CN113869349B - Schematic question-answering method based on hierarchical multi-task learning - Google Patents
Schematic question-answering method based on hierarchical multi-task learning Download PDFInfo
- Publication number
- CN113869349B CN113869349B CN202110892487.9A CN202110892487A CN113869349B CN 113869349 B CN113869349 B CN 113869349B CN 202110892487 A CN202110892487 A CN 202110892487A CN 113869349 B CN113869349 B CN 113869349B
- Authority
- CN
- China
- Prior art keywords
- question
- sequence
- answering
- module
- answer
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 29
- 238000010586 diagram Methods 0.000 claims abstract description 41
- 238000012549 training Methods 0.000 claims abstract description 41
- 238000004458 analytical method Methods 0.000 claims abstract description 34
- 230000000007 visual effect Effects 0.000 claims abstract description 15
- 238000000605 extraction Methods 0.000 claims abstract description 7
- 238000012545 processing Methods 0.000 claims abstract description 4
- 238000012360 testing method Methods 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 6
- 239000013598 vector Substances 0.000 claims description 5
- 241001061076 Melanonus zugmayeri Species 0.000 claims description 3
- 238000001514 detection method Methods 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000012916 structural analysis Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/284—Lexical analysis, e.g. tokenisation or collocates
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
-
- 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
- G06N3/045—Combinations of 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
- G06N3/047—Probabilistic or stochastic 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention relates to the technical field of image question answering, in particular to a schematic question answering method based on hierarchical multi-task learning, which comprises the following steps of: s1, detecting image components of a training image and a pre-training target detector, and performing position coding and visual feature extraction; s2, taking the training set images and the region characteristic sequence as the input of a diagram analysis module, and predicting the relationship between components; s3, forming a statement sentence by the training set questions and the alternative answers, performing word instalation processing, and extracting language features to obtain a language sequence formed by the alternative answers and the questions; s4, for the training set, splicing the output of the graph analysis module and the language sequence and inputting the spliced output into a question-answering module, and training network parameters; and S5, coding the graph, the question and the alternative answer to the test set image to form a region feature sequence and a language sequence. The invention combines the graph analysis module and the question and answer module together, utilizes multi-level tasks to train, and realizes a multi-task learning framework based on two levels of analysis and question and answer.
Description
Technical Field
The invention relates to the technical field of image question answering, in particular to a schematic diagram question answering method based on hierarchical multi-task learning.
Background
In recent years, with the continuous development of the fields of deep learning and computer vision, the comprehension ability of a computer to natural images has greatly progressed, for example, in the field of visual question answering, the current method can reach higher precision. However, the computer has difficulty in achieving a good semantic understanding of the schematic diagram, such as in the task of question answering the schematic diagram. The schematic question-answering task requires the computation mechanism to solve semantic knowledge in the schematic and deduce a correct answer from a plurality of candidate answers according to the knowledge. Schematic question-answering can be considered as an evaluation task for intelligent reasoning on diagram semantics, and is very challenging. First, the schematic diagram is an important media form for explanation, which is widely used in textbooks, slides, documents, and the like. While natural images hardly play the same role. Unlike natural images, schematics have highly structured semantic information, such as arrows, that represent a connection. Second, visually similar structures may have very different semantics in different schematics. These cause that the question-answering method applied to the natural image is difficult to apply to the schematic.
In order to answer questions related to the schematic diagrams, the existing method generally divides the question answering into two independent stages, and the analysis diagram module firstly identifies components and then generates a structure diagram by pairwise matching and classifying the components. The question-answering module generates facts, questions and alternative options according to the structure diagram and selects the option with the highest possibility. The method has no feedback between the generation of the diagram and the question reasoning, the accuracy of the analyzed structure diagram only depends on the image information of the training set, and the question-answering module selects the correct answer according to the generated structure diagram. During training, errors of structure analysis can cause errors of the question answering module to be very large, but the errors are only used for optimizing the question answering module and are not propagated to the analysis module in a reverse direction. Such two-stage modules cannot feed back each other, and naturally, it cannot be guaranteed that a globally optimal solution of the problem is found. The schematic diagrams often appear in materials and documents, contain rich knowledge and structural information, and are more complex than natural images, so that the common visual question answering method is difficult to apply to schematic diagram question answering. Also, similar schematics may contain disparate semantic information, which also makes schematic question-answering very challenging.
Disclosure of Invention
The invention aims to provide a schematic question-answering method based on hierarchical multi-task learning aiming at the problems in the background technology, which comprises the following steps:
s1, detecting image components of a training image and a pre-training target detector, performing position coding and visual feature extraction, and then coding to form a region feature sequence;
s2, taking the training set images and the region characteristic sequences as the input of a graph analysis module, predicting the relationship between components and training network parameters;
s3, forming a statement sentence by the training set questions and the alternative answers, performing word instalation processing, and extracting language features to obtain a language sequence formed by the alternative answers and the questions;
s4, for the training set, splicing the output of the graph analysis module and the language sequence and inputting the output and the language sequence into a question-answering module, predicting correct options of the question and training network parameters;
and S5, coding the graph, the question and the alternative answer into a regional characteristic sequence and a language sequence for the test set image, inputting the regional characteristic sequence and the language sequence into a depth network, and predicting the correct options of the question.
Preferably, S2 and S4 use a multitask learning framework to solve the question-answering task by illustrating two learning tasks, namely the analysis task and the question-answering task.
Preferably, the training graph analysis module and the question and answer module are combined in the step S4, so that the training loss of the question and answer can be correctly fed back to the graph analysis module and the question and answer module.
Preferably, the graph analysis task is firstly carried out, then the question and answer task is carried out by utilizing the output in the graph analysis task, and a hierarchical multi-task framework is constructed.
Preferably, in S1, the pre-trained target detector is a YOLO v3 target detector pre-trained based on the COCO dataset and the schematic question-and-answer image dataset; the detected schematic diagram components comprise four categories of characters, object areas, arrow heads and arrow tails; the region signature sequence coding can be divided into the following two sub-steps:
s1.1, for scheme I, the detected composition is O = { O = 1 ,o 2 ,...,o m Obtaining a visual feature sequence [ z ] with the dimension of 2048 through a deep network feature extraction module ResNet101 1 ,z 2 ,...,z m ]Then the whole image is also passed through the z obtained by the same feature extractor 0 Put in the first bit of the sequence as global information, constituting [ z ] 0 ,...,z m ]Upper left coordinate (x) of each detected component min ,y min ) And lower right coordinate (x) max ,y max ) According to (x) min ,y min ,x max ,y max ) Width, height, area constitute a 7-dimensional sequence of positional features [ q ] 0 ,q 1 ,...,q m ];
S1.2, visual characteristic sequence [ z 0 ,...,z m ]And obtaining 1024-dimensional region feature sequence through a visual feature and position feature fusion module according to the position features
Preferably, in S2, the module consists of multiple layers of transformers and one layer of GRU, and the region characteristic sequence H is formed by in Inputting into a Transfomer encoder to obtain 1024-dimensional outputForming a pair of regions pairwise, predicting whether a relationship exists between the two regions, and determining a relationship candidate pair<o i ,o j >Is characterized byByConcatenating to obtain, where i, j =1, 2., m and i ≠ j, the featuresThe composed sequences are input into GRU, and each pair is predicted<o i ,o j >If a relationship exists, then the loss is calculated according to:
where N is the number of pairs of relationship candidates, y n Is the true value of the nth pair of relationships,is the predicted value of the model to the nth pair relation.
Preferably, in S3, the tokenization and coding process uses RoBERTA, and for a question q of the image I, the question contains K candidate answers { a } k |k=1,..K, questions q and a k Statement sentence s composed of space connections k Inputting Roberta to carry out word segmentation and coding to obtain s k Characteristic of a languageWhereinn is the maximum number of words in the sentence.
Preferably, in S4, the question answering module is composed of a multi-layer Transformer module TB dqa And a full connection layer, a region characteristic sequence splicing language sequence,by TB dqa To obtainMultiplying the 0 th and the m +1 th vectors by corresponding elements, and inputting the multiplied vectors into the full-link layer and the softmax layer to obtain a k Score of candidate itemThen, the score with the highest score is selected as a predicted value, and the question-answer loss is calculated according to the following formula:
wherein, t correct A label of correct answer indicating the current question, in indicates if k and t correct If the two are the same, the value is 1, otherwise, the value is 0;
global loss of L = al sp +βL dqa And the alpha and the beta are adjustable hyper-parameters which are used for learning the balanced question-answering module and the balanced question-answering module, and global parameter adjustment is carried out through a back propagation algorithm so as to optimize the network parameters by taking the minimized global loss function as a target until the function value is not reduced any more.
Preferably, in S5, the trained network parameters are fixed to reason about the schematic diagram and the problem, and the method specifically includes the following substeps:
s5.1, carrying out component detection and feature coding according to the schematic diagram S1, and then passing through TB according to the schematic diagram S2 sp Obtain a characteristic output H out ;
S5.2, preprocessing the question and the candidate answer according to S3 to obtain the question and the candidate answer a k Characteristic of a language
S5.3、H out And characteristics of languageSplicing the characteristic sequences that make up the unionThen press through TB in S4 dqa And softmax layer get a k Score of candidate itemAnd selecting the option with the highest score as an output answer.
Compared with the prior art, the invention has the following beneficial technical effects:
the invention combines the graph analysis module and the question-answering module together, utilizes multi-level tasks for training, realizes a multi-task learning framework based on two levels of analysis and question-answering, and has better semantic understanding and reasoning capability compared with the prior method.
Drawings
FIG. 1 is a schematic diagram of a training flow diagram of a question-answering method based on hierarchical multi-task learning according to the present invention;
fig. 2 is a diagram illustrating a model structure of a question-and-answer.
Detailed Description
The images of the data set involved in the method proposed by the present invention are schematic diagrams, each schematic diagram comprises a plurality of questions related to the contents of the schematic diagram, and each question has four candidate answers.
In order to solve the multi-stage problem of the previous work, the invention provides a schematic question-answering method based on hierarchical multi-task learning.
As shown in fig. 1-2, the schematic question-answering method based on hierarchical multitask learning provided by the present invention includes the following steps:
s1, detecting image components of a training image and a pre-training target detector, performing position coding and visual feature extraction, and then coding to form a region feature sequence; the pre-training target detector is a YOLO v3 target detector pre-trained on the basis of a COCO data set and a schematic question and answer image data set; the detected schematic diagram components comprise four categories of characters, object areas, arrow heads and arrow tails; the region signature sequence coding can be divided into the following two sub-steps: s1.1, for scheme I, the detected composition is O = { O = 1 ,o 2 ,...,o m Obtaining a visual feature sequence [ z ] with dimension 2048 through a depth network feature extraction module ResNet101 1 ,z 2 ,...,z m ]Then the whole image is also passed through the z obtained by the same feature extractor 0 Put in the first bit of the sequence as global information, constituting [ z ] 0 ,...,z m ]Upper left coordinate (x) of each detected component min ,y min ) And lower right coordinate (x) max ,y max ) According to (x) min ,y min ,x max ,y max ) Width, height, area make up a 7-dimensional sequence of positional features [ q [ ] 0 ,q 1 ,...,q m ](ii) a S1.2, visual characteristic sequence [ z 0 ,...,z m ]And obtaining a 1024-dimensional region feature sequence through a visual feature and position feature fusion module according to the position features
S2, inputting the training set images and the region characteristic sequences as an image analysis module, predicting the relationship between components and training network parameters; the module consists of a plurality of layers of transformers and a layer of GRU is composed of a region signature sequence H in Inputting into a Transfomer encoder to obtain 1024-dimensional outputForming a pair of regions pairwise, predicting whether a relationship exists between the two regions, and determining a relationship candidate pair<o i ,o j >Is characterized byByConcatenating to obtain, where i, j =1, 2., m and i ≠ j, the featuresThe composed sequences are input into GRU, and each pair is predicted<o i ,o j >If a relationship exists, then the loss is calculated according to:
where N is the number of pairs of relationship candidates, y n Is the true value of the nth pair of relationships,is the predicted value of the model to the nth pair relation.
S3, forming a statement sentence by the training set questions and the alternative answers, then performing word instantiation processing and extracting language features to obtain a language sequence formed by the alternative answers and the questions; the tokenization and encoding process utilizes RoBERTA, for a question q of an image I, the question containing K candidate answers { a } k I K =1, say, K }, will question q and a k Statement sentence s composed of space connections k Inputting Roberta to carry out word segmentation and coding to obtain s k Characteristic of a languageWhereinn is the maximum number of words in the sentence.
S4, for the training set, splicing the output of the graph analysis module and the language sequence and inputting the spliced output into a question-answering module, predicting correct options of the question and training network parameters; s2 and S4 use a multi-task learning framework to solve the question and answer task through two learning tasks, namely a diagram analysis task and a question and answer task. And S4, the training graph analysis module and the question and answer module are combined, so that the training loss of the question and answer can be correctly fed back to the graph analysis module and the question and answer module. The question-answering module is composed of a multi-layer Transformer module TB dqa And a full connection layer, a region characteristic sequence splicing language sequence,by TB dqa To obtainMultiplying the 0 th and the (m + 1) th vectors by corresponding elements, and inputting the multiplied vectors into a full connection layer and a softmax layer to obtain a k Score of candidate itemThen, the score with the highest score is selected as a predicted value, and the question-answering loss is calculated according to the following formula:
wherein, t correct A correct answer label indicating the current question, in indicates if k and t correct If the two are the same, the value is 1, otherwise, the value is 0; global loss L = al sp +βL dqa And the alpha and the beta are adjustable hyper-parameters which are used for learning the balanced question-answering module and the balanced question-answering module, and global parameter adjustment is carried out through a back propagation algorithm so as to optimize the network parameters by taking the minimized global loss function as a target until the function value is not reduced any more.
And S5, coding the graph, the question and the alternative answer into a regional characteristic sequence and a language sequence for the test set image, inputting the regional characteristic sequence and the language sequence into a depth network, and predicting the correct options of the question. The method is characterized in that the diagram and the problem are reasoned by fixing trained network parameters, and the method specifically comprises the following substeps:
s5.1, carrying out component detection and feature coding according to the schematic diagram S1, and then passing through TB according to the schematic diagram S2 sp Obtain a characteristic output H out ;
S5.2, preprocessing the question and the candidate answer according to S3 to obtain the question and the candidate answer a k Characteristic of a language
S5.3、H out And characteristics of languageSplicing the characteristic sequences that make up the unionThen according to the passage TB in S4 dqa And softmax layer to obtain a k Score of candidate itemAnd selecting the option with the highest score as an output answer.
In the embodiment, the graph analysis task is performed first, and then the question and answer task is performed by using the output in the graph analysis task, so that a hierarchical multi-task framework is constructed.
In the invention, the model framework comprises a diagram analysis module and a question-answering module and is based on hierarchical multi-task learning training. In our multi-task learning paradigm, the two tasks of graph analysis and question answering are at different levels in the training and use different neural network modules to form a hierarchy. First, a pre-trained target detector is used to detect meaningful regions in the schematic and extract features such as text, arrows, objects. Then, the image feature sequence and the word sequence of the question are spliced into an input sequence. After the sequence is input into the model, the model is simultaneously trained by a structure analysis task and a question-answering task of the graph, the graph analysis module encodes components in the graph and relation information thereof, and the question-answering module deduces correct answers by combining structure signals with answers of questions. The schematic diagram question-answering method provided by the invention can be used for simultaneously carrying out structural analysis and deducing question answers on the diagram, realizes a multi-task learning framework based on two layers of analysis and question-answering, and has better semantic understanding and reasoning capabilities compared with the existing method.
The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.
Claims (5)
1. The schematic question answering method based on hierarchical multitask learning is characterized by comprising the following steps of:
s1, detecting image components of a training image and a pre-training target detector, performing position coding and visual feature extraction, and then coding to form a region feature sequence;
s2, taking the training set images and the region characteristic sequences as the input of a graph analysis module, predicting the relationship between components and training network parameters;
s3, forming a statement sentence by the training set questions and the alternative answers, performing word instalation processing, and extracting language features to obtain a language sequence formed by the alternative answers and the questions;
s4, for the training set, splicing the output of the graph analysis module and the language sequence and inputting the spliced output into a question-answering module, predicting correct options of the question and training network parameters;
s5, coding the graph, the question and the alternative answer into a regional characteristic sequence and a language sequence for the test set image, inputting a depth network, and predicting correct options of the question;
in S1, a pre-training target detector is a YOLO v3 target detector pre-trained on the basis of a COCO data set and a schematic question and answer image data set; the detected schematic diagram components comprise four categories of characters, object areas, arrow heads and arrow tails; the region signature sequence coding can be divided into the following two sub-steps:
s1.1 for scheme I, the detected composition is O = { O 1 ,o 2 ,...,o m Obtaining a visual feature sequence [ z ] with the dimension of 2048 through a deep network feature extraction module ResNet101 1 ,z 2 ,...,z m ]Then the whole image is also passed through the z obtained by the same feature extractor 0 Put in the first bit of the sequence as global information, constituting [ z ] 0 ,...,z m ]Upper left coordinate (x) of each detected component min ,y min ) And lower right coordinate (x) max ,y max ) According to (x) min ,y min ,x max ,y max ) Width, height, area constitute a 7-dimensional sequence of positional features [ q ] 0 ,q 1 ,...,q m ];
S1.2, visual characteristic sequence [ z 0 ,...,z m ]And obtaining a 1024-dimensional region feature sequence through a visual feature and position feature fusion module according to the position features
In S2, the module consists of a plurality of layers of transformers and a layer of GRU, and the region characteristic sequence H is divided into in Inputting into a Transfomer encoder to obtain 1024-dimensional outputForming a pair of regions pairwise, predicting whether a relation exists between the two regions, and enabling the relation candidate pair to be less than o i ,o j Characteristic ofByq i ,qj are spliced, where i, j =1, 2.. M and i ≠ j, the featuresThe composed sequences are input into GRU, and each pair is predicted to be < o i ,o j If a relationship exists, then the loss is calculated according to:
where N is the number of pairs of relationship candidates, y n Is the true value of the nth pair of relationships,predicting the nth pair relation for the model;
in S3, the tokenization and coding process utilizes RoBERTA, and for a question q of an image I, the question contains K candidate answers { a } k I K =1,.., K }, questions q and a are solved k Statement sentence s composed of space connections k Inputting Roberta to carry out word segmentation and coding to obtain s k Characteristic of a languageWhereinn is the maximum number of words in the sentence;
in S4, the question answering module is composed of a multi-layer Transformer module TB dqa And a full connection layer, a region characteristic sequence splicing language sequence,by TB dqa To obtainMultiplying the 0 th and the m +1 th vectors by corresponding elements and outputtingInto the fully-connected layer and the softmax layer to obtain a k Score of candidate itemThen, the score with the highest score is selected as a predicted value, and the question-answer loss is calculated according to the following formula:
wherein, t correct A label of correct answer indicating the current question, in indicates if k and t correct If the two are the same, the value is 1, otherwise, the value is 0;
global loss of L = al sp +βL dqa And the alpha and the beta are adjustable hyper-parameters which are used for learning the balanced question-answering module and the balanced question-answering module, and global parameter adjustment is carried out through a back propagation algorithm so as to optimize the network parameters by taking the minimized global loss function as a target until the function value is not reduced any more.
2. The method for learning question and answer based on hierarchical multitask according to claim 1, characterized by that S2 and S4 use multitask learning frame to solve the question and answer task by means of two learning tasks of diagram analysis task and question and answer task.
3. The schematic diagram question-answering method based on hierarchical multitask learning according to claim 1, characterized in that in S4, a training diagram analysis module and a question-answering module are combined, so that the training loss of the question-answering can be correctly fed back to the diagram analysis module and the question-answering module.
4. The method for learning and answering questions and answers based on hierarchical multitask as claimed in claim 1, wherein a graph analysis task is performed first, and then the question-answering task is performed by using output in the graph analysis task to construct a hierarchical multitask framework.
5. The schematic diagram question-answering method based on hierarchical multitask learning according to claim 1, characterized in that in S5, the trained network parameters are fixed to reason about the schematic diagram and the question, and the method specifically comprises the following substeps:
s5.1, carrying out component detection and feature coding according to S1 in the schematic diagram, and then passing through TB according to S2 sp Obtain a characteristic output H out ;
S5.2, preprocessing the question and the candidate answer according to S3 to obtain the question and the candidate answer a k Characteristic of a language
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110892487.9A CN113869349B (en) | 2021-08-04 | 2021-08-04 | Schematic question-answering method based on hierarchical multi-task learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110892487.9A CN113869349B (en) | 2021-08-04 | 2021-08-04 | Schematic question-answering method based on hierarchical multi-task learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113869349A CN113869349A (en) | 2021-12-31 |
CN113869349B true CN113869349B (en) | 2022-10-14 |
Family
ID=78990250
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110892487.9A Expired - Fee Related CN113869349B (en) | 2021-08-04 | 2021-08-04 | Schematic question-answering method based on hierarchical multi-task learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113869349B (en) |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10909329B2 (en) * | 2015-05-21 | 2021-02-02 | Baidu Usa Llc | Multilingual image question answering |
CN111008293A (en) * | 2018-10-06 | 2020-04-14 | 上海交通大学 | Visual question-answering method based on structured semantic representation |
CN110348535B (en) * | 2019-07-17 | 2022-05-31 | 北京金山数字娱乐科技有限公司 | Visual question-answering model training method and device |
CN111782839B (en) * | 2020-06-30 | 2023-08-22 | 北京百度网讯科技有限公司 | Image question-answering method, device, computer equipment and medium |
-
2021
- 2021-08-04 CN CN202110892487.9A patent/CN113869349B/en not_active Expired - Fee Related
Also Published As
Publication number | Publication date |
---|---|
CN113869349A (en) | 2021-12-31 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wu et al. | Image-to-markup generation via paired adversarial learning | |
Atienza | Advanced Deep Learning with Keras: Apply deep learning techniques, autoencoders, GANs, variational autoencoders, deep reinforcement learning, policy gradients, and more | |
Davis et al. | End-to-end document recognition and understanding with dessurt | |
Li et al. | Improving attention-based handwritten mathematical expression recognition with scale augmentation and drop attention | |
CN110866542B (en) | Depth representation learning method based on feature controllable fusion | |
CN111160037A (en) | Fine-grained emotion analysis method supporting cross-language migration | |
CN112115238A (en) | Question-answering method and system based on BERT and knowledge base | |
Sharma et al. | A survey of methods, datasets and evaluation metrics for visual question answering | |
CN111831789A (en) | Question-answer text matching method based on multilayer semantic feature extraction structure | |
CN111680484B (en) | Answer model generation method and system for visual general knowledge reasoning question and answer | |
CN110781672A (en) | Question bank production method and system based on machine intelligence | |
Davis et al. | Visual FUDGE: form understanding via dynamic graph editing | |
CN115953569A (en) | One-stage visual positioning model construction method based on multi-step reasoning | |
Kovvuri et al. | Pirc net: Using proposal indexing, relationships and context for phrase grounding | |
Palash et al. | Bangla image caption generation through cnn-transformer based encoder-decoder network | |
Deng et al. | A position-aware transformer for image captioning | |
CN117609536A (en) | Language-guided reference expression understanding reasoning network system and reasoning method | |
CN113869349B (en) | Schematic question-answering method based on hierarchical multi-task learning | |
Sharma et al. | Visual question answering model based on the fusion of multimodal features by a two-way co-attention mechanism | |
CN115409028A (en) | Knowledge and data driven multi-granularity Chinese text sentiment analysis method | |
CN114691848A (en) | Relational triple combined extraction method and automatic question-answering system construction method | |
CN112364654A (en) | Education-field-oriented entity and relation combined extraction method | |
Pei et al. | Visual relational reasoning for image caption | |
Xie et al. | Enhancing multimodal deep representation learning by fixed model reuse | |
Cai et al. | Multimodal Visual Question Answering Model Enhanced with Image Emotional Information |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20221014 |