Learning Text Similarity with Siamese Recurrent Networks

“In this equation and below the logistic function is denoted by σ The Long Short-Term Memory (Hochreiter and Schmidhuber, 1997) variant of RNNs in particular has had success in tasks related to natural language processing, such as text classification (Graves, 2012) and language translation (Sutskever…”

“The parameters of the model are optimized using the Adam method (Kingma and Ba, 2014) and each model is trained until convergence.”

“, 2013) and Glove (Pennington et al., 2014) have seen much use in such models, but some go beyond the word level and represent text as a sequence of characters (Kim et”

“In the last few years, neural network models have been applied to tasks ranging from machine translation (Zou et al., 2013; Cho et al., 2014) to question answering (Weston et al.”

“For example, word similarity models based on dense embeddings (Mikolov et al., 2013) have recently been applied in diverse settings, such as sentiment analysis (dos Santos and Gatti, 2014) and recommender systems (Barkan and Koenig-stein, 2016).”

“…and below the logistic function is denoted by σ The Long Short-Term Memory (Hochreiter and Schmidhuber, 1997) variant of RNNs in particular has had success in tasks related to natural language processing, such as text classification (Graves, 2012) and language translation (Sutskever et al., 2014).”

“(1) Bidirectional RNNs (Schuster and Paliwal, 1997) incorporate both future and past context by running the reverse of the input through a separate RNN.”

“The use of (deep) neural networks for NLP has recently received much attention, starting from the seminal papers employing convolutional networks on traditional NLP tasks (Collobert et al., 2011) and the availability of high quality semantic word representations (Mikolov et al., 2013).”

“[120] present a similar model that uses character-level Bi-LSTMs for f1 and f2, and the”

“Table 13: Comparing the F1 accuracy of the Hybrid and RNN models in our submission with the model proposed in [56] for an NLP task.”

“[120] present a similar model that uses character-level Bi-LSTMs for f1 and f2, and the”

“Deep learning models such as recurrent neural networks (RNN) and its variants [20]–[22] and convolution neural networks (CNN) [23] have been used effectively in many NLP tasks that share similarities to fake news and consist of calculating semantic similarity between sentences [24], [25] and…”

“Building neural networks with two towers has recently become a popular approach in several natural language tasks including modeling sentence similarities [31], response suggestions [24], and text-based information retrieval [17, 37].”

“Language Processing) to learn the similarity between words [37] or sentences [36] whose goals share a lot in common with ours.”

“People started to adopt two-tower DNNs to learn representation from content features in language models [6, 16, 19].”

“…preserve information from both past and future, BiLSTM can better understand context than LSTM and show its advantages in many different text mining tasks, such as sentiment analysis [Chen et al. 2017a], sentence similarity [Neculoiu et al. 2016], and name entity recognition [Lample et al. 2016].”

“In literature [38], stacked BLSTM networks have been used for better semantic representation of sequences compared to single-layer BLSTM networks.”

“Deep learning models such as recurrent neural networks (RNN) and its variants [7, 13, 17] and convolution neural networks (CNN) [14] have been efficacious in solving many NLP problems that share similarities to fake news which includes but not limited to - computing semantic similarity between…”

“For instance, some studies [2, 27, 32, 33, 45, 46] exploit RNNs to model the sequential information of documents while other work [19, 38, 46, 54] applies CNNs to capture the important local messages in documents for semantic text matching.”

“matching queries with indexed documents [25].”

“They have been proven successful in tasks like signature verification (Bromley et al., 1993), face verification (Chopra et al., 2005), image similarity (Koch et al., 2015) and have been recently used successfully in sentence similarity (Neculoiu et al., 2016).”

“The technology to train a model over such a dataset is already available (Mueller & Thyagarajan, 2016; Neculoiu et al., 2016).”

“This one-shot recognition approach has proven to be useful in tasks like drug discovery [2], natural language processing [25], audio recognition [7] and image segmentation [36].”

“Since then, this neural network topology has been used in a broad range of applications, including classification tasks [26], online object tracking [27], and similarity function learning for images [28], human faces [29], and text [30].”

“Then gensim-based sentence similarity algorithm [71] was used to predict the participant’s opinion.”

“Moreover, they have confirmed that the Siamese networks are good at learning with a little data is available [5], [8], [28].”

“Siamese Encoder : we apply a Siamese neural network encoder (Neculoiu et al., 2016) onto each vertex v ∈ G AB to convert the word embeddings (Mikolov et al., 2013) of {S A ( v ) , S B ( v ) } into a fixed-sized hidden feature vector m AB ( v ) , which we call the match vector .”

“The second component of our model is a Siamese net [13] over two textual features in the dataset.”

“…[29], face verification [30], localization [31], image descriptors [32], zero-shot recognition [33], one-shot recognition [34], few-shot recognition [35], signature verification [36], sentence similarity [37], text similarity [38], detection [39], image segmentation [40], and image retrieval [41].”

“We use the contrastive loss [22] for this objective.”

“The first one is the Siamese network where two symmetric networks such as the convolutional neural network (CNN) and recurrent neural network (RNN) with sharing parameters are used to model the compared sentences [2, 18, 20, 22].”

“More recently, Siamese architecture has been applied to measure text similarity[20].”

“We adopt the siamese architecture [6], which has been widely used in vision and natural language applications [25] for similarity computation.”

“In [23, 24] a recurrent Siamese network with long-short term memory (LSTM) cells was proposed to learn semantic similarities between sentences.”

“skill-matching products have been developed in the last few years, for instance, BurningGlass, Workday, Pluralsight, EmployInsight, and TextKernel (see, for example, [19]).”

“…product matching task; a simple shallow neural network model based on the fastText (Joulin et al., 2016b; Bo-janowski et al., 2016; Joulin et al., 2016a) library and a deep siamese network model based on bidirectional long short term memory (Neculoiu et al., 2016; Mueller and Thyagarajan, 2016).”

“The Siamese structure (Bromley et al., 1994) for metric learning achieve great success and is widely applied (Amiri et al., 2016; Liu et al., 2018; Mueller and Thyagarajan, 2016; Neculoiu et al., 2016; Wan et al., 2016; He et al., 2015).”

“Moreover, we also bring forward its use from sentence/character-level to document-level representation learning and semantic relatedness measure, and this makes our work distinct from [24].”

“It has been successfully applied to transfer learning such as one-shot image recognition (Koch et al., 2015), text similarity (Neculoiu et al., 2016).”

“Siamese neural networks [6] constitute a special class of deep learning architectures, which are used in tasks involving similarity comparison, such as image or text matching [7, 9, 29].”

“Such approaches have been used to achieve job title normalization and classification [28] or learning sentence similarity [29].”

“It is widely used in vision (Chopra et al., 2005), and in NLP for semantic similarity, entailment, query normalization and QA (Mueller and Thyagarajan, 2016; Neculoiu et al., 2016; Das et al., 2016).”

“Later, researchers discovered that this method can also be applied to predict text similarity such as [28].”

“Matching network approaches such as Siamese networks (Nicosia and Moschitti 2017; Neculoiu et al. 2016), Bilateral Multi-Perspective Matching (BiMPM) (Wang et al. 2017) and other matching network proposals (Duan et al. 2018; Nie and Bansal 2017), also seem promising for tasks related to sentence…”

“sed at the last layer to denote similarity or dissimilarity of the objects fed to each tower. This architecture have been shown to be eective in learning similarities in several domains such as text [2] and images [3]. Figure2depicts an overview of the network used in our framework. ! ! ! a separate LSTM. The output o…”

“[33], presents a deep architecture for learning the similarity of variable-length sentences.”

“We train the LSTMs in Modeling Layer and the Distance Layer as an end-to-end Siamese Network[29].”

“as an optimal ratio for training Siamese networks [57].”

“The inputs to this Siamese Network are anchor , positive and negative tests, which are transformed by each sub-network of the Siamese Network into high-dimensional embeddings [46].”

“A two-layer bidirectional LSTM network on characters was used in [14] to normalize job titles.”

“It was originally used for signature verification [5] and has since been applied to various applications such as face verification [6], unsupervised acoustic modeling [43], and learning semantic entailment [32], as well as text similarity [34].”

“2020b), and text classification (Mueller and Thyagarajan 2016, Neculoiu et al .”

“[27] constructed a patent model tree and compared the advantages and disadvantages of CNN, RNN, LSTM, and Siamese LSTM, and established that Siamese LSTM [28,29] has obvious advantages among them.”

“Neculoiu et al. [7] proposed a deep architecture for learning a similarity metric on variable length character sequences.”

“The second component of our model is a Siamese net [13] over two textual features in the dataset.”

“[18] used a siamese RNN for job title normalization.”

“Specially, the pairs of threads used for training are sampled from the dataset in a 3:1 ratio of negative pairs (two threads with different classes) to the positive ones (two threads with the same class), which is considered as an optimal ratio for training siamese networks [49].”

“Neculoiu et al. [41] has used cosine similarity as the distance vector to find the text-similarity in their Sia-mese recurrent network.”

“…(DNN) is now a fundamental facet of applications ranging from visual search (Jing et al., 2015; Hadi Kiapour et al., 2015), semantic text matching (Neculoiu et al., 2016), oneshot classification (Koch et al., 2015), clustering (Oh Song et al., 2017), and recommendation (Shankar et al., 2017).”

“…been recently used in different application domains featuring sequential data, including natural language processing (Mueller and Thyagarajan, 2016; Neculoiu et al, 2016), computer vision (McLaughlin et al, 2016; Wu et al, 2018) and speaker identification (Li et al, 2017; Chung et al, 2018), but…”

“Neculoiu [17] et al. used the LSTM framework to propose a similarity measure method for variable-length character sequences, and projecting variable-length strings into a fixed-dimensional embedded space.”

“In NLP, contrastive learning has been applied to learn similarities between text (Mueller and Thyagarajan, 2016; Neculoiu et al., 2016) and speech utterances (Kamper et al.”

“The basic building blocks of all deep models are recurrent networks such as recurrent neural networks (RNN) [23], long short-term memory networks (LSTM) [16] and gated recurrent units (GRU) [11], and convolution networks such as convolution neural networks (CNN) [17].”

“As Siamese networks have already been shown to work well in areas where similarities between different entities need to be evaluated [38, 45], we chose this type of network to address the task of duplicate detection.”

“Furthermore, recognizing similarities in 1-D signal data, such as radar, speech, and natural language process (NLP) [2, 15, 32, 35], has also gained many usages in different applications.”

“[13] combined a stack of character-level bidirectional LSTM with Siamese architecture to compare the relevance of two words or phrases.”

“On the contrary, these methods using global features of sentences [10], [15] fail to consider local features of sentences, resulting in the problem of insufficient features.”

“Moreover, they proved that Siamese networks are good at learning on a dataset where a small amount of data is available [46, 62, 78].”

“Among the NLP community, the closest line of applicable work are the generic approaches of textual similarity (Cer et al., 2017; Wang et al., 2017; Neculoiu et al., 2016; Yin et al., 2016).”

“We further utilize Siamese recurrent architectures (Neculoiu et al., 2016) for learning sentence similarity, in which the architecture allows us to learn useful representations with limited supervision.”

“The cosine similarity between the two vector representations can then be used to determine the similarity between the two pieces of text [27].”

“A Siamese network, sharing the parameters between two RNNs or other neural network models [9] was better than one.”

“Siamese RNNs have been used for learning sentence similarity [8], human re-identification [9], and visual tracking [10], all of which solve the problem of sequential data comparison.”

“RNN has been successfully used to solve time-related issues, such as natural language processing [20], speech recognition [21], computation of text similarity [22].”

“…editor coordinating the review of this manuscript and approving it for publication was Jesus Felez . researchers employed Siamese neural networks (SNN) [11] consisting of dual recurrent neural networks with shared parameters, to model sentence pairs and compute the similarity via distance function.”

“1) Siamese-based networks (Neculoiu et al., 2016; Mueller and Thya-garajan, 2016): the input phrases are first encoded by the same network; the encoded vectors are then used to compute similarities by metrics like Co-sine.”

“Siamese neural networks (SNN)[7] with attention mechanism[8] for estimation semantic similarity, achieving good results on the general datasets, but unsatisfactory performance on biomedical datasets due to the long-dependency[15] problem.”

“[41] Word2vec LSTM’s Cos Input sentences”

“The approach of finding similar texts by using Siamese networks is described in [8].”

“Since the successful application in unsupervised feature learning, many more feature learning tasks begin to use the Siamese network recently [33, 34].”

“Siamese Network (SN) [3] is an architecture that allows learning distance metrics and low-dimensional representation with information about the similarity between pairs [22] of objects.”

“Following a prior work on learning text similarity (Neculoiu et al., 2016), we adopt the contrastive loss function for training.”