Learning scientific document representations can be substantially improved through contrastive learning objectives, where the challenge lies in creating positive and negative training samples that encode the desired similarity semantics. Prior work relies on discrete citation relations to generate contrast samples. However, discrete citations enforce a hard cut-off to similarity. This is counter-intuitive to similarity-based learning, and ignores that scientific papers can be very similar despite lacking a direct citation - a core problem of finding related research. Instead, we use controlled nearest neighbor sampling over citation graph embeddings for contrastive learning. This control allows us to learn continuous similarity, to sample hard-to-learn negatives and positives, and also to avoid collisions between negative and positive samples by controlling the sampling margin between them. The resulting method SciNCL outperforms the state-of-the-art on the SciDocs benchmark. Furthermore, we demonstrate that it can train (or tune) models sample-efficiently, and that it can be combined with recent training-efficient methods. Perhaps surprisingly, even training a general-domain language model this way outperforms baselines pretrained in-domain.
Contrastive learning has been the dominant approach to training dense retrieval models. In this work, we investigate the impact of ranking context - an often overlooked aspect of learning dense retrieval models. In particular, we examine the effect of its constituent parts: jointly scoring a large number of negatives per query, using retrieved (query-specific) instead of random negatives, and a fully list-wise loss. To incorporate these factors into training, we introduce Contextual Document Embedding Reranking (CODER), a highly efficient retrieval framework. When reranking, it incurs only a negligible computational overhead on top of a first-stage method at run time (delay per query in the order of milliseconds), allowing it to be easily combined with any state-of-the-art dual encoder method. After fine-tuning through CODER, which is a lightweight and fast process, models can also be used as stand-alone retrievers. Evaluating CODER in a large set of experiments on the MS~MARCO and TripClick collections, we show that the contextual reranking of precomputed document embeddings leads to a significant improvement in retrieval performance. This improvement becomes even more pronounced when more relevance information per query is available, shown in the TripClick collection, where we establish new state-of-the-art results by a large margin.
Sentiment Analysis is one of the most classical and primarily studied natural language processing tasks. This problem had a notable advance with the proposition of more complex and scalable machine learning models. Despite this progress, the Brazilian Portuguese language still disposes only of limited linguistic resources, such as datasets dedicated to sentiment classification, especially when considering the existence of predefined partitions in training, testing, and validation sets that would allow a more fair comparison of different algorithm alternatives. Motivated by these issues, this work analyzes the predictive performance of a range of document embedding strategies, assuming the polarity as the system outcome. This analysis includes five sentiment analysis datasets in Brazilian Portuguese, unified in a single dataset, and a reference partitioning in training, testing, and validation sets, both made publicly available through a digital repository. A cross-evaluation of dataset-specific models over different contexts is conducted to evaluate their generalization capabilities and the feasibility of adopting a unique model for addressing all scenarios.
Keyphrase extraction (KPE) automatically extracts phrases in a document that provide a concise summary of the core content, which benefits downstream information retrieval and NLP tasks. Previous state-of-the-art (SOTA) methods select candidate keyphrases based on the similarity between learned representations of the candidates and the document. They suffer performance degradation on long documents due to discrepancy between sequence lengths which causes mismatch between representations of keyphrase candidates and the document. In this work, we propose a novel unsupervised embedding-based KPE approach, Masked Document Embedding Rank (MDERank), to address this problem by leveraging a mask strategy and ranking candidates by the similarity between embeddings of the source document and the masked document. We further develop a KPE-oriented BERT (KPEBERT) model by proposing a novel self-supervised contrastive learning method, which is more compatible to MDERank than vanilla BERT. Comprehensive evaluations on six KPE benchmarks demonstrate that the proposed MDERank outperforms state-of-the-art unsupervised KPE approach by average 1.80 $F1@15$ improvement. MDERank further benefits from KPEBERT and overall achieves average 3.53 $F1@15$ improvement over the SOTA SIFRank. Our code is available at \url{https://github.com/LinhanZ/mderank}.
Embedding based methods are widely used for unsupervised keyphrase extraction (UKE) tasks. Generally, these methods simply calculate similarities between phrase embeddings and document embedding, which is insufficient to capture different context for a more effective UKE model. In this paper, we propose a novel method for UKE, where local and global contexts are jointly modeled. From a global view, we calculate the similarity between a certain phrase and the whole document in the vector space as transitional embedding based models do. In terms of the local view, we first build a graph structure based on the document where phrases are regarded as vertices and the edges are similarities between vertices. Then, we proposed a new centrality computation method to capture local salient information based on the graph structure. Finally, we further combine the modeling of global and local context for ranking. We evaluate our models on three public benchmarks (Inspec, DUC 2001, SemEval 2010) and compare with existing state-of-the-art models. The results show that our model outperforms most models while generalizing better on input documents with different domains and length. Additional ablation study shows that both the local and global information is crucial for unsupervised keyphrase extraction tasks.
We develop a Multimodal Spatiotemporal Neural Fusion network for Multi-Task Learning (MSNF-MTCL) to predict 5 important students' retention risks: future dropout, next semester dropout, type of dropout, duration of dropout and cause of dropout. First, we develop a general purpose multi-modal neural fusion network model MSNF for learning students' academic information representation by fusing spatial and temporal unstructured advising notes with spatiotemporal structured data. MSNF combines a Bidirectional Encoder Representations from Transformers (BERT)-based document embedding framework to represent each advising note, Long-Short Term Memory (LSTM) network to model temporal advising note embeddings, LSTM network to model students' temporal performance variables and students' static demographics altogether. The final fused representation from MSNF has been utilized on a Multi-Task Cascade Learning (MTCL) model towards building MSNF-MTCL for predicting 5 student retention risks. We evaluate MSNFMTCL on a large educational database consists of 36,445 college students over 18 years period of time that provides promising performances comparing with the nearest state-of-art models. Additionally, we test the fairness of such model given the existence of biases.
Over the last few years, neural network derived word embeddings became popular in the natural language processing literature. Studies conducted have mostly focused on the quality and application of word embeddings trained on public available corpuses such as Wikipedia or other news and social media sources. However, these studies are limited to generic text and thus lack technical and scientific nuances such as domain specific vocabulary, abbreviations, or scientific formulas which are commonly used in academic context. This research focuses on the performance of word embeddings applied to a large scale academic corpus. More specifically, we compare quality and efficiency of trained word embeddings to TFIDF representations in modeling content of scientific articles. We use a word2vec skip-gram model trained on titles and abstracts of about 70 million scientific articles. Furthermore, we have developed a benchmark to evaluate content models in a scientific context. The benchmark is based on a categorization task that matches articles to journals for about 1.3 million articles published in 2017. Our results show that content models based on word embeddings are better for titles (short text) while TFIDF works better for abstracts (longer text). However, the slight improvement of TFIDF for larger text comes at the expense of 3.7 times more memory requirement as well as up to 184 times higher computation times which may make it inefficient for online applications. In addition, we have created a 2-dimensional visualization of the journals modeled via embeddings to qualitatively inspect embedding model. This graph shows useful insights and can be used to find competitive journals or gaps to propose new journals.
Transformer is important for text modeling. However, it has difficulty in handling long documents due to the quadratic complexity with input text length. In order to handle this problem, we propose a hierarchical interactive Transformer (Hi-Transformer) for efficient and effective long document modeling. Hi-Transformer models documents in a hierarchical way, i.e., first learns sentence representations and then learns document representations. It can effectively reduce the complexity and meanwhile capture global document context in the modeling of each sentence. More specifically, we first use a sentence Transformer to learn the representations of each sentence. Then we use a document Transformer to model the global document context from these sentence representations. Next, we use another sentence Transformer to enhance sentence modeling using the global document context. Finally, we use hierarchical pooling method to obtain document embedding. Extensive experiments on three benchmark datasets validate the efficiency and effectiveness of Hi-Transformer in long document modeling.
Explicitly modelling field interactions and correlations in complex document structures has recently gained popularity in neural document embedding and retrieval tasks. Although this requires the specification of bespoke task-dependent models, encouraging empirical results are beginning to emerge. We present the first in-depth analyses of non-linear multi-field interaction (NL-MFI) ranking in the cooking domain in this work. Our results show that field-weighted factorisation machines models provide a statistically significant improvement over baselines in recipe retrieval tasks. Additionally, we show that sparsely capturing subsets of field interactions based on domain knowledge and feature selection heuristics offers significant advantages over baselines and exhaustive alternatives. Although field-interaction aware models are more elaborate from an architectural basis, they are often more data-efficient in optimisation and are better suited for explainability due to mirrored document and model factorisation.
We present a contrasting learning approach with data augmentation techniques to learn document representations in an unsupervised manner. Inspired by recent contrastive self-supervised learning algorithms used for image and NLP pretraining, we hypothesize that high-quality document embedding should be invariant to diverse paraphrases that preserve the semantics of the original document. With different backbones and contrastive learning frameworks, our study reveals the enormous benefits of contrastive augmentation for document representation learning with two additional insights: 1) including data augmentation in a contrastive way can substantially improve the embedding quality in unsupervised document representation learning, and 2) in general, stochastic augmentations generated by simple word-level manipulation work much better than sentence-level and document-level ones. We plug our method into a classifier and compare it with a broad range of baseline methods on six benchmark datasets. Our method can decrease the classification error rate by up to 6.4% over the SOTA approaches on the document classification task, matching or even surpassing fully-supervised methods.
Collecting supporting evidence from large corpora of text (e.g., Wikipedia) is of great challenge for open-domain Question Answering (QA). Especially, for multi-hop open-domain QA, scattered evidence pieces are required to be gathered together to support the answer extraction. In this paper, we propose a new retrieval target, hop, to collect the hidden reasoning evidence from Wikipedia for complex question answering. Specifically, the hop in this paper is defined as the combination of a hyperlink and the corresponding outbound link document. The hyperlink is encoded as the mention embedding which models the structured knowledge of how the outbound link entity is mentioned in the textual context, and the corresponding outbound link document is encoded as the document embedding representing the unstructured knowledge within it. Accordingly, we build HopRetriever which retrieves hops over Wikipedia to answer complex questions. Experiments on the HotpotQA dataset demonstrate that HopRetriever outperforms previously published evidence retrieval methods by large margins. Moreover, our approach also yields quantifiable interpretations of the evidence collection process.
One of the current state-of-the-art multilingual document embedding model LASER is based on the bidirectional LSTM neural machine translation model. This paper presents a transformer-based sentence/document embedding model, T-LASER, which makes three significant improvements. Firstly, the BiLSTM layers is replaced by the attention-based transformer layers, which is more capable of learning sequential patterns in longer texts. Secondly, due to the absence of recurrence, T-LASER enables faster parallel computations in the encoder to generate the text embedding. Thirdly, we augment the NMT translation loss function with an additional novel distance constraint loss. This distance constraint loss would further bring the embeddings of parallel sentences close together in the vector space; we call the T-LASER model trained with distance constraint, cT-LASER. Our cT-LASER model significantly outperforms both BiLSTM-based LASER and the simpler transformer-based T-LASER.
Due to the increasing amount of data on the internet, finding a highly-informative, low-dimensional representation for text is one of the main challenges for efficient natural language processing tasks including text classification. This representation should capture the semantic information of the text while retaining their relevance level for document classification. This approach maps the documents with similar topics to a similar space in vector space representation. To obtain representation for large text, we propose the utilization of deep Siamese neural networks. To embed document relevance in topics in the distributed representation, we use a Siamese neural network to jointly learn document representations. Our Siamese network consists of two sub-network of multi-layer perceptron. We examine our representation for the text categorization task on BBC news dataset. The results show that the proposed representations outperform the conventional and state-of-the-art representations in the text classification task on this dataset.
Canonical automatic summary evaluation metrics, such as ROUGE, focus on lexical similarity which cannot well capture semantics nor linguistic quality and require a reference summary which is costly to obtain. Recently, there have been a growing number of efforts to alleviate either or both of the two drawbacks. In this paper, we present a proof-of-concept study to a weakly supervised summary evaluation approach without the presence of reference summaries. Massive data in existing summarization datasets are transformed for training by pairing documents with corrupted reference summaries. In cross-domain tests, our strategy outperforms baselines with promising improvements, and show a great advantage in gauging linguistic qualities over all metrics.
Text classification is fundamental in natural language processing (NLP), and Graph Neural Networks (GNN) are recently applied in this task. However, the existing graph-based works can neither capture the contextual word relationships within each document nor fulfil the inductive learning of new words. In this work, to overcome such problems, we propose TextING for inductive text classification via GNN. We first build individual graphs for each document and then use GNN to learn the fine-grained word representations based on their local structures, which can also effectively produce embeddings for unseen words in the new document. Finally, the word nodes are aggregated as the document embedding. Extensive experiments on four benchmark datasets show that our method outperforms state-of-the-art text classification methods.
Understanding intrinsic patterns and predicting spatiotemporal characteristics of cities require a comprehensive representation of urban neighborhoods. Existing works relied on either inter- or intra-region connectivities to generate neighborhood representations but failed to fully utilize the informative yet heterogeneous data within neighborhoods. In this work, we propose Urban2Vec, an unsupervised multi-modal framework which incorporates both street view imagery and point-of-interest (POI) data to learn neighborhood embeddings. Specifically, we use a convolutional neural network to extract visual features from street view images while preserving geospatial similarity. Furthermore, we model each POI as a bag-of-words containing its category, rating, and review information. Analog to document embedding in natural language processing, we establish the semantic similarity between neighborhood ("document") and the words from its surrounding POIs in the vector space. By jointly encoding visual, textual, and geospatial information into the neighborhood representation, Urban2Vec can achieve performances better than baseline models and comparable to fully-supervised methods in downstream prediction tasks. Extensive experiments on three U.S. metropolitan areas also demonstrate the model interpretability, generalization capability, and its value in neighborhood similarity analysis.
Graph and language embedding models are becoming commonplace in large scale analyses given their ability to represent complex sparse data densely in low-dimensional space. Integrating these models' complementary relational and communicative data may be especially helpful if predicting rare events or classifying members of hidden populations - tasks requiring huge and sparse datasets for generalizable analyses. For example, due to social stigma and comorbidities, mental health support groups often form in amorphous online groups. Predicting suicidality among individuals in these settings using standard network analyses is prohibitive due to resource limits (e.g., memory), and adding auxiliary data like text to such models exacerbates complexity- and sparsity-related issues. Here, I show how merging graph and language embedding models (metapath2vec and doc2vec) avoids these limits and extracts unsupervised clustering data without domain expertise or feature engineering. Graph and language distances to a suicide support group have little correlation (\r{ho} < 0.23), implying the two models are not embedding redundant information. When used separately to predict suicidality among individuals, graph and language data generate relatively accurate results (69% and 76%, respectively); however, when integrated, both data produce highly accurate predictions (90%, with 10% false-positives and 12% false-negatives). Visualizing graph embeddings annotated with predictions of potentially suicidal individuals shows the integrated model could classify such individuals even if they are positioned far from the support group. These results extend research on the importance of simultaneously analyzing behavior and language in massive networks and efforts to integrate embedding models for different kinds of data when predicting and classifying, particularly when they involve rare events.
One of the principal tasks of machine learning with major applications is text classification. This paper focuses on the legal domain and, in particular, on the classification of lengthy legal documents. The main challenge that this study addresses is the limitation that current models impose on the length of the input text. In addition, the present paper shows that dividing the text into segments and later combining the resulting embeddings with a BiLSTM architecture to form a single document embedding can improve results. These advancements are achieved by utilising a simpler structure, rather than an increasingly complex one, which is often the case in NLP research. The dataset used in this paper is obtained from an online public database containing lengthy legal documents with highly domain-specific vocabulary and thus, the comparison of our results to the ones produced by models implemented on the commonly used datasets would be unjustified. This work provides the foundation for future work in document classification in the legal field.
PubMed is the biggest and most used bibliographic database worldwide, hosting more than 26M biomedical publications. One of its useful features is the "similar articles" section, allowing the end-user to find scientific articles linked to the consulted document in term of context. The aim of this study is to analyze whether it is possible to replace the statistic model PubMed Related Articles (pmra) with a document embedding method. Doc2Vec algorithm was used to train models allowing to vectorize documents. Six of its parameters were optimised by following a grid-search strategy to train more than 1,900 models. Parameters combination leading to the best accuracy was used to train models on abstracts from the PubMed database. Four evaluations tasks were defined to determine what does or does not influence the proximity between documents for both Doc2Vec and pmra. The two different Doc2Vec architectures have different abilities to link documents about a common context. The terminological indexing, words and stems contents of linked documents are highly similar between pmra and Doc2Vec PV-DBOW architecture. These algorithms are also more likely to bring closer documents having a similar size. In contrary, the manual evaluation shows much better results for the pmra algorithm. While the pmra algorithm links documents by explicitly using terminological indexing in its formula, Doc2Vec does not need a prior indexing. It can infer relations between documents sharing a similar indexing, without any knowledge about them, particularly regarding the PV-DBOW architecture. In contrary, the human evaluation, without any clear agreement between evaluators, implies future studies to better understand this difference between PV-DBOW and pmra algorithm.
In this paper we present a model for unsupervised topic discovery in texts corpora. The proposed model uses documents, words, and topics lookup table embedding as neural network model parameters to build probabilities of words given topics, and probabilities of topics given documents. These probabilities are used to recover by marginalization probabilities of words given documents. For very large corpora where the number of documents can be in the order of billions, using a neural auto-encoder based document embedding is more scalable then using a lookup table embedding as classically done. We thus extended the lookup based document embedding model to continuous auto-encoder based model. Our models are trained using probabilistic latent semantic analysis (PLSA) assumptions. We evaluated our models on six datasets with a rich variety of contents. Conducted experiments demonstrate that the proposed neural topic models are very effective in capturing relevant topics. Furthermore, considering perplexity metric, conducted evaluation benchmarks show that our topic models outperform latent Dirichlet allocation (LDA) model which is classically used to address topic discovery tasks.