Paraphrase detection is useful in many natural language understanding applications. Current works typically formulate this problem as a sentence pair binary classification task. However, this setup is not a good fit for many of the intended applications of paraphrase models. In particular, such applications often involve finding the closest paraphrases of the target sentence from a group of candidate sentences where they exhibit different degrees of semantic overlap with the target sentence. To apply models to this paraphrase retrieval scenario, the model must be sensitive to the degree to which two sentences are paraphrases of one another. However, many existing datasets ignore and fail to test models in this setup. In response, we propose adversarial paradigms to create evaluation datasets, which could examine the sensitivity to different degrees of semantic overlap. Empirical results show that, while paraphrase models and different sentence encoders appear successful on standard evaluations, measuring the degree of semantic overlap still remains a big challenge for them.

Compositionality and inference are essential features of human language, and should hence be simultaneously accessible to a model of meaning. Despite being theory-grounded, distributional models can only be directly tested on compositionality, usually through similarity judgements, while testing for inference requires external resources. Recent work has shown that knowledge graph embeddings (KGE) architectures can be used to train distributional models capable of learning syntax-aware compositional representations, by training on syntactic graphs. We propose to expand such work with Multi-Graphs embedding (MuG) models, a new set of models learning from syntactic and knowledge-graphs. Using a phrase-level inference task, we show how MuGs can simultaneously handle syntax-aware composition and inference, and remain competitive distributional models with respect to lexical and compositional similarity.

Paraphrase identification involves identifying whether a pair of sentences express the same or similar meanings. While cross-encoders have achieved high performances across several benchmarks, bi-encoders such as SBERT have been widely applied to sentence pair tasks. They exhibit substantially lower computation complexity and are better suited to symmetric tasks. In this work, we adopt a bi-encoder approach to the paraphrase identification task, and investigate the impact of explicitly incorporating predicate-argument information into SBERT through weighted aggregation. Experiments on six paraphrase identification datasets demonstrate that, with a minimal increase in parameters, the proposed model is able to outperform SBERT/SRoBERTa significantly. Further, ablation studies reveal that the predicate-argument based component plays a significant role in the performance gain.

This paper addresses a deficiency in existing cross-lingual information retrieval (CLIR) datasets and provides a robust evaluation of CLIR systems’ disambiguation ability. CLIR is commonly tackled by combining translation and traditional IR. Due to translation ambiguity, the problem of ambiguity is worse in CLIR than in monolingual IR. But existing auto-generated CLIR datasets are dominated by searches for named entity mentions, which does not provide a good measure for disambiguation performance, as named entity mentions can often be transliterated across languages and tend not to have multiple translations. Therefore, we introduce a new evaluation dataset (MuSeCLIR) to address this inadequacy. The dataset focusses on polysemous common nouns with multiple possible translations. MuSeCLIR is constructed from multilingual Wikipedia and supports searches on documents written in European (French, German, Italian) and Asian (Chinese, Japanese) languages. We provide baseline statistical and neural model results on MuSeCLIR which show that MuSeCLIR has a higher requirement on the ability of systems to disambiguate query terms.

Previous work has demonstrated that pre-trained large language models (LLM) acquire knowledge during pre-training which enables reasoning over relationships between words (e.g, hyponymy) and more complex inferences over larger units of meaning such as sentences. Here, we investigate whether lexical entailment (LE, i.e. hyponymy or the is a relation between words) can be generalised in a compositional manner. Accordingly, we introduce PLANE (Phrase-Level Adjective-Noun Entailment), a new benchmark to test models on fine-grained compositional entailment using adjective-noun phrases. Our experiments show that knowledge extracted via In–Context and transfer learning is not enough to solve PLANE. However, a LLM trained on PLANE can generalise well to out–of–distribution sets, since the required knowledge can be stored in the representations of subwords (SW) tokens.

Previous works have demonstrated the effectiveness of utilising pre-trained sentence encoders based on their sentence representations for meaning comparison tasks. Though such representations are shown to capture hidden syntax structures, the direct similarity comparison between them exhibits weak sensitivity to word order and structural differences in given sentences. A single similarity score further makes the comparison process hard to interpret. Therefore, we here propose to combine sentence encoders with an alignment component by representing each sentence as a list of predicate-argument spans (where their span representations are derived from sentence encoders), and decomposing the sentence-level meaning comparison into the alignment between their spans for paraphrase identification tasks. Empirical results show that the alignment component brings in both improved performance and interpretability for various sentence encoders. After closer investigation, the proposed approach indicates increased sensitivity to structural difference and enhanced ability to distinguish non-paraphrases with high lexical overlap.

Recently, impressive performance on various natural language understanding tasks has been achieved by explicitly incorporating syntax and semantic information into pre-trained models, such as BERT and RoBERTa. However, this approach depends on problem-specific fine-tuning, and as widely noted, BERT-like models exhibit weak performance, and are inefficient, when applied to unsupervised similarity comparison tasks. Sentence-BERT (SBERT) has been proposed as a general-purpose sentence embedding method, suited to both similarity comparison and downstream tasks. In this work, we show that by incorporating structural information into SBERT, the resulting model outperforms SBERT and previous general sentence encoders on unsupervised semantic textual similarity (STS) datasets and transfer classification tasks.

The automatic detection of hypernymy relationships represents a challenging problem in NLP. The successful application of state-of-the-art supervised approaches using distributed representations has generally been impeded by the limited availability of high quality training data. We have developed two novel data augmentation techniques which generate new training examples from existing ones. First, we combine the linguistic principles of hypernym transitivity and intersective modifier-noun composition to generate additional pairs of vectors, such as “small dog - dog” or “small dog - animal”, for which a hypernymy relationship can be assumed. Second, we use generative adversarial networks (GANs) to generate pairs of vectors for which the hypernymy relation can also be assumed. We furthermore present two complementary strategies for extending an existing dataset by leveraging linguistic resources such as WordNet. Using an evaluation across 3 different datasets for hypernymy detection and 2 different vector spaces, we demonstrate that both of the proposed automatic data augmentation and dataset extension strategies substantially improve classifier performance.

HTML tags are typically discarded in free text Named Entity Recognition from Web pages. We investigate whether these discarded tags might be used to improve NER performance. We compare Text+Tags sentences with their Text-Only equivalents, over five datasets, two free text segmentation granularities and two NER models. We find an increased F1 performance for Text+Tags of between 0.9% and 13.2% over all datasets, variants and models. This performance increase, over datasets of varying entity types, HTML density and construction quality, indicates our method is flexible and adaptable. These findings imply that a similar technique might be of use in other Web-aware NLP tasks, including the enrichment of deep language models.

In this paper, we investigate whether an a priori disambiguation of word senses is strictly necessary or whether the meaning of a word in context can be disambiguated through composition alone. We evaluate the performance of off-the-shelf single-vector and multi-sense vector models on a benchmark phrase similarity task and a novel task for word-sense discrimination. We find that single-sense vector models perform as well or better than multi-sense vector models despite arguably less clean elementary representations. Our findings furthermore show that simple composition functions such as pointwise addition are able to recover sense specific information from a single-sense vector model remarkably well.

Count-based distributional semantic models suffer from sparsity due to unobserved but plausible co-occurrences in any text collection. This problem is amplified for models like Anchored Packed Trees (APTs), that take the grammatical type of a co-occurrence into account. We therefore introduce a novel form of distributional inference that exploits the rich type structure in APTs and infers missing data by the same mechanism that is used for semantic composition.

Non-compositional phrases such as red herring and weakly compositional phrases such as spelling bee are an integral part of natural language (Sag, 2002). They are also the phrases that are difficult, or even impossible, for good compositional distributional models of semantics. Compositionality detection therefore provides a good testbed for compositional methods. We compare an integrated compositional distributional approach, using sparse high dimensional representations, with the ad-hoc compositional approach of applying simple composition operations to state-of-the-art neural embeddings.

We address the issue of how to associate frequency information with lexicalized grammar formalisms, using Lexicalized Tree Adjoining Grammar as a representative framework. We consider systematically a number of alternative probabilistic frameworks, evaluating their adequacy from both a theoretical and empirical perspective using data from existing large treebanks. We also propose three orthogonal approaches fo r backing off probability estimates to cope with the large number of parameters involved.

In wide-coverage lexicalized grammars many of the elementary structures have substructures in common. This means that during parsing some of the computation associated with different structures is duplicated. This paper explores ways in which the grammar can be precompiled into finite state automata so that some of this shared structure results in shared computation at run-time.