CN110765755A - A Semantic Similarity Feature Extraction Method Based on Double Selection Gate - Google Patents

Abstract

The invention discloses a semantic similarity feature extraction method based on double selection gate, which relates to the field of natural language processing. Input the two-way long-term memory network to obtain the context information vector of the two sentences, and then obtain the core feature vector of the sentence pair through the double selection gate, and then input the vector into the multi-angle semantic feature matching network to obtain the sentence pair. Feature matching vector, Finally, the matching vectors are combined with two semantic feature matching vectors through the bidirectional long-short-term memory network aggregation layer respectively, and the similarity prediction of sentence pairs is performed. The method effectively alleviates the problem of low matching efficiency caused by information redundancy, and at the same time avoids the cost problem of manually extracting core information.

Description

A Semantic Similarity Feature Extraction Method Based on Double Selection Gate

technical field

The invention relates to the field of natural language processing, in particular to a method for extracting semantic similarity features based on a double selection gate.

Background technique

Today's world is full of massive amounts of information, most of which are stored in the form of text, and an important topic of artificial intelligence is to organize and "express" this text information, so that computers can "understand" like humans. these messages. Due to the existence of many uncertain factors in the language, such as many words with multiple meanings, the same concept can be expressed in different ways, etc., the traditional text similarity calculation method based on string matching is used in search engines and question answering systems, etc. It has been difficult to meet user needs. When a user enters a keyword to find information that matches the keyword, the content returned by the search may correspond to the content that does not match the search. Extremely invariant, so deeper semantic understanding computing text similarity has become a hotspot in current natural language research.

There are many sentence semantic similarity matching methods in the prior art, which basically focus on the matching of strings at first. Map to a vector representation, and then use the obtained two sentence vectors to judge the similarity of the two sentences through the cosine distance. Although using the traditional string method to judge the similarity of sentence pairs helps people to filter out some irrelevant information when searching for related questions to a certain extent, the quality of the search results is still unsatisfactory. Because judging the similarity between sentences by strings is only to calculate the distance between words at the word level, without contextual semantic information, resulting in incorrect matching and ambiguity of information, and end users cannot quickly find relevant information about keywords.

Therefore, it is necessary to invent a new semantic similarity feature extraction method.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a semantic similarity feature extraction method based on a double selection gate, which can automatically determine the semantic similarity of two sentences, and effectively reduces the redundant information of sentences through double automatic selection of core information, and improves the performance of sentences. Similarity accuracy and judgment efficiency.

Its technical solutions are:

S100, performing word segmentation processing on the sentence pairs P and Q to be processed, and performing vectorized representation on the words after word segmentation processing to obtain word vectors;

S200. Input all word vectors of the sentence pairs P and Q obtained in step S100 into the first cyclic neural network in order to obtain a context information vector, wherein the last context information vector of the sentence represents the sentence vector of the sentence;

S300, input the sentence vectors of the sentence pairs P and Q into the first-level selection gate to obtain core information features;

S400, input the core information obtained in step S300 into the secondary selection gate, and obtain the core information feature again;

S500. Input the core information obtained in step S400 into the multi-angle semantic matching network, wherein the multi-angle semantic matching network includes four methods: full matching, maximum pooling matching, attention matching and maximum attention matching, to obtain sentence pairs feature matching vector;

S600. Pass the matching vector obtained in step S500 through the second neural network to fuse the feature matching vector into a fixed-length vector, and input it to the prediction layer to calculate the probability distribution of the similarity of sentence pairs.

Preferably, the first recurrent neural network is used to generate a state vector of context information.

Preferably, the first layer of the first recurrent neural network is a single-term long-short-term memory network, the second layer is a bidirectional long-short-term memory network, and each hierarchical structure includes a plurality of connected LSTM cell modules.

Preferably, the first recurrent neural network includes two hierarchical structures;

The first layer of the first recurrent neural network is used to generate word-level vectors;

The second layer of the first recurrent neural network is used to generate context information vectors.

Preferably, the primary selection gate and the secondary selection gate respectively comprise a plurality of primary selection gate units and secondary selection gate units;

The first-level selection gate and the second-level selection gate have different structures and different parameters.

Preferably, in the step S200, all the word vectors of the sentence pair obtained in the step S100 are input into the first cyclic network in order, so as to obtain the sentence state vector after each word is input, specifically:

Input the i-th word vector and the output word vector at the i-1th time into the i-th LSTM cell module, and obtain the state of the sentence after the i-th word vector is processed by the i-th LSTM cell module vector.

Preferably, in the step S300, the sentence vector of the sentence pair is input into the first-level selection gate, and the acquisition of core information features includes:

The context information vector at each moment of sentence P and the i-th sentence vector of sentence Q are input into the first-level selection gate unit, and the core information is obtained through processing by the i-th first-level selection gate unit.

Preferably, in step S400, the core information obtained in step S300 is input into the secondary selection gate, and obtaining the core information features again includes:

The core information processed by the i-th first-level selection gate unit is input into the i-th second-level selection gate unit, and the core information features are obtained through processing by the i-th second-level selection gate unit.

Preferably, in the step S500, the core information obtained in the step S400 is input into the multi-angle semantic matching network, and the obtained feature matching vector includes:

The full matching performs cosine similarity calculation on the context information vector described in sentence P and the sentence vector described in sentence Q at each moment to obtain a feature matching vector;

The maximum pooling matching performs cosine similarity calculation on the context information vector at every moment of sentence P and the context information vector at every moment of sentence Q, and selects the maximum value as a feature matching vector;

The attention matching performs cosine calculation on the context information vector at the ith moment of sentence P and the context information vector at the ith moment of sentence Q, respectively, to obtain i cosine values of sentence P, and weight the i cosine values as attention. The force weight is multiplied with the context information described in each moment of sentence Q, and the obtained result is then subjected to cosine calculation with the context information vector described in each moment of sentence P to obtain a feature matching vector;

The maximum attention matching performs cosine calculation on the context information vector at the i-th moment of sentence P and the context information vector at the i-th moment in sentence Q, respectively, to obtain i cosine values of sentence P, from the i cosine values. The largest value is selected as the attention weight and multiplied by the context information of sentence Q, and the result obtained is then cosine calculation with the context information vector described in sentence P at each moment to obtain the feature matching vector.

Preferably, the second neural network includes two bidirectional long-term and short-term memory networks, and the feature matching vectors for processing sentence pairs are aggregated into a fixed-length vector.

Preferably, in the step S600, the matching vector obtained in the step S500 is passed through the second neural network, so that the feature matching vector is fused into a fixed-length vector, and input to the prediction layer to calculate the similarity probability distribution of the sentence pair, including:

The four feature matching vectors obtained by the four matchings of the sentence P are aggregated into a fixed-length feature matching vector through the second cyclic neural network;

The four feature matching vectors obtained by the four matchings of the sentence Q are aggregated into a fixed-length feature matching vector through the bidirectional long-short-term memory network;

The two feature matching vectors of sentence P and sentence Q are input to the prediction layer to obtain the sentence pair similarity.

Preferably, in step S100, Word2Vec is used to perform vectorized representation on the words processed by Jieba word segmentation. Word2Vec is a predictive model that can efficiently learn embedded words. The basic idea of Word2Vec is to represent each word in natural language as a short vector with a unified meaning and dimension.

The beneficial effects brought by the technical solutions provided in the embodiments of the present invention are:

1. The semantic similarity feature extraction method based on the double selection gate of the present invention does not need to rely on manual removal of redundant information, automatically obtains the core information in the sentence, and can automatically determine the semantic similarity of two sentences through the semantic similarity model, and The sentence similarity determined by this model is more accurate and efficient, and can help users find more matching results in question answering or search systems.

2. The semantic similarity feature extraction method based on the double selection gate of the present invention utilizes a bidirectional long-short-term memory network to perform vectorized representation of context information on sentences. The network has a cell state that can capture long-distance dependencies of text, can remember long-term states, realize information update, forgetting, filtering, better express contextual relationships, and can solve the problem of network gradient disappearance and explosion. The traditional RNN network connects the past output and the current input to control the output of both through the activation function, and can only consider the state of the recent moment.

3. The semantic similarity feature extraction method based on the double selection gate of the present invention uses two selection gates to automatically obtain the core semantic information in the sentence, thereby avoiding the influence of redundant information on the judgment of the semantic similarity of the sentence, and improving the matching. efficiency.

4. The semantic similarity feature extraction method based on the double selection gate of the present invention uses a multi-angle semantic matching network to perform four matching methods of full matching, maximum pooling matching, attention matching and maximum attention matching for two sentences, The four matching methods make full use of the context information vector to perform multi-angle and more detailed matching, which effectively avoids the problem of low accuracy in determining the similarity by only the cosine distance between two sentences in the traditional method. The memory network integrates the matching vector into a fixed-length vector, which effectively controls the dimension of the matching vector, which is beneficial for the prediction layer to calculate the similarity of sentence pairs.

5. The semantic similarity feature extraction method based on the double selection gate of the present invention can effectively improve the accuracy and efficiency of sentence semantic similarity judgment, and is suitable for Chinese and English sentence pair corpus.

Description of drawings

FIG. 1 is a flowchart of a method according to an embodiment of the present invention.

FIG. 2 is a structural diagram of a dual selection gate module according to an embodiment of the present invention.

FIG. 3 is a structural diagram of a multi-angle semantic matching network according to an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

It should be noted that the embodiments of the present invention and the features of the embodiments may be combined with each other under the condition of no conflict.

Example 1

Referring to Fig. 1, the present invention provides a method for extracting semantic similarity features based on double selection gate, including:

S100: Perform word segmentation processing on the sentence pairs P and Q to be processed, and perform vectorized representation on the words after word segmentation processing to obtain word vectors.

The word segmentation process in step S100 is the process of dividing the words in the sentence into reasonable word sequences that conform to the contextual meaning. It is one of the key technologies and difficulties in natural language understanding and text information processing. an important part of the processing. The problem of Chinese word segmentation is more complicated. The reason is that there is no obvious mark between words, and the use of words is flexible, varied, and semantically rich, which is prone to ambiguity. According to research, the main difficulty of Chinese text segmentation based on statistics lies in ambiguity resolution, proper noun and new word discovery. The present invention uses Jieba to segment Chinese text and Nltk to segment English text, thereby improving the correct rate of word segmentation.

The models for vectorized representation of words include One-hot model and Distributed model. Among them, the One-hot model is simple, but the dimensions cannot be controlled, and the relationship between words cannot be well represented. Therefore, this method adopts the Distributed model, and specifically uses Word2Vec to vectorize the words.

S200. Input all word vectors of the sentence pairs P and Q obtained in step S100 into the first cyclic neural network in order to obtain a context information vector, wherein the last context information vector of the sentence represents the sentence vector of the sentence;

Among them, the first recurrent neural network is used to generate the state vector of context information; the first recurrent neural network includes two hierarchical structures, the first layer is a single-term long-term memory network, which is used to generate word-level vectors; the second layer is a bidirectional long-short-term memory network for generating context information vectors; each hierarchical structure includes multiple connected LSTM cell modules; modules in different hierarchical structures have different parameters to generate word-level and context information vectors.

Input all the word vectors of the sentence pair obtained in step S100 into the first cyclic network in sequence, so as to obtain the sentence state vector after each word is input, specifically:

Input the i-th word vector and the output word vector at the i-1th time into the i-th LSTM cell module, and obtain the state vector of the sentence after the i-th word vector after processing by the i-th LSTM cell module.

S300, input the sentence vectors of the sentence pairs P and Q into the first-level selection gate to obtain core information features;

Specifically, the context information vector of sentence P at each moment and the ith sentence vector of sentence Q are input into the first-level selection gate unit, and the core information is obtained after processing by the i-th first-level selection gate unit.

S400. Input the core information obtained in step S300 into the secondary selection gate, and obtain the core information feature again; specifically, input the core information processed by the i-th primary selection gate unit into the i-th secondary selection gate In the unit, the core information features are obtained through processing by the i-th secondary selection gate unit.

The primary selection gate and the secondary selection gate respectively comprise a plurality of primary selection gate units and secondary selection gate units;

The structure of the primary selection gate and the secondary selection gate are different, and the parameters are different.

S500. Input the core information obtained in step S400 into the multi-angle semantic matching network, wherein the multi-angle semantic matching network includes four methods: full matching, maximum pooling matching, attention matching and maximum attention matching, to obtain sentence pairs Feature matching vector; specifically,

In full matching, the cosine similarity calculation is performed between the context information vector of sentence P and the sentence vector of sentence Q at each moment, and the feature matching vector is obtained;

The maximum pooling matching performs cosine similarity calculation on the context information vector of sentence P at each moment and the context information vector of sentence Q at each moment, and selects the maximum value as the feature matching vector;

Attention matching performs cosine calculation on the context information vector of sentence P at the ith moment and the context information vector of sentence Q at the ith moment, respectively, to obtain i cosine values of sentence P, and weight the i cosine values as the attention weight and combine it with sentence Q. The context information at each moment is multiplied, and the result obtained is then subjected to cosine calculation with the context information vector of sentence P at each moment to obtain a feature matching vector;

The maximum attention matching performs cosine calculation on the context information vector at the ith moment of sentence P and the context information vector at the ith moment of sentence Q, respectively, to obtain i cosine values of sentence P, and selects the largest value from the i cosine values as attention. The force weight is multiplied with the context information of sentence Q, and the result obtained is then subjected to cosine calculation with the context information vector of sentence P at each moment to obtain the feature matching vector.

Among them, the second neural network includes two bidirectional long and short-term memory networks, and the feature matching vectors used to process sentence pairs are aggregated into a fixed-length vector.

S600. Pass the matching vector obtained in step S500 through the second neural network, so that the feature matching vector is fused into a fixed-length vector, and input to the prediction layer to calculate the similarity probability distribution of sentence pairs, specifically,

The four feature matching vectors obtained by the four matchings of the sentence P are aggregated into a fixed-length feature matching vector through the second cyclic neural network;

The four feature matching vectors obtained by the four matchings of the sentence Q are also aggregated into a fixed-length feature matching vector through the bidirectional long-short-term memory network;

The two feature matching vectors of sentence P and sentence Q are input to the prediction layer to obtain the sentence pair similarity.

In step S100, Word2Vec is used to vectorize the words processed by Jieba word segmentation.

Example 2

On the basis of Example 1, the first recurrent neural network consists of a layer of unidirectional LSTM network and a layer of bidirectional LSTM network, each layer includes a plurality of connected LSTM cell modules, according to the input gate in the LSTM cell module, The forget gate, update gate and filter output gate process the current input information and the output information of the previous moment. The first layer of the first recurrent neural network consists of multiple connected unidirectional LSTM cell modules for obtaining the state vector for each word. The second layer of the first recurrent neural network consists of multiple connected bidirectional LSTM cell modules for sentence context information vectors.

In this method, the first cyclic neural network is used to model the words and context information of the sentence, and the state vector of the corresponding moment of each word of the sentence and the context information vector of the sentence at each moment are obtained. Wherein, as shown in Figure 2, in step S200, the first cyclic neural network adopts the Long Short Term Memory Network (LSTM), and the calculation formula of the network is as follows:

f _t =σ(W _f w _t +U _f h _t-1 +b _f );

i _t =σ(W _i w _t +U _i h _t-1 +b _i );

o _t =σ(W _o w _t +U _o h _t-1 +b _o );

h _t =o _t tanh(c _t );

In the above formula, f _t is the output of the forget gate; i _t is the output of the input gate; _{o t} is the output of the output gate; W _f , Wi , W _o , W _c , b _f , b _i , b _o , b _c , are the weight matrix and bias vector of the forget gate, input gate, output gate, and selection gate; is the new memory information; c _t is the memory content of the updated LSTM network unit; σ is the sigmoid function; ⊙ is the element product; h _t-1 is the output of the hidden layer at time t-1, and W _t is the input information at time t .

In the method of the present invention, since the context of the sentence is modeled by the cyclic neural network, the state vector of the corresponding sentence after the word is input at time t theoretically includes the information of all words before the time, that is, the input of the last word The sentence state vector h _n obtained after the word contains all the information of the whole sentence. Therefore, h _n represents the state vector of the whole sentence, that is, the sentence vector.

Example 3

On the basis of Embodiment 1 or 2, the double selection gate includes two selection gate structures, and the two selection gates have different structures and different parameters. Through different selection gates, it is beneficial to filter out redundant information in sentences and obtain core information more accurately. The calculation formula of the first layer selection gate is as follows:

s= _hn ;

sGate _i =σ(W _{s hi} +U _s s+b);

是元素之间的点乘。In the above formula, use the sentence context hidden vector to construct its sentence vector, take the hidden layer h _n of the sentence as the sentence vector s, sGate _i as the gate vector, W _s and U _s are the weight matrix, b is the bias vector, σ is the sigmoid activation function,

is the dot product between elements.

The second-layer selection gate calculates the selection gate weight by calculating the context vector at time t, using the sentence vector at the previous time and the hidden layer state h′ _i of the selection gate, and finally normalizes the selection gate weight. The calculation formula is as follows:

e _i,j = v _a ^T tanh(W _a s _t-1 +U _a h' _i );

为权值矩阵，a_i,j为选择门选中归一化，

为第k个语句的核心特征向量，k＝1,2,...,L，L为文本中的语句数量。In the above formula, h′ _i is the context hidden vector;

is the weight matrix, a _i,j is the normalization selected by the selection gate,

is the core feature vector of the kth sentence, k=1,2,...,L, where L is the number of sentences in the text.

和Q语句的上下文向量表达式矩阵

经过步骤S300、S400中的两层选择门获取核心信息，语句P核心特征特征表达式 同理可得，语句Q表达式

Referring to Fig. 2, the statement P is P=[p ₁ , p ₂ ,...,pi ,...,p _n ], and the statement Q _is expressed as Q=[q ₁ ,q ₂ ,..., _qi ,...,q _m ] represents the input sentence pair sequence, the model inputs the words once and obtains the context information vector representation of each moment of the sentence through step S200, and the implicit vector expression matrix of the P sentence context

and the context vector expression matrix of the Q statement

The core information is obtained through the two-layer selection gate in steps S300 and S400, and the expression of the core feature of the statement P is Similarly, the statement Q expression

The method of the invention obtains the sentence context information vector through the cyclic neural network, so that the context semantic correlation of the two sentences is stronger, and the semantic similarity of the two sentences is better judged.

As shown in Fig. 3, the second recurrent neural network is a bidirectional LSTM neural network, including multiple bidirectional LSTM cell modules connected. In order to make the feature matching vector generated by the multi-angle matching network into a fixed-length vector input to the prediction layer, the matching vector needs to be input into the bidirectional LSTM network and fused into a fixed-length vector.

In order to obtain the similarity judgment of two sentences, the present invention uses a second cyclic neural network, and inputs the four feature matching vectors of sentence P and sentence Q into the second cyclic neural network to fuse to obtain a fixed-length vector, sentence Q and sentence P The four feature matching vectors of , use the same operations as above to obtain two fixed-length matching vectors respectively, and input the vectors into the prediction layer to obtain the probability distribution of sentence pair similarity.

The semantic similarity of sentences determined by the method of the present invention not only uses the context information between sentences, but also automatically extracts core information features from the sentences as the input of the matching network, which improves the matching accuracy and reduces the matching network for the same time. The processing of redundant information improves the matching efficiency. For some words with the same meaning and different expressions in the sentence, the model can also be used to determine that they are similar, such as the words "computer" and "computer". When determining the similarity of the two words, not only the similarity between the words is considered distance, but use the context information of the sentence in which the word is located to determine the similarity.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention. Inside.

Claims (10)

1. a semantic similarity feature extraction method based on double selection gate, is characterized in that, comprises the steps S100, performing word segmentation processing on the sentence pairs P and Q to be processed, and performing vectorized representation on the words after word segmentation processing to obtain word vectors; S200. Input all word vectors of the sentence pairs P and Q obtained in step S100 into the first cyclic neural network in order to obtain a context information vector, wherein the last context information vector of the sentence represents the sentence vector of the sentence; S300, input the sentence vectors of the sentence pairs P and Q into the first-level selection gate to obtain core information features; S400, input the core information obtained in step S300 into the secondary selection gate, and obtain the core information feature again; S500. Input the core information obtained in step S400 into the multi-angle semantic matching network, wherein the multi-angle semantic matching network includes four methods: full matching, maximum pooling matching, attention matching and maximum attention matching, to obtain sentence pairs feature matching vector; S600. Pass the matching vector obtained in step S500 through the second neural network to fuse the feature matching vector into a fixed-length vector, and input it to the prediction layer to calculate the probability distribution of the similarity of sentence pairs. 2 . The semantic similarity feature extraction method based on double selection gate according to claim 1 , wherein the first recurrent neural network is used to generate a state vector of context information. 3 . 3. The semantic similarity feature extraction method based on double selection gate according to claim 1, is characterized in that, the first layer of the first recurrent neural network is a single-item long-short-term memory network, and the second layer is a bidirectional long-short-term memory network A network, each hierarchical structure consists of multiple connected LSTM cell modules. 4. the semantic similarity feature extraction method based on double selection gate according to claim 3, is characterized in that, The first recurrent neural network includes two hierarchical structures; The first layer of the first recurrent neural network is used to generate word-level vectors; The second layer of the first recurrent neural network is used to generate context information vectors. 5. The semantic similarity feature extraction method based on double selection gate according to claim 1, characterized in that, the primary selection gate and the secondary selection gate respectively comprise a plurality of primary selection gate units and secondary selection gates unit; 6. The semantic similarity feature extraction method based on double selection gate according to claim 3, is characterized in that, In the step S200, all the word vectors of the sentence pair obtained in the step S100 are input into the first cyclic network in sequence, so as to obtain the sentence state vector after each word is input, specifically: Input the i-th word vector and the output word vector at the i-1th time into the i-th LSTM cell module, and obtain the state of the sentence after the i-th word vector is processed by the i-th LSTM cell module vector. 7. The semantic similarity feature extraction method based on double selection gate according to claim 5, is characterized in that, In the step S300, the sentence vector of the sentence pair is input into the first-level selection gate, and the acquisition of core information features includes: The context information vector at each moment of sentence P and the i-th sentence vector of sentence Q are input into the first-level selection gate unit, and the core information is obtained after processing by the i-th first-level selection gate unit. 8. the semantic similarity feature extraction method based on double selection gate according to claim 1-7, is characterized in that, In step S400, the core information obtained in step S300 is input into the secondary selection gate, and the characteristics of the core information obtained again include: The core information processed by the i-th first-level selection gate unit is input into the i-th second-level selection gate unit, and the core information features are obtained through processing by the i-th second-level selection gate unit. 9. The semantic similarity feature extraction method based on double selection gate according to claim 1-8, wherein in the step S500, the core information obtained in the step S400 is input into the multi-angle semantic matching network, The obtained feature matching vector includes: The full matching performs cosine similarity calculation on the context information vector described in sentence P and the sentence vector described in sentence Q at each moment to obtain a feature matching vector; The maximum pooling matching performs cosine similarity calculation on the context information vector at every moment of sentence P and the context information vector at every moment of sentence Q, and selects the maximum value as a feature matching vector; The attention matching performs cosine calculation on the context information vector at the ith moment of sentence P and the context information vector at the ith moment of sentence Q, respectively, to obtain i cosine values of sentence P, and weight the i cosine values as attention. The force weight is multiplied with the context information described in each moment of sentence Q, and the obtained result is then subjected to cosine calculation with the context information vector described in each moment of sentence P to obtain a feature matching vector; The maximum attention matching performs cosine calculation on the context information vector at the i-th moment of sentence P and the context information vector at the i-th moment in sentence Q, respectively, to obtain i cosine values of sentence P, from the i cosine values. The largest value is selected as the attention weight and multiplied by the context information of sentence Q, and the result obtained is then cosine calculation with the context information vector described in sentence P at each moment to obtain the feature matching vector. 10. The semantic similarity feature extraction method based on the double selection gate according to claim 1-9, wherein the step S600 passes the matching vector obtained in the step S500 through the second neural network, so that the feature matching vector is fused into a A fixed-length vector and input to the prediction layer to calculate the probability distribution of the similarity of sentence pairs including: The four feature matching vectors obtained by the four matchings of the sentence P are aggregated into a fixed-length feature matching vector through the second cyclic neural network; The four feature matching vectors obtained by the four matchings of the sentence Q are aggregated into a fixed-length feature matching vector through the bidirectional long-short-term memory network; The two feature matching vectors of sentence P and sentence Q are input to the prediction layer to obtain the sentence pair similarity.

Images