CN111552773A - A method and system for finding key sentences of question-like or not in reading comprehension task - Google Patents

Abstract

The method and system for finding whether or not a question-like key sentence in a reading comprehension task provided by the present invention includes: selecting the existing reading comprehension question and answer data, preprocessing the question and answer data to obtain a data set, and then mining the data set based on the coding layer network. The semantic information of sentences in sentences and chapter paragraphs is obtained, and the word embedding representation of each word is obtained; and an algorithm model is constructed, and the question sentences and chapter paragraphs mined through the coding layer network are used to calculate whether the question is a question or not by using the neural network model and TFIDF. Key sentence; input the question and answer data to be read into the trained algorithm model, and predict the key sentence of whether it is a question or not. The present invention can provide more key sentence support, calculate the weight of the key sentence through the combination of the bidirectional gated cyclic network and TF-IDF, and improve the answering efficiency and accuracy of whether or not the question is a question.

Description

A method and system for finding key sentences of question-like or not in reading comprehension task

technical field

The invention belongs to the technical field of processing natural language data, and in particular relates to a method and system for finding key sentences of whether or not a question is in a reading comprehension task.

Background technique

With the explosive growth of network information, all kinds of information flood the entire network environment. People are now accustomed to go to the Internet to find some solutions to their problems. When users are not very familiar with some search techniques, they often need to spend a lot of time filtering the results returned by search engines. The birth of reading comprehension system effectively solves the problem of complicated information mentioned above. The reading comprehension system uses natural language processing to analyze the questions submitted by users, obtain relevant answers and return them to users.

The search for key sentences of question-like or not has always been a key issue in reading comprehension tasks. How to find the key sentence of the question is a key point to answer the question. It is also a key point in evaluating whether this reading comprehension system is excellent. The search for key sentences of whether or not is a project that integrates natural language processing and natural language understanding. How to automatically and correctly find the key sentence of the question or not without human participation, and how to correctly answer the key sentence is a common problem faced by reading comprehension researchers at present. If the above problems can be solved correctly, the reading comprehension question answering system will be applied to all aspects of life. At the same time, we found that due to the lack of the support of relevant questions in the current reading comprehension question answering system, the content of the reading comprehension question answering system is too single, which is not suitable for the needs of current people. At the same time, the accuracy rate of finding key sentences related to the question is too low, so that the model cannot find the key information matching the question when answering the question, and the scalability of the traditional method is poor.

SUMMARY OF THE INVENTION

In order to solve the problem of the low accuracy rate of the existing traditional method in finding key sentences related to the problem, the model cannot find the key information matching the problem when answering the question, and the scalability of the traditional method is poor. Sentence query field limitations. To this end, the present invention proposes a method for finding key sentences of question-like or not in a reading comprehension task, including:

Selecting the existing reading comprehension question and answer data, preprocessing the question and answer data to obtain a data set, and then dividing the data set into a training set, a verification set and a test set;

Based on the constructed coding layer network, mine the semantic information of the sentences in the questions and paragraphs in the training set and the test set, and obtain the word embedding representation of each word in the sentence;

Build an algorithm model, and use the neural network model and TFIDF to calculate the key sentences of whether or not the question is a question or not.

Input the data to be read comprehension question-and-answer data into the trained algorithm model, and predict the key sentences of whether or not in the to-be-read comprehension question and answer data.

Preferably, preprocessing the question and answer data includes:

Perform multi-scale and multi-fine-grained small sample data processing on the question and answer data;

The data processing also includes the processing of web pages, website addresses, pictures, and dirty words.

Preferably, the data set is divided into training set, validation set and test set, including:

The data set is divided based on relevant semantic understanding, and 80% of the samples are randomly selected as the training set; 10% of the samples are selected as the verification set; and the remaining 10% of the samples in the training data are divided into the test set;

At the same time, generate word vectors that meet the needs of the model;

The segmentation includes data segmentation corrected for the original article paragraphs and questions.

Preferably, the key sentence of whether the question is a question or not is calculated by using the neural network model and TFIDF to calculate the question sentence and chapter paragraph after mining through the coding layer network, including:

Input the questions and paragraphs mined through the coding layer network into the neural network for training;

At the same time, TFIDF calculation is performed on questions and paragraphs;

Multiply the weight of each word obtained by tf-idf by the final vector representation of each sentence output by the neural network, and then calculate the cosine similarity;

The key sentence with the highest similarity is selected as the key sentence of the whether question sentence.

Preferably, the final vector representation of each sentence obtained by multiplying the weight of each word obtained by tf-idf by the output of the neural network includes:

The tf-idf weights of each word in question sentences and discourse paragraphs are combined with the contextual features learned through the bidirectional gated recurrent unit GRU network to obtain the final vector representation of each sentence.

Preferably, the word embedding is expressed as follows:

w _i =word2vec(t _i )

where w _i is the word embedding representation of the ith word t _i in the question or text, and word2vec() is the word embedding formula

Preferably, the calculation formula for learning the context feature through the bidirectional gated recurrent unit GRU network is as follows:

为正序向量，

为逆序向量。Among them, GRU is the gated recurrent unit network, hi is the hidden layer vector output by the gated recurrent unit network of the _ith word, ";" is the splicing operation, _wi is the question or the ith word t _i in the text The word embedding representation of , h _i-1 is the hidden layer vector output by the gated recurrent unit network of the word before the i-th word, h _i+1 is the hidden layer vector output by the gated recurrent unit network of the i-th word,

is a positive sequence vector,

is a vector in reverse order.

Preferably, the TFIDF calculation formula for the question and the paragraph is as follows:

tfidf(i,j)=tf(i,j)×idf(i)

where tf(i,j) represents the word frequency of word i in document j, where n _i,j represents the number of occurrences of word i in document j, ∑ _k n _i,j is the total number of words in document j, idf(i ) is divided by the total number of files D by the number of files containing the word i, t _i is the word i; d _j is the file j containing the word; and then the obtained quotient is obtained by taking the logarithm.

Preferably, while inputting the mined questions and text paragraphs into the neural network for training, perform TFIDF calculation on the questions and text paragraphs to obtain the most suitable sentence as the key sentence of the question whether it is a class or not, and then include:

Correct the predicted results to obtain correct results;

Save the corrected results, use the corrected data to train the model, and update the algorithm model.

Based on the same inventive concept, the present invention also proposes a system for finding question-like key sentences in reading comprehension tasks, including:

The data preparation module is used to select the existing reading comprehension question and answer data, preprocess the question and answer data to obtain a data set, and then divide the data set into a training set, a verification set and a test set;

The word and sentence encoding module is used to mine the semantic information of the sentences in the questions and paragraphs in the training set and the test set based on the constructed encoding layer network, and obtain the word embedding representation of each word in the sentence;

The key sentence matching module is used to calculate the key sentence of whether the question is a question or not by using the neural network model and TFIDF to calculate the question sentence and chapter paragraph after mining through the coding layer network;

The key sentence determination module is used for inputting the question and answer data to be read comprehension into the trained algorithm model, and predicts the key sentence of the question whether or not in the question and answer data to be read comprehension.

The beneficial effects of the present invention are:

The invention provides a method and system for finding key sentences of whether or not a question in a reading comprehension task, comprising: selecting existing reading comprehension question and answer data, preprocessing the question and answer data to obtain a data set, and then dividing the data set into Training set, validation set and test set; based on the constructed coding layer network, mine the semantic information of the sentences in the questions and paragraphs in the training set and test set, and obtain the word embedding representation of each word in the sentence; build the algorithm model, The question sentences and chapter paragraphs mined through the coding layer network are calculated by using the neural network model and TFIDF to obtain the key sentences of whether or not the question is a question; input the question-and-answer data to be read into the trained algorithm model, and the question-and-answer data to be read comprehension To predict the key sentences of the whether or not problem, more key sentence supports can be provided through the technical solution provided by the present invention. At the same time, because the present invention does not need manual annotation, it is more suitable for application in real scenes in life, and improves the efficiency and accuracy of the question answering system.

In the technical solution provided by the present invention, the method for finding key sentences in the reading comprehension task based on neural network and TF-IDF calculates the weight of the key sentences by combining the bidirectional gated cyclic network and TF-IDF, so as to further improve the problem of whether or not The efficiency of answering, greatly improves the correct rate of whether or not. It not only greatly reduces manual annotation, but also improves the accuracy of reading comprehension.

Description of drawings

Fig. 1 is a kind of reading comprehension task provided by the present invention is a kind of question key sentence search method flow chart;

Fig. 2 is based on tf-idf of the present invention whether the key sentence of class question is searched for method steps schematic diagram;

FIG. 3 is a structural diagram of a system for finding key sentences in a reading comprehension task provided by the present invention.

Detailed ways

The object of the present invention is to provide a method for finding key sentences in the reading comprehension task based on neural network and TF-IDF, which is to search for key sentences through the combination of bidirectional gated recurrent network and TF-IDF, wherein TF -IDF (term frequency-inverse document frequency) is a weighting technique for information retrieval and data mining. TF means Term Frequency, and IDF means InverseDocument Frequency.

The present invention will be described in detail below with reference to specific embodiments.

Example 1:

As shown in Figure 1, the present invention provides a method for finding a key sentence of a question or not in a reading comprehension task, which is carried out according to the following steps:

S1 selects the existing reading comprehension question and answer data, preprocesses the question and answer data to obtain a data set, and then divides the data set into a training set, a verification set and a test set;

Based on the constructed coding layer network, S2 mines the semantic information of the sentences in the questions and paragraphs in the training set and the test set, and obtains the word embedding representation of each word in the sentence;

S3 builds an algorithm model, and uses the neural network model and TFIDF to calculate the key sentences of whether the question is a question or not by using the question sentences and chapter paragraphs mined by the coding layer network;

S4 inputs the question-and-answer data to be read comprehension into the trained algorithm model, and predicts the key sentences of whether or not in the question-and-answer data to be read comprehension.

The technical scheme that the present invention adopts as shown in Figure 2 is to carry out according to the following steps:

Step 1. Select the existing reading comprehension question and answer data, and preprocess the data.

The selected reading comprehension question-and-answer data have the same attributes, such as the same data format, the same professional direction, and whether it is a question or not. Data preprocessing includes multi-scale and multi-fine-grained small sample data processing on reading comprehension question and answer data, including sentence segmentation, word segmentation, regularization, cleaning, etc. It also includes processing of web pages, URLs, pictures, and dirty text. Its purpose is to optimize and expand existing data.

Step 2: Split the training set and the test set, perform pre-training of the embedding layer model, and generate word vectors that meet the needs of the model.

Further, the data set is divided based on relevant semantic understanding, and 80% of the samples are randomly selected as the training set; 10% of the samples are selected to be divided into the verification set; and the remaining 10% of the samples in the training data are divided into the test set. At the same time, word vectors that meet the needs of the model are generated; the segmentation includes the data segmentation after correcting the original article paragraphs and questions. The data set is segmented based on relevant semantic understanding, so that the class information in the training set and the test set can achieve a perfect segmentation ratio, which greatly improves the training speed and convergence rate.

Step 3. The coding layer network mines the semantic information of the sentences in the questions and the paragraphs.

The coding layer network uses the word embedding method to express the word embedding for the question and the paragraph;

w _i =word2vec(t _i )

Where w _i is the word embedding representation of the ith word t _i in the question or text, and word2vec() is the word embedding formula.

Step 4. Carry out model improvement based on TF-IDF fusion, including: while inputting questions and chapter paragraphs into the neural network, perform TFIDF calculation on questions and chapter paragraphs. It is used to control the weight of each sentence in the question sentence and the text paragraph, and then calculate the cosine similarity between the question sentence and the text paragraph, and obtain the most suitable sentence as the key sentence of the question.

specific,

Input the questions and paragraphs mined through the coding layer network into the neural network for training;

At the same time, TFIDF calculation is performed on questions and paragraphs;

Multiply the weight of each word obtained by tf-idf by the final vector representation of each sentence output by the neural network, and then calculate the cosine similarity;

The key sentence with the highest similarity is selected as the key sentence of the whether question sentence.

Here, the weight of each word obtained by tf-idf is multiplied by the final vector representation of each sentence output by the neural network, including: combining the tf-idf weight of each word of the question and chapter paragraph with the bidirectional gated recurrent unit The GRU network learns the contextual features to obtain the final vector representation of each sentence.

The calculation formula for learning contextual features through the bidirectional gated recurrent unit GRU network is as follows:

为正序向量，

为逆序向量。Among them, GRU is the gated recurrent unit network, hi is the hidden layer vector output by the gated recurrent unit network of the _ith word, ";" is the splicing operation, _wi is the question or the ith word t _i in the text The word embedding representation of , h _i-1 is the hidden layer vector output by the gated recurrent unit network of the word before the i-th word, h _i+1 is the hidden layer vector output by the gated recurrent unit network of the i-th word,

is a positive sequence vector,

is a vector in reverse order.

The TFIDF calculation formula for questions and chapter paragraphs is as follows:

tfidf(i,j)=tf(i,j)×idf(i)

Among them, tf(i, j) represents the word frequency of word i in document j, where n _{i, j} represents the number of occurrences of word i in document j, ∑ _k n _{i, j} is the total number of words in document j, idf(i ) is divided by the total number of files D by the number of files containing the word i, t _i is the word i; d _j is the file j containing the word; and then the obtained quotient is obtained by taking the logarithm.

Step 5. Correct the predicted result to obtain the correct result.

The abnormal generated results are discarded, not used for subsequent updates and upgrades, and the normal questions are generated.

Step 6: Save the corrected result, use the corrected data to train the model, and update the algorithm model. Using the corrected algorithm model greatly improves the training speed and accuracy.

Step 7: Input the new question-and-answer data to be read and comprehend, and predict the key sentences of whether or not in the data.

First, analyze whether its attributes are the same as the data attributes of the reading comprehension question and answer data selected in step 1. If the attributes are different, try to make them consistent with the data attributes selected in step 1 in the subsequent preprocessing operation. Then, preprocessing is performed on the reading comprehension question answering data text.

Example 2

Given questions and sentences, is Beijing the capital of China? Beijing is the capital of China

Participle: <Beijing> <whether><is><China><the><capital><Beijing><is><China><the><capital>

Feed it into the model and obtain the word embedding representation through neural network and TF-IDF

Calculate the cosine similarity between a question sentence and a text paragraph

Return the result, the sentence is the key sentence of the question sentence

Given questions and sentences, is Beijing the capital of China? This winter vacation has been extended to November

Participle: <Beijing><Whether><Yes><China><The><Capital><This Year><Winter Holiday><Extended><To><November>

Feed it into the model and obtain the word embedding representation through neural network and TF-IDF

Calculate the cosine similarity between a question sentence and a text paragraph

Return the result, the sentence is not the key sentence of the question sentence

Further, model update can choose automatic update upgrade or manual one-click update upgrade.

Example 3

In order to realize the above-mentioned method, the present invention also proposes a system for finding key sentences of question or not in the reading comprehension task, as shown in the figure, including:

The data preparation module is used to select the existing reading comprehension question and answer data, preprocess the question and answer data to obtain a data set, and then divide the data set into a training set, a verification set and a test set;

The word and sentence encoding module is used to mine the semantic information of the sentences in the questions and paragraphs in the training set and the test set based on the constructed encoding layer network, and obtain the word embedding representation of each word in the sentence;

The key sentence matching module is used to calculate the key sentence of whether the question is a question or not by using the neural network model and TFIDF to calculate the question sentence and chapter paragraph after mining through the coding layer network;

The key sentence determination module is used for inputting the question and answer data to be read comprehension into the trained algorithm model, and predicts the key sentence of the question whether or not in the question and answer data to be read comprehension.

Specifically, preprocessing the question and answer data includes:

Perform multi-scale and multi-fine-grained small sample data processing on the question and answer data;

The data processing also includes the processing of web pages, website addresses, pictures, and dirty words.

Split the dataset into training, validation, and test sets, including:

The data set is divided based on relevant semantic understanding, and 80% of the samples are randomly selected as the training set; 10% of the samples are selected as the verification set; and the remaining 10% of the samples in the training data are divided into the test set;

At the same time, generate word vectors that meet the needs of the model;

The segmentation includes data segmentation corrected for the original article paragraphs and questions.

Using the neural network model and TFIDF to calculate the question sentences and chapter paragraphs mined by the coding layer network, the key sentences of whether or not the question are classified, including:

Input the questions and paragraphs mined through the coding layer network into the neural network for training;

At the same time, TFIDF calculation is performed on questions and paragraphs;

Multiply the weight of each word obtained by tf-idf by the final vector representation of each sentence output by the neural network, and then calculate the cosine similarity;

The key sentence with the highest similarity is selected as the key sentence of the whether question sentence.

Multiply the resulting per-word weights from tf-idf by the neural network's output for the final vector representation of each sentence, including:

The tf-idf weights of each word in question sentences and discourse paragraphs are combined with the contextual features learned through the bidirectional gated recurrent unit GRU network to obtain the final vector representation of each sentence.

The word embedding is expressed as:

w _i =word2vec(t _i )

Where w _i is the word embedding representation of the ith word t _i in the question or text, and word2vec() is the word embedding formula

The calculation formula for learning contextual features through the bidirectional gated recurrent unit GRU network is as follows:

Among them, GRU is the gated recurrent unit network, hi is the hidden layer vector output by the gated recurrent unit network of the _ith word, ";" is the splicing operation, ω _i is the question or the ith word t _i in the text word embedding representation of .

The TFIDF calculation formula for questions and chapter paragraphs is as follows:

tfidf(i,j)=tf(i,j)×idf(i)

where tf(i, j) represents the word frequency of word i in document j, where n _{i, j} represents the number of occurrences of word i in document j, ∑ _k n _{i, j} is the total number of words in document j, idf(i ) is divided by the total number of files D by the number of files containing the word i, t _i is the word i; d _j is the file j containing the word; and then the obtained quotient is obtained by taking the logarithm.

While inputting the mined questions and chapter paragraphs into the neural network for training, the TFIDF calculation is performed on the questions and chapter paragraphs to obtain the most suitable sentence as the key sentence for the question whether it is a class or not, and then includes:

Correct the predicted results to obtain correct results;

Save the corrected results, use the corrected data to train the model, and update the algorithm model.

The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Any simple modifications, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention belong to the present invention. within the scope of the technical solution of the invention.

Claims (10)

1. A method for searching key sentences of whether problems are similar or not in a reading and understanding task is characterized by comprising the following steps:

selecting existing reading understanding question and answer data, preprocessing the question and answer data to obtain a data set, and then dividing the data set into a training set, a verification set and a test set;

based on the constructed coding layer network, mining semantic information of sentences in question sentences and chapter paragraphs in a training set and a test set to obtain word embedded representation of each word in the sentences;

constructing an algorithm model, and calculating the question and chapter paragraphs mined by the coding layer network by using a neural network model and a TFIDF (fuzzy finite field Format) to obtain a key sentence of whether the question is similar to the question;

inputting the question-answer data to be read into the trained algorithm model, and predicting whether key sentences similar to problems in the question-answer data to be read are predicted.

2. The key sentence searching method according to claim 1, wherein the preprocessing of the question-answer data includes:

processing the question-answer data by using a small sample with multiple scales and multiple fine granularities;

the data processing also comprises the processing of web pages, websites, pictures and dirty and messy characters.

3. The key sentence finding method of claim 1, wherein the segmenting the data set into a training set, a validation set, and a test set comprises:

segmenting the data set based on relevant semantic understanding, and randomly selecting 80% samples as a training set; selecting 10% of samples to be divided into verification sets; dividing the data and the rest 10% of samples in the training data into a test set;

generating word vectors meeting the requirements of the model;

the segmentation includes data segmentation after correcting the original chapter paragraphs and the problem.

4. The method for finding key sentences according to claim 1, wherein the calculating the question sentences and chapter paragraphs mined by the coding layer network to obtain the key sentences of question-or-not classes by using the neural network model and TFIDF comprises:

inputting the question sentences and discourse paragraphs mined by the coding layer network into a neural network for training;

simultaneously performing TFIDF calculation on the question and the chapter paragraphs;

multiplying each word weight obtained by tf-idf by the final vector representation of each sentence output by the neural network, and then performing cosine similarity calculation;

and selecting the key sentence with the highest similarity as the question sentence whether to be classified.

5. A key sentence finding method as claimed in claim 4, wherein the multiplying the resulting per word weight of tf-idf by the final vector representation of each sentence of the output of the neural network comprises:

and (3) learning the context characteristics by the tf-idf weight of each word of the question and chapter paragraphs and the bidirectional gating circulation unit GRU network to obtain the final vector representation of each sentence.

6. The key sentence searching method of claim 4, wherein the word embedding expression is as follows:

w_i＝word2υec(t_i)

wherein w_iFor the ith word t in question or text_iWord2 υ ec () is a word embedding formula.

7. The key sentence finding method of claim 6, wherein the learning of the context feature through the bi-directional gated round-robin unit GRU network is calculated as follows:

wherein GRU is a gated cyclic unit network, h_iHidden layer vector output by gated cyclic unit network for ith word, h_i-1Hidden layer vector output by gated cyclic unit network of word preceding ith word, h_i+1The hidden layer vector output by the gated round-robin unit network for the ith word,

is a vector of a positive sequence, and is,

is the reverse order vector, "; "is a splicing operation, w_iFor the ith word t in question or text_iThe word embedded representation.

8. The method of claim 7, wherein TFIDF is calculated for question and chapter paragraphs as follows:

tfidf(i，j)＝tf(i，j)×idf(i)

where tf (i, j) represents the word frequency of word i in document j, where n_i，jRepresenting the number of occurrences of word i in document j, ∑_kn_i,jThe total number of words in document j, idf (i) is the total number of files D divided by the number of files containing the word i, t_iIs the word i; d_jIs a file j containing the word; and taking the logarithm of the obtained quotient to obtain the product.

9. The method for finding key sentences according to claim 1, wherein TFIDF calculation is performed on the question and the chapter paragraphs while inputting the mined question and chapter paragraphs into the neural network for training to obtain the most suitable sentence as the key sentence for question classification, and then further comprising:

correcting the predicted result to obtain a correct result;

and storing the corrected result, training the model by using the corrected data, and updating the algorithm model.

10. A system for searching key sentences whether to be similar to problems in reading and understanding tasks is characterized by comprising the following steps:

the data preparation module is used for selecting the existing reading comprehension question and answer data, preprocessing the question and answer data to obtain a data set, and then dividing the data set into a training set, a verification set and a test set;

the word and sentence coding module is used for mining semantic information of the sentences in the question sentences and the chapter paragraphs in the training set and the test set based on the constructed coding layer network to obtain word embedded representation of each word in the sentences;

the key sentence matching module is used for calculating the question sentences and chapter paragraphs mined by the coding layer network by utilizing a neural network model and a TFIDF (trivial text field function) to obtain key sentences of which the question types are not similar;

and the key sentence determining module is used for inputting the question and answer data to be read into the trained algorithm model and predicting the key sentences which are similar to the questions in the question and answer data to be read.

Images