Text Matching as Image Recognition

By Liang Pang, Yanyan Lan, Jiafeng Guo, Jun Xu, Shengxian Wan, Xueqi Cheng. 2016

578 citations recorded by the source corpus.

Matching two texts is a fundamental problem in many natural language processing tasks. An effective way is to extract meaningful matching patterns from words, phrases, and sentences to produce the matching score. Inspired by the success of convolutional neural network in image recognition, where neurons can capture many complicated patterns based on the extracted elementary visual patterns such as oriented edges and corners, we propose to model text matching as the problem of image recognition. Firstly, a matching matrix whose entries represent the similarities between words is constructed and viewed as an image. Then a convolutional neural network is utilized to capture rich matching patterns in a layer-by-layer way. We show that by resembling the compositional hierarchies of patterns in image recognition, our model can successfully identify salient signals such as n-gram and n-term matchings. Experimental results demonstrate its superiority against the baselines.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

Citations in the corpus

End-to-End Neural Ad-hoc Ranking with Kernel Pooling (2017, background): “bedding similarity with the query. „e interaction based neural models, on the other hand, learn query-document matching pa−erns from word-level interactions. For example, ARC-II [10] and MatchPyramid [20] build hierarchical Convolutional Neural Networks (CNN) on the interactions of two texts’ word e”
A Deep Relevance Matching Model for Ad-hoc Retrieval (2016, methodology): “MP: MatchPyramid [19] is another state-of-the-art interaction-focused deep matching model and has been tested on two NLP tasks including paraphrase identification and paper citation matching.”
Challenges in Building Intelligent Open-domain Dialog Systems (2019, methodology): “Other matching models that were proposed originally for non-dialog tasks such as paraphrase detection, language inference, and reading comprehension [97, 149], have also been adapted and applied to dialog response ranking.”
Learning to Match using Local and Distributed Representations of Text for Web Search (2016, methodology): “[31] propose the use of matching matrices to represent the similarity of short texts, then apply a convolutional neural network inspired by those in computer vision.”
A Survey of the Usages of Deep Learning for Natural Language Processing (2020, background): “Some representation-focused approaches build deep learning models to produce good representations for the texts and then match the representations straightforwardly [133], [165], [166], whereas interactionfocused approaches first build local interactions directly and then use DNNs to learn how the t”
A Modern Perspective on Query Likelihood with Deep Generative Retrieval Models (2021, unknown): “The first group captures patterns of similarity values across terms that appear close together within the query and within the document [11, 20, 21, 34].”
Incorporating Query Term Independence Assumption for Efficient Retrieval and Ranking using Deep Neural Networks (2019, background): “However, recent performance breakthroughs on many IR tasks have been achieved by neural models [Hu et al., 2014, Pang et al., 2016, Mitra et al., 2017, Dai et al., 2018, Nogueira and Cho, 2019] that learn latent representations of the query or inspect interaction patterns between query and document…”
Ranking Distillation: Learning Compact Ranking Models With High Performance for Recommender System (2018, methodology): “As in [10, 27, 28, 37], three different evaluation metrics used are Precision@ n (Prec@ n ), nDCG@ n , and Mean Average Precision (MAP).”
Fine-grained Fact Verification with Kernel Graph Attention Network (2019, background): “The recent development of neural information retrieval models, especially the interaction based ones, have shown promising effectiveness in extracting soft match patterns from query-document interactions (Hu et al., 2014; Pang et al., 2016; Guo et al., 2016; Xiong et al., 2017; Dai et al., 2018).”
Match-SRNN: Modeling the Recursive Matching Structure with Spatial RNN (2016, methodology): “MatchPyramid [Pang et al., 2016] uses a CNN to model the generation of the global interaction as an abstraction of the word level and phrase level interactions.”
Response Ranking with Deep Matching Networks and External Knowledge in Information-seeking Conversation Systems (2018, methodology): “e deep matching networks used in our research belong to the interaction focused models due to their better performances on a variety of text matching tasks compared with representation focused models [6, 7, 20, 37, 38, 43]. We study different ways to build the interaction matching matrices to captur”
Zero-shot Neural Retrieval via Domain-targeted Synthetic Query Generation (2020, background): “As a result, neural retrieval models achieve new state-of-the-art performance on several information retrieval (IR) benchmarks (Guo et al., 2016; Pang et al., 2016; Hui et al., 2017; Dai et al., 2018; MacAvaney et al., 2019; Yang et al., 2019a; Gillick et al., 2018; Yang et al., 2019b).”
Beyond 512 Tokens: Siamese Multi-depth Transformer-based Hierarchical Encoder for Document Matching (2020, methodology): “A number of neural matching models have been proposed for information retrieval [7, 11, 12, 18, 19, 32–34 , 39 ].”
Neural information retrieval: at the end of the early years (2017, methodology): “…made in computer vision through employing a very large number of hidden layers (hence ‘‘deep’’ learning), such deep structures have typically been less effective in NLP and IR than more shallow architectures (Pang et al. 2016a), though again with notable recent exceptions (Conneau et al. 2016).”
Co-PACRR: A Context-Aware Neural IR Model for Ad-hoc Retrieval (2017, methodology): “herefore, models developed for natural language processing tasks can also be used as retrieval models by assigning a similarity score to individual query-document pairs. For example, ARC-I and ARC-II [19] are two such models developed for the tasks of sentence completion, identifying the response to”
Interpretable & Time-Budget-Constrained Contextualization for Re-Ranking (2020, methodology): “In the MatchPyramid model, we set the number of CNN layers to 5, each with kernel size 3 × 3 and 16 convolution channels.”
Layer-Wise Coordination between Encoder and Decoder for Neural Machine Translation (2018, background): “Such an approach has been shown to be effective for other NLP tasks such as text matching [16, 10, 19] or non-autoregressive machine translation [8].”
Advances and Challenges in Conversational Recommender Systems: A Survey (2021, methodology): “An alternative strategy, in contrast, is to combine the representation vectors of query and response first, and then a neural method is used on the combined representation pair to further learn the interaction [180, 173, 126, 105].”
An anatomy for neural search engines (2019, methodology): “Several works in the literature have shown that interaction-focused models work well for Neural IR, often beating traditional ranking algorithms [Guo et al., 2016; Lu and Li, 2013; Pang et al., 2016; Mitra et al., 2017].”