CN113239703A - Deep logical reasoning financial text analysis method and system based on multivariate factor fusion - Google Patents

Deep logical reasoning financial text analysis method and system based on multivariate factor fusion Download PDF

Info

Publication number
CN113239703A
CN113239703A CN202110562999.9A CN202110562999A CN113239703A CN 113239703 A CN113239703 A CN 113239703A CN 202110562999 A CN202110562999 A CN 202110562999A CN 113239703 A CN113239703 A CN 113239703A
Authority
CN
China
Prior art keywords
vector
text
vectors
memory
financial
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110562999.9A
Other languages
Chinese (zh)
Other versions
CN113239703B (en
Inventor
李鑫
王智
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen International Graduate School of Tsinghua University
Original Assignee
Shenzhen International Graduate School of Tsinghua University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen International Graduate School of Tsinghua University filed Critical Shenzhen International Graduate School of Tsinghua University
Priority to CN202110562999.9A priority Critical patent/CN113239703B/en
Publication of CN113239703A publication Critical patent/CN113239703A/en
Application granted granted Critical
Publication of CN113239703B publication Critical patent/CN113239703B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Machine Translation (AREA)

Abstract

The invention discloses a deep logical reasoning financial text analysis method and system based on multi-factor fusion. Through the three modules, after a financial text and a plurality of external features enter the model proposed by the method, text coding, memory storage and information interactive understanding are sequentially carried out, and deep information contained in the text is extracted step by step through three processes, so that a reasonable inference is made.

Description

Deep logical reasoning financial text analysis method and system based on multivariate factor fusion
Technical Field
The invention relates to financial text analysis, in particular to a deep logical reasoning financial text analysis method and system based on multi-factor fusion.
Background
In the aspect of financial text analysis, the mainstream technology has been shifted from the conventional method mainly for feature extraction to the method mainly for neural network driving, and the traditional features, such as emotional word counting, key segment matching, text stream reconstruction and the like, need to consume too much manpower in the construction process, and need to reconstruct features for texts in different fields, so that the mobility is insufficient. The neural network-based method can well complete the financial text analysis task under a specific scene only by training on specific data and consuming computing power resources, so the neural network-based method is a mainstream technology in recent years.
At present, more work focuses on the selection level of the neural network in the aspect of application and research in financial texts, and such methods focus on the representation learning of texts by using the neural network, learn vectorized representations of texts by designing the neural network, and expect to find out the neural network capable of better realizing text modeling, but the methods have the following two problems:
1) these methods focus on representation learning of the text surface layer, ignoring reasoning on text semantics. However, the financial text analysis needs to help financial services, so when building a neural network model, attention should be paid to the capability of the designed calculation method on semantic understanding and logical reasoning.
2) The financial phenomenon is generated by the combined influence of a plurality of external factors, the inference is easy to fall into local optimum by using a single factor, the existing work focuses more on the text per se, the inference is carried out by using the single text factor, and the attention to some potential external factors is not enough, so that the model can form certain one-sidedness when the inference is carried out by drawing a conclusion, and the data mixed with noise lacks enough generalization.
Disclosure of Invention
In order to make up for the defects of the prior art, the invention provides a deep logical reasoning financial text analysis method and system based on multi-factor fusion, which extracts deep information contained in a text so as to make a reasonable inference.
The technical problem of the invention is solved by the following technical scheme:
a deep logic reasoning financial text analysis method based on multi-element fusion is characterized by comprising the following steps: s1, converting the input text into semantic vector representation; s2, semantic parsing is carried out on the text through a coding module; s3, converting the external factors into memory vectors; and S4, calculating the cross representation form of the text vector and the external factors by adopting a cross attention module, and re-expressing the cross representation form into a higher-level semantic feature to realize abstract understanding of semantics.
In some embodiments, the following features are also included:
the step S1 includes: s1-1, converting the input text into word embedding vectors;
using a dual embedding vector as an initial said word embedding vector; the dual embedding vector includes two types: generic word embedding vectors and domain-specific word embedding vectors, which are distinguished from each other by training the word embedding vectors in a corpus of specified domains.
The general word embedding vector is trained according to internet text data, the specific field word embedding vector is trained by utilizing the financial text data set constructed by the method, and the two vectors are spliced together according to specified dimensions.
During training, data are read according to batches, and the lengths of all texts are unified for the data in the same batch.
In step S2, the text is semantically parsed by an encoding module, where the encoding module is composed of a stacked cyclic neural network, the stacked cyclic neural network is a four-layer neural network structure, and each layer is composed of a bidirectional stack flow gate control unit.
The main body of the cascade control unit is composed of two gate structures, the flow directions of the input vector and the hidden layer vector are controlled by the two gate structures, and semantic information understanding in the text is realized by adjusting the information interaction process of the two vectors.
The calculation flow of the stack flow control unit comprises the following steps: given an input vector x for a current time ttAnd memory cell c of last time period t-1t-1Memory cell of current time periodctCalculated from the following formula:
Figure BDA0003079760890000021
Figure BDA0003079760890000022
Figure BDA0003079760890000023
wherein, WiIs a trainable parameter of the neural network, tanh is a hyperbolic tangent function, which is a vector point product operation. f. oftThe output of the forgetting gate is used for forgetting a part of unimportant historical memory information in the SFGU unit, and the function of compressing information flow is achieved.
In step S3, the memory module with slots is used to encode external factors and store the encoded external factors in a memory unit for use by a following module; in step S4, when calculating the high-level vector representation, weights are reassigned to each vector according to the importance of each position.
The invention also proposes a deep logical reasoning financial text analysis system based on multivariate factor fusion, characterized in that it comprises a processor and a memory, said memory having stored therein a computer program executable by the processor for implementing the method according to any one of claims 1 to 9.
Compared with the prior art, the invention has the advantages that:
the deep logical reasoning model based on the multi-dimensional data fusion provided by the method respectively realizes semantic analysis, multi-factor fusion and deep logical reasoning through three different mechanisms, namely a stacking recurrent neural network, a slot memory module and a cross attention module. Through the three modules, after a financial text and a plurality of external features enter the model proposed by the method, text coding, memory storage and information interactive understanding are sequentially carried out, and deep information contained in the text is extracted step by step through three processes, so that a reasonable inference is made.
Drawings
FIG. 1 is a diagram of a deep logical inference model based on multi-metadata fusion according to an embodiment of the present invention.
FIG. 2 is a schematic diagram of a stack flow control unit according to an embodiment of the invention.
Detailed Description
The invention will be further described with reference to the accompanying drawings and preferred embodiments. It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict.
It should be noted that the terms of orientation such as left, right, up, down, top and bottom in the present embodiment are only relative concepts to each other or are referred to the normal use state of the product, and should not be considered as limiting.
Aiming at the defects of the prior art, the method of the following embodiment of the application designs a new technical scheme, and the method is used for mining logical connection from a cause-and-effect evidence-based relation chain from the perspective of utilizing various external information, and performing deep logical reasoning on financial texts, thereby obtaining a reliable and interpretable financial conclusion. The method designs a brand-new interactive modeling mode of the external factors and the text, and further performs deep logical reasoning in an interactive fusion mode to solve the problem of cooperative understanding of the external factors and the text.
The deep logical reasoning model based on multivariate data fusion provided by the method mainly comprises three modules: the connection relation and the technical details of the coding module, the slot memory module and the cross attention module are shown in figure 1.
1. Problem definition
The method is focused on solving the financial text analysis task of multi-element factor fusion, aims to give the multi-element external factors, and carries out deep logical reasoning according to the external factors to obtain a final analysis conclusion. Specifically, a financial text X containing n words is given, i.e., X ═ w i1, 2, 3, …, n, given m external factors T ═ T1，t2，t3，…，tmWhere tiThe method can be a Chinese character string or a number, then deducing a conclusion Y according to X and T, wherein Y belongs to { -1, 0, 1}, the process is marked as Y ═ f (X, T) by using a mathematical notation, and the method designs a neural network model M which models an input<X，T>Mapping to output Y.
2. Technical scheme
1) Coding module
The encoder module is responsible for converting the input text into semantic vector representation, thereby achieving the semantic parsing function of the text. The first step of converting semantic vectors is to convert the input text into word-embedded vectors. The word embedding vector is a general technical scheme in the field of natural language processing and is realized by a word2vec algorithm. The method uses a dual embedding vector as the initial word embedding vector. Dual embedded vectors contain two types: the general word embedding vector and the specific field word embedding vector are distinguished by training a corpus in a specified field according to the word embedding vector. On the mathematical symbolic representation, each word wiWill use the feature vector
Figure BDA0003079760890000041
Carrying out initialization in which dGAnd dDRespectively universal word embedded vector
Figure BDA0003079760890000042
And domain-specific word embedding vector
Figure BDA0003079760890000043
Is measured in the first dimension. | V | represents the size of the vocabulary. Thus the final input word is embedded in the vector representation
Figure BDA0003079760890000044
Calculated from the following formula:
Figure BDA0003079760890000045
wherein
Figure BDA0003079760890000051
And representing a vector splicing operation, and splicing the two vectors together according to the specified dimension. The method uses a 300-dimensional Glove word embedded vector as an initialized word vector matrix, a general word embedded vector is trained according to internet text data, and a specific field word embedded vector is trained by using a financial text data set constructed by the method.
When obtaining word-embedded vectors for the entire input text, there are special cases that need to be handled: since the input text may contain words other than vocabulary V, the method uses a special identifier < unk > for all words not in vocabulary, and generates a word-embedded vector for this identifier by random initialization.
In addition, when the neural network model is trained, data is read according to batches, and for data in the same batch, the lengths of the texts need to be unified. The unified mode is as follows: all texts in the same batch are taken, the maximum sentence length L is obtained, and the longest sentence length L of one text is presetmaxFor excess length LmaxPerforming post-truncation, and discarding the text with the maximum length LmaxPart (c) of (a). If the text length is smaller than L, the text needs to be filled, and the zero vector is used for filling at the rear end of the text, so that the filled text length is ensured to reach L. With the above word-embedded vector conversion, the original input word sequence X ═ w i1, 2, 3, …, n } is converted into a word-embedded vector sequence
Figure BDA0003079760890000052
After the vector sequence of the text is obtained, the method carries out semantic analysis on the text through a coding module. The coding module is composed of a stacked recurrent neural networks (S-RNN), the S-RNN is a four-layer neural network structure, and each layer is composed of a bidirectional stacked flow gate control unit (SFGU). The SFGU body is shown in fig. 2.
The SFGU main body is composed of two gate structures, the flow directions of an input vector and a hidden layer vector are controlled by the two gate structures, and semantic information understanding in a text is realized by adjusting the information interaction process of the two vectors. The calculation flow of the overlap flow gate control unit is substantially the same as that of a gate control loop unit (GRU), but is modified to some extent as follows: given an input vector x for a current time ttAnd memory cell c of last time period t-1t-1Memory cell c of the current time periodtCalculated from the following formula:
Figure BDA0003079760890000053
Figure BDA0003079760890000054
Figure BDA0003079760890000055
wherein, WiIs a trainable parameter of the neural network, tanh is a hyperbolic tangent function, which is a vector point product operation. f. oftThe SFGU is the output of a forgetting gate, is used for forgetting a part of unimportant historical memory information in the SFGU unit, and plays a role in compressing information flow, and a specific calculation formula is shown as the following formula:
ft＝σ(Wfxt+Ufct-1)
wherein, WfAnd UfAre trainable parameters of the neural network and σ is the Sigmoid function. The foldover gate control unit has another output, namely the hidden layer vector h of the current time periodtThe calculation formula is shown as the following formula:
Figure BDA0003079760890000061
Figure BDA0003079760890000062
Figure BDA0003079760890000063
wherein s istThe output of the flow folding gate is used for reading input text information in the SFGU unit, and the information flow of the whole SFGU unit is dominated by a specific calculation formula shown as the following formula:
st＝σ(Wsxt+Usct-1)
wherein WsAnd UsAre trainable parameters of the neural network. The forgetting gate and the flow folding gate together form a complete SFGU unit, and the forgetting gate controls the information flow to be conducted from the last time segment to the current time segment and selectively filters partial information. The cascading valves control information to be transmitted from the SFGU unit of the current layer to the SFGU unit of the next layer, and the information of the input text is understood to play a role in abstracting text semantics.
Each layer of the stacked recurrent neural network follows a bi-directional vector coding approach, and this bi-directional representation is similar to a bidirectional long-short term memory network (LSTM). In the method, hidden layer vectors output in the forward direction and the reverse direction of the SFGU unit are spliced to be used as a final output vector H of a coding moduletThe formula is shown as follows:
Figure BDA0003079760890000064
wherein the content of the first and second substances,
Figure BDA0003079760890000065
and
Figure BDA0003079760890000066
the difference in the calculations is the flow of the input information: one flow is from left to right and one flow is from right to left. The method takes the last vector of the output vectors as a semantic vector representation C of the input text.
2) Slot memory module
The goal of the slot memory module is to set the external factor T to { T }1，t2，t3，…，tmThe step of encoding and then storing the encoded data in a memory unit for use by a following module is essentially to convert external factors into memory vectors. External information tiPossibly a single word or a sequence of words, the method handling both numbers and time as strings, and thus tiCan also be counted as ti＝{c i，j1, 2, 3, …, k }. The method uses an LSTM network to map tiConversion into a vector representation miThe specific formula is shown as the following formula:
mi＝LSTM(ci，j，hi-1)
wherein h isi-1Is the last hidden layer output of the LSTM, and uses a zero vector as the initialization vector of the LSTM, and the output of the last state of the LSTM is the memory vector mi. For each external factor, performing code conversion through the same LSTM network to obtain a series of memory vectors M ═ Mi|i＝1，2，3，…，m}。
3) Cross attention module
This section describes the cross attention module proposed by the present method. The basic idea of the attention mechanism is that when computing the high level vector representation, it will re-assign weights to each vector based on the importance of each location. Based on the basic idea, the method provides a cross attention module to calculate the cross representation form of the text vector and the external factors, and expresses the cross representation form of the text vector and the external factors into a higher-level semantic feature again to realize abstract understanding of semantics. Intuitively speaking, the semantic importance of words at different positions in the input text to a sentence is different, and if different external factors are looked at, the importance of different words applied to the input text is also different.
Given a series of memory vectors M ═ M i1, 2, 3, …, n and semantic vector representation C of the input text, and selecting the memory slice m which is most relevant to the text vector from the memory unit through the cross attention moduleiAnd outputting a continuous vector representation v, the output vector being obtained by weighted summation of the memory sequences, the specific calculation formula being shown as follows:
Figure BDA0003079760890000071
where k is the size of the memory sequence, aiIs a correlation weight coefficient with a value range of 0,1]And sigmaiai1. To calculate the correlation coefficient aiThe method designs a fully-connected neural network, which reads the memory vector miAnd a text vector C, and then calculating a similarity score, wherein the specific formula is shown as the following formula:
Figure BDA0003079760890000081
wherein WaIs a trainable parameter of the neural network and,
Figure BDA0003079760890000082
representing a vector stitching operation. When all the weight coefficients g of the memory sequence are obtained1，g2，…，gmThen, the method uses a Softmax function to normalize the weight coefficient to obtain a correlation coefficient aiThe specific calculation formula is shown as the following formula, where exp is an exponential function with a natural constant e as the base:
Figure BDA0003079760890000083
the cross-attention module has two advantages: one is that it can assign an importance score to each memory vector based on the semantic relevance; the other is that it is differentiable, so it can be easily trained in an end-to-end fashion with other neural network modules.
After obtaining the memory vector representation v, the method uses a full-link layer to perform final fusion on the text vector C and the memory vector representation, so as to make a final conclusion inference, and the specific formula is as follows:
p＝Softmax(Wo(v+C))
wherein WoIs a trainable parameter of the neural network, p is the final output of the model, and in the method, p is a three-dimensional vector, and each dimension represents a probability value of a corresponding label.
4) Training method
The deep layer reasoning model based on the multiple factors is an end-to-end neural network model, so the model is trained by minimizing the classified cross entropy error in a supervision training mode, and the loss function is shown as the following formula:
Loss＝-∑x∈D∑c∈Cy*log(M(X，T))
where D represents all training instances, C is the set of categories, (X, T) is the "text-factor set" pair entered, y is the correct label of the entered text, the value is 1 or 0, indicating whether the current category C is the correct answer, and p is the probability value predicted by the model.
The method uses a back propagation algorithm to calculate the gradients of all parameters, and updates the model parameters by a StochasticgradientDescription (SGD) optimizer. For all trainable parameters used in the neural network, initialized randomly by uniformly distributing U (-0.01, 0.01), the learning rate of the optimizer is set to 0.01.
In some variations, the slot memory module may be replaced with other types, the method uses a long-short term memory network as the main structure of the slot memory module, and word vector accumulation, a cyclic neural network and a gated cyclic unit may be alternatives to the slot memory module.
The foregoing is a more detailed description of the invention in connection with specific preferred embodiments and it is not intended that the invention be limited to these specific details. For those skilled in the art to which the invention pertains, several equivalent substitutions or obvious modifications can be made without departing from the spirit of the invention, and all the properties or uses are considered to be within the scope of the invention.

Claims (10)

1. A deep logic reasoning financial text analysis method based on multi-element fusion is characterized by comprising the following steps:
s1, converting the input text into semantic vector representation;
s2, semantic parsing is carried out on the text through a coding module;
s3, converting the external factors into memory vectors;
and S4, calculating the cross representation form of the text vector and the external factors by adopting a cross attention module, and re-expressing the cross representation form into a higher-level semantic feature to realize abstract understanding of semantics.
2. The method for deep logical inference financial text analysis based on multi-factor fusion as claimed in claim 1, wherein said step S1 comprises: the input text is first converted into word-embedded vectors.
3. The method of claim 2, wherein a double-embedded vector is used as the initial word-embedded vector; the dual embedding vector includes two types: generic word embedding vectors and domain-specific word embedding vectors, which are distinguished from each other by training the word embedding vectors in a corpus of specified domains.
4. The method of claim 3, wherein the generic word embedded vectors are trained according to Internet text data, and the domain-specific word embedded vectors are trained using the financial text data set constructed by the method, and the two vectors are spliced together according to a specified dimension.
5. The method as claimed in claim 4, wherein the training process reads the data in batches, and the lengths of the texts are unified for the data in the same batch.
6. The method according to claim 1, wherein in step S2, the text is semantically parsed by a coding module, wherein the coding module is composed of a stacked recurrent neural network, the stacked recurrent neural network is a four-layer neural network structure, and each layer is composed of a bidirectional stack flow control unit.
7. The method as claimed in claim 6, wherein the overlay flow control unit body is composed of two gate structures, the two gate structures control the flow direction of the input vector and the hidden layer vector, and the semantic information understanding in the text is realized by adjusting the information interaction process of the two vectors.
8. The method for deep logical inference financial text analysis based on multi-factor fusion as claimed in claim 7, wherein the calculation process of the stack flow control unit comprises: given an input vector x for a current time ttAnd memory cell c of last time period t-1t-1Memory cell c of the current time periodtCalculated from the following formula:
Figure FDA0003079760880000021
Figure FDA0003079760880000022
Figure FDA0003079760880000023
wherein, WiIs a trainable parameter of the neural network, tanh is a hyperbolic tangent function, which is a vector point product operation. f. oftThe output of the forgetting gate is used for forgetting a part of unimportant historical memory information in the SFGU unit, and the function of compressing information flow is achieved.
9. The method for deep logical inference financial text analysis based on multi-element fusion as claimed in claim 1, wherein in step S3, the slot memory module is used to encode external elements and then store them in the memory unit for later module; in step S4, when calculating the high-level vector representation, weights are reassigned to each vector according to the importance of each position.
10. A deep logical inference financial text analysis system based on multivariate factor fusion, comprising a processor and a memory, the memory having stored therein a computer program executable by the processor for carrying out the method according to any one of claims 1-9.
CN202110562999.9A 2021-05-24 2021-05-24 Deep logic reasoning financial text analysis method and system based on multi-element factor fusion Active CN113239703B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110562999.9A CN113239703B (en) 2021-05-24 2021-05-24 Deep logic reasoning financial text analysis method and system based on multi-element factor fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110562999.9A CN113239703B (en) 2021-05-24 2021-05-24 Deep logic reasoning financial text analysis method and system based on multi-element factor fusion

Publications (2)

Publication Number Publication Date
CN113239703A true CN113239703A (en) 2021-08-10
CN113239703B CN113239703B (en) 2023-05-02

Family

ID=77138520

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110562999.9A Active CN113239703B (en) 2021-05-24 2021-05-24 Deep logic reasoning financial text analysis method and system based on multi-element factor fusion

Country Status (1)

Country Link
CN (1) CN113239703B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108415977A (en) * 2018-02-09 2018-08-17 华南理工大学 One is read understanding method based on the production machine of deep neural network and intensified learning
CN108920587A (en) * 2018-06-26 2018-11-30 清华大学 Merge the open field vision answering method and device of external knowledge
CN109389091A (en) * 2018-10-22 2019-02-26 重庆邮电大学 The character identification system and method combined based on neural network and attention mechanism
CN110188167A (en) * 2019-05-17 2019-08-30 北京邮电大学 A kind of end-to-end session method and system incorporating external knowledge
CN111241807A (en) * 2019-12-31 2020-06-05 浙江大学 Machine reading understanding method based on knowledge-guided attention
CN111274820A (en) * 2020-02-20 2020-06-12 齐鲁工业大学 Intelligent medical named entity identification method and device based on neural network

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108415977A (en) * 2018-02-09 2018-08-17 华南理工大学 One is read understanding method based on the production machine of deep neural network and intensified learning
CN108920587A (en) * 2018-06-26 2018-11-30 清华大学 Merge the open field vision answering method and device of external knowledge
CN109389091A (en) * 2018-10-22 2019-02-26 重庆邮电大学 The character identification system and method combined based on neural network and attention mechanism
CN110188167A (en) * 2019-05-17 2019-08-30 北京邮电大学 A kind of end-to-end session method and system incorporating external knowledge
CN111241807A (en) * 2019-12-31 2020-06-05 浙江大学 Machine reading understanding method based on knowledge-guided attention
CN111274820A (en) * 2020-02-20 2020-06-12 齐鲁工业大学 Intelligent medical named entity identification method and device based on neural network

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
ARTHUR SZLAM等: "End-To-End Memory Networks" *
LI XIN等: "Exploiting BERT for End-to-End Aspect-based Sentiment Analysis" *
RUIDAN HE等: "An interactive multi-task learning network for end-to-end aspect-based sentiment analysis" *
张芮;杨煦晨;琚生根;刘宁宁;谢正文;王婧妍;: "基于多层次动态门控推理网络的文本蕴含识别" *

Also Published As

Publication number Publication date
CN113239703B (en) 2023-05-02

Similar Documents

Publication Publication Date Title
CN110188176B (en) Deep learning neural network, and training and predicting method, system, device and medium
Sukhbaatar et al. End-to-end memory networks
CN111914067B (en) Chinese text matching method and system
CN110110324B (en) Biomedical entity linking method based on knowledge representation
CN108628935B (en) Question-answering method based on end-to-end memory network
CN110210032A (en) Text handling method and device
CN114443827A (en) Local information perception dialogue method and system based on pre-training language model
CN112784532B (en) Multi-head attention memory system for short text sentiment classification
CN116204674B (en) Image description method based on visual concept word association structural modeling
CN112347756A (en) Reasoning reading understanding method and system based on serialized evidence extraction
CN111881292A (en) Text classification method and device
CN111145914B (en) Method and device for determining text entity of lung cancer clinical disease seed bank
Yang et al. Sequence-to-sequence prediction of personal computer software by recurrent neural network
Caglayan Multimodal machine translation
CN109948163B (en) Natural language semantic matching method for dynamic sequence reading
CN115186147A (en) Method and device for generating conversation content, storage medium and terminal
CN110633473A (en) Implicit discourse relation identification method and system based on conditional random field
Tekir et al. Deep learning: Exemplar studies in natural language processing and computer vision
CN116629361A (en) Knowledge reasoning method based on ontology learning and attention mechanism
CN114661874B (en) Visual question-answering method based on multi-angle semantic understanding and self-adaptive double channels
CN113010662B (en) Hierarchical conversational machine reading understanding system and method
CN113239703A (en) Deep logical reasoning financial text analysis method and system based on multivariate factor fusion
Kandi Language Modelling for Handling Out-of-Vocabulary Words in Natural Language Processing
Shumsky et al. ADAM: a prototype of hierarchical neuro-symbolic AGI
Kvochick Fundamental aspects of pattern recognition in architectural drawing

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant