchmarking DE datasets (one generic, one domain-specific), show that these models obtain consistent state of the art results. Furthermore, we demonstrate that models trained on clean Wikipedia-like definitions can successfully be applied to more noisy domain-specific corpora.
In this paper we consider semantic spaces consisting of objects from some particular domain (e.g. IMDB movie reviews). Various authors have observed that such semantic spaces often model salient features (e.g. how scary a movie is) as directions. These feature directions allow us to rank objects according to how much they have the corresponding feature, and can thus play an important role in interpretable classifiers, recommendation systems, or entity-oriented search engines, among others. Methods for learning semantic spaces, however, are mostly aimed at modelling similarity. In this paper, we argue that there is an inherent trade-off between capturing similarity and faithfully modelling features as directions. Following this observation, we propose a simple method to fine-tune existing semantic spaces, with the aim of improving the quality of their feature directions. Crucially, our method is fully unsupervised, requiring only a bag-of-words representation of the objects as input.
Cross-lingual word embeddings are becoming increasingly important in multilingual NLP. Recently, it has been shown that these embeddings can be effectively learned by aligning two disjoint monolingual vector spaces through linear transformations, using no more than a small bilingual dictionary as supervision. In this work, we propose to apply an additional transformation after the initial alignment step, which moves cross-lingual synonyms towards a middle point between them. By applying this transformation our aim is to obtain a better cross-lingual integration of the vector spaces. In addition, and perhaps surprisingly, the monolingual spaces also improve by this transformation. This is in contrast to the original alignment, which is typically learned such that the structure of the monolingual spaces is preserved. Our experiments confirm that the resulting cross-lingual embeddings outperform state-of-the-art models in both monolingual and cross-lingual evaluation tasks.
Human language has evolved towards newer forms of communication such as social media, where emojis (i.e., ideograms bearing a visual meaning) play a key role. While there is an increasing body of work aimed at the computational modeling of emoji semantics, there is currently little understanding about what makes a computational model represent or predict a given emoji in a certain way. In this paper we propose a label-wise attention mechanism with which we attempt to better understand the nuances underlying emoji prediction. In addition to advantages in terms of interpretability, we show that our proposed architecture improves over standard baselines in emoji prediction, and does particularly well when predicting infrequent emojis.
Vector representations of word meaning have found many applications in the field of natural language processing. Word vectors intuitively represent the average context in which a given word tends to occur, but they cannot explicitly model the diversity of these contexts. Although region representations of word meaning offer a natural alternative to word vectors, only few methods have been proposed that can effectively learn word regions. In this paper, we propose a new word embedding model which is based on SVM regression. We show that the underlying ranking interpretation of word contexts is sufficient to match, and sometimes outperform, the performance of popular methods such as Skip-gram. Furthermore, we show that by using a quadratic kernel, we can effectively learn word regions, which outperform existing unsupervised models for the task of hypernym detection.
We propose a new word embedding model, inspired by GloVe, which is formulated as a feasible least squares optimization problem. In contrast to existing models, we explicitly represent the uncertainty about the exact definition of each word vector. To this end, we estimate the error that results from using noisy co-occurrence counts in the formulation of the model, and we model the imprecision that results from including uninformative context words. Our experimental results demonstrate that this model compares favourably with existing word embedding models.