CN112988959A - False news interpretability detection system and method based on evidence inference network - Google Patents

False news interpretability detection system and method based on evidence inference network Download PDF

Info

Publication number
CN112988959A
CN112988959A CN202110045012.6A CN202110045012A CN112988959A CN 112988959 A CN112988959 A CN 112988959A CN 202110045012 A CN202110045012 A CN 202110045012A CN 112988959 A CN112988959 A CN 112988959A
Authority
CN
China
Prior art keywords
sequence
news
false news
false
sequences
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
CN202110045012.6A
Other languages
Chinese (zh)
Other versions
CN112988959B (en
Inventor
饶元
吴连伟
兰玉乾
孙菱
郑鹏怡
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.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong 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 Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN202110045012.6A priority Critical patent/CN112988959B/en
Publication of CN112988959A publication Critical patent/CN112988959A/en
Application granted granted Critical
Publication of CN112988959B publication Critical patent/CN112988959B/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
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3344—Query execution using natural language analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/205—Parsing
    • G06F40/211—Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
    • 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
    • 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
    • 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

Landscapes

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

Abstract

The invention discloses a false news interpretability detection system and method based on evidence inference network, comprising the following steps: inputting a false news content sequence, a splicing sequence of all related articles and a plurality of different single related article sequences; enabling the false news content sequence to interact with all relevant article characteristics, and capturing two types of sequence characteristics; interacting the whole viewpoint sequence with the sequence of each related article, and exploring fine-grained potential sequence conflicts in each related article; and carrying out consistency modeling on the core sequence fragments and potential sequence conflicts, and detecting the interpretable false news by surrounding the conflicts of the core sequence fragments of the false news by the spleen and the stomach. The invention focuses on the false core semantic segment of the news to be detected in a fine-grained manner, and explores interpretable evidence from related articles, thereby improving the interpretability of false news detection. The method not only improves the false news detection performance, but also provides effective evidence to realize the interpretability of the detection result.

Description

False news interpretability detection system and method based on evidence inference network
Technical Field
The invention relates to a false news interpretability detection system and method based on evidence inference network.
Background
Studies have shown that while the percentage of false news is only 1% of the total consumption of all media news, the ratio of false news to all tweet news in social media is as high as 6%. Therefore, accurately and efficiently detecting false news on social media, preventing their spread and dissemination, and finding relevant evidence to rumor them is a very critical task in the current field of social media analysis.
False news detection is a difficult and challenging task facing today in the industry and academia. Previous research has focused on extracting various rich linguistic semantic features and manually extracted features around the textual content of the news to be detected and its associated metadata features, which have achieved superior performance. However, these methods have the general drawback that it is difficult to provide effective interpretation of the detection result of false news, colloquially, why is a news detected as false, where did it go wrong? For this problem, current research is beginning to focus on interpretable false news detection by constructing interactions between false news and its related articles and seeking conflicting or questioning semantics between them as evidence and interpreting the corresponding detection results. Currently, it is a reasonable idea to explore and construct an interaction model to solve the interpretable false news detection task, mainly because: different related articles, similar to "crowdsourcing" bartering, as opinions of different users about specific news, often challenge the wrong part of the false news, even though sometimes it cannot be specifically revealed why this news is false. However, these methods usually ignore the fact that captured conflicting semantics are not all spread around the spurious part of the spurious news, but also contain questioning semantics for non-spurious parts of some news, which makes it less fine-grained to interpret the detection of spurious news by many conflicting or questioning semantics as evidence. Therefore, how to capture the conflict semantics between the false news and the related articles so as to focus on the fine-grained core false part in the false news is also the core problem of the interpretability research of the false news.
Disclosure of Invention
The invention aims to solve the problems in the prior art and provides a false news interpretability detection system and method based on evidence inference network,
in order to achieve the purpose, the invention adopts the following technical scheme to realize the purpose:
a false news interpretability detection method for inferring a network based on evidence comprises the following steps:
step 1, inputting a false news content sequence, a splicing sequence of all related articles and a plurality of different single related article sequences, and respectively capturing the internal context hidden layer sequences of the false news content sequence, the splicing sequence of all related articles and the different single related article sequences through BilSTM;
step 2, enabling the false news content sequence to interact with all the related article characteristics, capturing two types of sequence characteristics, focusing core sequence segments concerned by users in the false news content, and learning the overall viewpoint sequences of all the users in the related articles;
step 3, the overall viewpoint sequence and the sequence of each related article are interacted with each other, and potential sequence conflicts with fine granularity in each related article are explored;
and 4, carrying out consistency modeling on the core sequence fragments and potential sequence conflicts, and detecting the interpretable false news by surrounding the conflicts of the core sequence fragments of the false news by the spleen and the stomach.
The invention further improves the following steps:
the specific method of the step 1 is as follows:
the BilSTM model encodes an input sequence of false news content, a mosaic sequence of all related articles, and several different sequences of individual related articles, and employs a hidden layer vector e of the last stepiAs a contextual representation of each sequence;
a sequence of false news content, a stitched sequence of all related articles, and several different individual sequences of related articles, all of which can be represented as X ═ X for sequences each containing k words1,x2,…,xk}, embedded representation of each word
Figure BDA0002896892890000031
Is a d-dimensional vector which can be initialized by embedding a pre-training word into the vector;
the coding characteristics of the false news, all related articles and each related article are respectively denoted as ec、
Figure BDA0002896892890000032
And
Figure BDA0002896892890000033
wherein j is more than or equal to 1 and less than or equal to R.
In the step 2, each false news content and all related articles are mutually interacted through the cross attention unit and the two gated affine absorption units, and the valuable characteristics which are suitable for the false news content and the related articles are screened out.
In the cross attention unit, a self-attention network is used as a cross attention ternary system to explicitly capture the dependency characteristics between words in the sequences and learn the internal structure information, so as to ensure the deep interaction between the two sequences, and the specific method is as follows:
Figure BDA0002896892890000034
wherein H is the result from the attention network, Q, K, V are query, key and value matrix, respectively;
Figure BDA0002896892890000035
and K ═ V ═ ecT is a transpose operation, dkThe hidden layer size of the BilSTM in the stage 1 is 2 h;
the self-attention network is a multi-head attention network and is used for performing linear mapping on Q, K and V for m times and then performing point-by-point attention in parallel; multi-head attention network headi：
Figure BDA0002896892890000041
Hs＝MultiHead(Q,K,V)＝Concat(head1,head2,…,headm)Wo (3)
Wherein the content of the first and second substances,
Figure BDA0002896892890000042
Figure BDA0002896892890000043
Figure BDA0002896892890000044
and
Figure BDA0002896892890000045
are all parameters that can be trained, d1Has a dimension of 2H/m and HsIs of a size of
Figure BDA0002896892890000046
HsIs an interactive feature of false news and all related articles.
The gated affine absorption unit is used for capturing characteristics beneficial to respective sequences so as to adaptively focus on remarkable sequence segment characteristics in the false news sequences and focus on user overall viewpoint sequences in the related article sequences; the structure of the gated affine absorption unit is as follows:
tc＝tanh(WcHc+bc) (4)
ts＝tanh(WsHs+bs) (5)
α(ec)＝Wαtc+bα (6)
β(Hs)＝Wβts+bβ (7)
γ(Hs)＝Wγts+bγ (8)
Figure BDA0002896892890000047
wherein, tcAs a result of the transformation based on the news sequence, c is the news sequence, tanh is the activation function, tsFor the transformation results based on the related article sequences, s is the concatenation sequence of all related articles, HcFor hidden layer representation based on news sequences, HsFor hidden layer representation based on related article sequences, α () is an affine result based on news sequences, β () is a first affine result based on related article sequences, γ () is a second affine result based on related article sequences, all W and b are learnable parameters, which indicate a dot product operation from element to element; the structure of gated G2 for all related articles is the same as gated G1 for spurious news, and
Figure BDA0002896892890000048
and
Figure BDA0002896892890000049
the output features corresponding to gates G1 and G2, respectively, are the core sequence segments in the bogus news and the overall opinion sequence in all related articles.
In said step 3, the cross attention unit is used to sequence the overall views
Figure BDA0002896892890000051
Semantics of local single related article
Figure BDA0002896892890000052
Full interaction and fusion; the cross attention network is:
Figure BDA0002896892890000053
wherein the content of the first and second substances,
Figure BDA0002896892890000054
representing the conflict semantics captured for the ith related article.
The specific method of the step 4 is as follows:
using two BilSTM modules to separately encode core sequence segments of false news
Figure BDA0002896892890000055
And conflicting semantics for each related article
Figure BDA0002896892890000056
And as its contextual representation by means of the hidden layer output of BilSTM, respectively
Figure BDA0002896892890000057
And
Figure BDA0002896892890000058
matching the false news and the salient features of each related article by means of an attention mechanism, wherein for the jth word in the ith related article, the attention mechanism is as follows:
Figure BDA0002896892890000059
Figure BDA00028968928900000510
wherein the content of the first and second substances,
Figure BDA00028968928900000511
is a conflict of the ith article
Figure BDA00028968928900000512
The number j of the word (a),
Figure BDA00028968928900000513
and
Figure BDA00028968928900000514
respectively measuring the original normalized relevance of the conflicted jth word to the whole false news sequence;
by interacting the entire sequence of false news with the conflicting sequence of related articles, a false news-directed conflict representation can be obtained
Figure BDA00028968928900000515
Figure BDA00028968928900000516
The summation calculation between the construction elements integrates the overall sequence of false news-guided conflicts and coding conflicts:
Figure BDA00028968928900000517
wherein the content of the first and second substances,
Figure BDA0002896892890000061
is a conflicting composite representation of the ith related article,
Figure BDA0002896892890000062
is a summation calculation operation between elements;
will be provided with
Figure BDA0002896892890000063
And
Figure BDA0002896892890000064
splicing is carried out, and then the low-dimensional prediction vector is obtained by inputting the low-dimensional prediction vector into only one full-connection layer and is used as a consistency representation between the core segment of the false news and the conflict of the related articles
Figure BDA0002896892890000065
Figure BDA0002896892890000066
For different individual related articles, respective consistent prediction vectors are obtained in a similar operation, i.e.
Figure BDA0002896892890000067
Integrating consistency prediction vectors of related articles through splicing operation, and predicting probability distribution p of the articles through the following formula:
Figure BDA0002896892890000068
wherein, WpAnd bpAre trainable parameters.
The model is trained by minimization of the cross entropy error:
loss＝-∑ylogp (17)
where y is based on the truth label.
A false news interpretability detection system for inferring networks based on evidence, comprising:
the input coding module is used for inputting a false news content sequence, a splicing sequence of all related articles and a plurality of different single related article sequences, and respectively capturing the internal context hidden layer sequence representations of the false news content sequence, the splicing sequence of all related articles and the different single related article sequences through the BilSTM;
the mutual interaction sharing module is used for enabling the false news content sequence to interact with all the related article characteristics, capturing two types of sequence characteristics, focusing core sequence segments concerned by users in the false news content, and learning the overall viewpoint sequences of all the users in the related articles;
the fine-grained conflict mining module is used for mutually interacting the whole viewpoint sequence and the sequence of each related article and exploring the potential fine-grained sequence conflict in each related article;
an evidence-based consistency module for consistency modeling of the core sequence segments against potential sequence conflicts surrounding the core sequence segments of the fake news by the spleen and stomach for detection of the interpretable fake news.
Compared with the prior art, the invention has the following beneficial effects:
the invention focuses on the false core semantic segment of the news to be detected in a fine-grained manner, and explores interpretable evidence from related articles, thereby improving the interpretability of false news detection. The method not only improves the false news detection performance, but also provides effective evidence to realize the interpretability of the detection result. The invention is used as an innovative fine-grained interpretable false news detection framework, and can capture the conflict of core contents which question false news as evidence, thereby verifying the result of false news detection. The mutual interaction sharing module provided by the invention can respectively and adaptively focus core semantic fragments of false news and overall viewpoint semantics of related articles by means of the gated imitation absorption module. The evidence-based consistency module provided by the invention can effectively match the conflict characteristics of suspected false news core contents by means of consistency modeling. The invention verifies its validity and interpretability on two popular false news detection data sets.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are required to be used in the embodiments will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present invention and therefore should not be considered as limiting the scope, and for those skilled in the art, other related drawings may be obtained according to these drawings without inventive effort.
FIG. 1 is an architectural diagram of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some embodiments of the present invention, but not all embodiments. The components of embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.
Thus, the following detailed description of the embodiments of the present invention, presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures.
In the description of the embodiments of the present invention, it should be noted that if the terms "upper", "lower", "horizontal", "inner", etc. are used for indicating the orientation or positional relationship based on the orientation or positional relationship shown in the drawings or the orientation or positional relationship which is usually arranged when the product of the present invention is used, the description is merely for convenience and simplicity, and the indication or suggestion that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and thus, cannot be construed as limiting the present invention. Furthermore, the terms "first," "second," and the like are used merely to distinguish one description from another, and are not to be construed as indicating or implying relative importance.
Furthermore, the term "horizontal", if present, does not mean that the component is required to be absolutely horizontal, but may be slightly inclined. For example, "horizontal" merely means that the direction is more horizontal than "vertical" and does not mean that the structure must be perfectly horizontal, but may be slightly inclined.
In the description of the embodiments of the present invention, it should be further noted that unless otherwise explicitly stated or limited, the terms "disposed," "mounted," "connected," and "connected" should be interpreted broadly, and may be, for example, fixedly connected, detachably connected, or integrally connected; can be mechanically or electrically connected; they may be connected directly or indirectly through intervening media, or they may be connected internally or indirectly. The specific meaning of the above terms in the present invention can be understood by those of ordinary skill in the art according to specific situations.
The invention is described in further detail below with reference to the accompanying drawings:
referring to fig. 1, an embodiment of the present invention discloses a false news interpretability detection system for inferring a network based on evidence, including:
module 1. input coding module: the model captures the semantic representation of the internal context hidden layer through BilSTM by means of three types of input, namely false news content, the overall semantics of all related articles, different single related articles and the like.
Module 2. mutual interaction share module: the module enables related news content semantics to interact with all related article features, so that two types of semantic features are captured in a self-adaptive manner, on one hand, core semantic fragments which are widely concerned by users in false news content are focused, and on the other hand, the overall viewpoint semantics of all users in related articles are learned.
Module 3, fine-grained conflict mining module: the overall viewpoint semantics obtained by the module 2 and the semantics of each related article interact with each other, so that the potential semantic conflict of fine granularity in each article is explored.
Module 4. evidence-based consistency module: in order to screen module 3 for conflicts that can become true evidence, the module constructs core semantic fragments of the false news (obtained by module 2) to model consistency with potential semantic conflicts (obtained by module 3) to match the conflicts around the core semantics of the false news, thereby performing interpretable false news detection by means of the conflicts.
The embodiment of the invention also discloses a false news interpretability detection method based on evidence inference network, which comprises the following steps:
step 0: given a number N of datasets
Figure BDA0002896892890000091
Wherein xiIndicates a news, R, to be tested for confidenceiFor a collection R containing m related articlesi＝{R1,R3,…,Rm}，yiRepresenting true and false binary labels;
step 1: the model contains three types of input sequences, false news content sequence xiA stitched sequence of all related articles (R in number) and a different single related article sequence. Of these three sequences, the sequence may be denoted X ═ X for sequences that each contain k words1,x2,…,xk}, embedded representation of each word
Figure BDA0002896892890000101
Is a d-dimensional vector that can be initialized by pre-training word embedding vectors. The three sequences are then encoded by means of the BilSTM model and the hidden layer vector e of the last step is usediAs a context representation for each sequence. Thus, the coding characteristics of the false news, the coding characteristics of all related articles, and the coding characteristics of each related article are respectively denoted as ec，
Figure BDA0002896892890000102
And
Figure BDA0002896892890000103
(1≤j≤R)。
step 2: in order to enable the model to focus on core semantic fragments of false news and overall semantic features of all related articles respectively, the invention constructs a mutual interaction sharing module which is composed of a cross attention unit and two gated affine absorbing units, and each false news content and all related articles interact with each other so as to screen out the valuable features suitable for each false news content and all related articles.
And step 3: in the cross attention unit, the present invention uses the self-attention network as a cross attention triplet to explicitly capture the dependency features from word to word in the sequence and learn their intrinsic structural information, ensuring deep interaction between the two sequences, which can be formulated as:
Figure BDA0002896892890000104
wherein Q, K, V are query, key, and value matrices, respectively. In the present invention, in the case of the present invention,
Figure BDA0002896892890000105
and K ═ V ═ ec，dkIs the hidden layer size of BiLSTM in phase 1, which has a value of 2 h.
And 4, step 4: in order to enhance the parallelism of the network, a multi-head attention network pair Q, K and V is designed to be subjected to linear mapping m times from the attention network, and then point-multiplied attention is performed in parallel. The structure of multi-head attention can be formulated as:
Figure BDA0002896892890000111
Hs＝MultiHead(Q,K,V)＝Concat(head1,head2,…,headm)Wo (3)
wherein the content of the first and second substances,
Figure BDA0002896892890000112
Figure BDA0002896892890000113
Figure BDA0002896892890000114
and
Figure BDA0002896892890000115
are all parameters that can be trained, d1Has a dimension of 2H/m and HsIs of a size of
Figure BDA0002896892890000116
HsIs an interactive feature of false news and all related articles.
And 5: in a gated affine absorption cell, consider the H captured in step 4sBelonging to shared interactive features of false news sequences and all related article sequence captures, which lack specific context semantics aiming at respective sequence objects, the invention designs a gated affine absorption unit to enable model capture to be beneficial to the features of respective sequences, so as to adaptively focus on remarkable semantic segment features in the false news sequences and focus on user overall viewpoint semantics in the related article sequences. Specifically, the structure of the gated affine absorption cell can be formulated as:
tc＝tanh(WcHc+bc) (4)
ts＝tanh(WsHs+bs) (5)
α(ec)＝Wαtc+bα (6)
β(Hs)＝Wβts+bβ (7)
γ(Hs)＝Wγts+bγ (8)
Figure BDA0002896892890000117
wherein all W and b are learnable parameters, which indicate a dot product operation from element to element. In particular, gating G2 for all relevant articles and gating G2 for all relevant articlesThe structure of the gate G1 for false news is the same, and
Figure BDA0002896892890000118
and
Figure BDA0002896892890000119
the output features corresponding to the gates G1 and G2, respectively, are the salient features in the false news and the overall opinion semantics in all related articles, respectively.
Step 6: in order to explore semantic conflicts aiming at the meaning of each relevant article, the invention designs a conflict mining module with fine granularity, the structure of the conflict mining module is also a cross attention unit, and the overall view of the user obtained in the stage 2 can be realized
Figure BDA0002896892890000121
Semantics of local single related article
Figure BDA0002896892890000122
Sufficient interaction and fusion. In particular, the cross attention network has been introduced in phase 2, which can be simplified to:
Figure BDA0002896892890000123
wherein the content of the first and second substances,
Figure BDA0002896892890000124
representing the conflict semantics captured for the ith related article.
And 7: in order to eliminate noise features in all relevant article capture conflicts that are not relevant to the false news content and infer which core semantics in the false news are questioned, the invention constructs an evidence-based consistency module to measure the core semantic fragments of the false news for consistency comparisons with the potential conflict semantics of each relevant article obtained in stage 3.
And 8: in this module, two BilSTM modules are first used to separately encode the cores of the fake newsHeart meaning fragment
Figure BDA0002896892890000125
And conflicting semantics for each related article
Figure BDA0002896892890000126
And as their context representation by means of the hidden layer output of BilSTM, which can be represented as
Figure BDA0002896892890000127
And
Figure BDA0002896892890000128
and step 9: then, the invention matches the false news and the salient features of each related article by means of an attention mechanism, which can be formalized as:
Figure BDA0002896892890000129
Figure BDA00028968928900001210
wherein the content of the first and second substances,
Figure BDA00028968928900001211
is a conflict of the ith article
Figure BDA00028968928900001212
The number j of the word (a),
Figure BDA00028968928900001213
and
Figure BDA00028968928900001214
respectively, the original normalized relevance measure of the conflicted jth word to the entire sequence of false news.
Step 10: this is achieved byBy interacting the semantics of the whole false news and the conflict semantics of the related articles, the invention can obtain a conflict representation of false news guidance
Figure BDA00028968928900001215
It can be formulated as:
Figure BDA00028968928900001216
step 11: in order to reinforce the conflict semantic features related to the core semantic segment of the false news, the invention constructs the summation calculation among elements to integrate the overall semantics of the conflict and the coding conflict of the false news guide.
Figure BDA0002896892890000131
Wherein the content of the first and second substances,
Figure BDA0002896892890000132
is a conflicting composite representation of the ith related article,
Figure BDA0002896892890000133
is a summation calculation operation between elements.
Step 12: the invention is to
Figure BDA0002896892890000134
With respect to that obtained in step 8
Figure BDA0002896892890000135
And splicing, and inputting the low-dimensional prediction vector into only one full-connection layer to obtain a low-dimensional prediction vector as a consistency representation between the core segment of the false news and the conflict of the related articles.
Figure BDA0002896892890000136
Step 13: for different single related articles, the invention can obtain their respective consistent prediction vectors in similar operations, i.e.
Figure BDA0002896892890000137
Step 14: finally, the consistency prediction vectors of the related articles are integrated through splicing operation, and the probability distribution of the relevant articles is predicted through the following formula.
Figure BDA0002896892890000138
Wherein, WpAnd bpAre trainable parameters.
Step 15: based on the truth label y, the model is trained by minimization of the cross entropy error:
loss＝-∑ylogp (17)
the invention is suitable for social network environment, and can provide social media network environment for widely discussing relevant articles of false news.
The superiority of the invention's performance was demonstrated by the present invention (EVIN) on two competitive datasets, Snopes and PolitiFact, as shown in Table 1:
TABLE 1 Experimental Performance of the invention under two data sets of Snaps and PolitiFact
Figure BDA0002896892890000141
And the effectiveness of its different separation modules was confirmed, as shown in table 2:
TABLE 2 comparison of separation Performance of different modules of the invention under two datasets of Snapes and PolitiFact
Figure BDA0002896892890000142
Wherein, shared represents removing mutually interactive shared modules, gate represents removing gated affine absorbing units, G1 represents removing gated affine absorbing units based on false news, G2 represents removing gated affine absorbing units based on related texts, conflict represents removing fine-grained conflict mining modules, and coherence represents removing consistency modules based on evidence.
The above is only a preferred embodiment of the present invention, and is not intended to limit the present invention, and it is obvious to those skilled in the art that various modifications and variations can be made in the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (8)

1. A false news interpretability detection method for inferring a network based on evidence is characterized by comprising the following steps:
step 1, inputting a false news content sequence, a splicing sequence of all related articles and a plurality of different single related article sequences, and respectively capturing the internal context hidden layer sequences of the false news content sequence, the splicing sequence of all related articles and the different single related article sequences through BilSTM;
step 2, enabling the false news content sequence to interact with all the related article characteristics, capturing two types of sequence characteristics, focusing core sequence segments concerned by users in the false news content, and learning the overall view sequences of all the users in the related articles;
step 3, the overall viewpoint sequence and the sequence of each related article are interacted with each other, and potential sequence conflicts with fine granularity in each related article are explored;
and 4, carrying out consistency modeling on the core sequence fragments and potential sequence conflicts, and detecting the interpretable false news by surrounding the conflicts of the core sequence fragments of the false news by the spleen and the stomach.
2. The evidence-based inference network false news interpretability detection method of claim 1, wherein the specific method of the step 1 is as follows:
bilstm model for inputting false news content sequences, all related articlesSplicing sequence and several different single related article sequences to code, and adopting hidden layer vector e of last stepiAs a contextual representation of each sequence;
a sequence of false news content, a stitched sequence of all related articles, and several different individual sequences of related articles, all of which can be represented as X ═ X for sequences each containing k words1，x2，...，xk}, embedded representation of each word
Figure FDA0002896892880000011
Is a d-dimensional vector which can be initialized by embedding a pre-training word into the vector;
the coding characteristics of the false news, all related articles and each related article are respectively denoted as ec、
Figure FDA0002896892880000012
And
Figure FDA0002896892880000013
wherein j is more than or equal to 1 and less than or equal to R.
3. The evidence inference network-based false news interpretability detection method of claim 1, wherein in the step 2, each false news content and all related articles are mutually interacted through a cross attention unit and two gated affine absorbing units, and are screened out to be suitable for respective valuable features.
4. The evidence inference network based false news interpretability detection method of claim 3, wherein in the cross attention unit, the self-attention network is used as a cross attention triplet to explicitly capture the dependency features between words in the sequence and learn the intrinsic structure information, ensuring deep interaction between two sequences, and the specific method is as follows:
Figure FDA0002896892880000021
wherein H is the result from the attention network, Q, K, V are query, key and value matrix, respectively;
Figure FDA0002896892880000022
and K ═ V ═ ecT is a transpose operation, dkThe hidden layer size of the BilSTM in the stage 1 is 2 h;
the self-attention network is a multi-head attention network and is used for performing linear mapping on Q, K and V for m times and then performing point-by-point attention in parallel; multi-head attention network headi：
Figure FDA0002896892880000023
Figure FDA0002896892880000024
Wherein the content of the first and second substances,
Figure FDA0002896892880000025
and
Figure FDA0002896892880000026
are trainable parameters, d1Has a dimension of 2H/m and HsIs of a size of
Figure FDA0002896892880000027
HsIs an interactive feature of false news and all related articles.
5. The evidence-inference-network-based false news interpretability detection method of claim 3, wherein the gated affine absorption unit is used for capturing features that favor respective sequences, so as to adaptively focus on salient sequence segment features in false news sequences and focus on user overall opinion sequences in related article sequences; the structure of the gated affine absorption unit is as follows:
tc＝tanh(WcHc+bc) (4)
ts＝tanh(WsHs+bs) (5)
α(ec)＝Wαtc+bα (6)
β(Hs)＝Wβts+bβ (7)
γ(Hs)＝Wγts+bγ (8)
Figure FDA0002896892880000031
wherein, tcAs a result of the transformation based on the news sequence, c is the news sequence, tanh is the activation function, tsFor the transformation results based on the related article sequences, s is the concatenation sequence of all related articles, HcFor hidden layer representation based on news sequences, HsFor hidden layer representation based on related article sequences, α () is an affine result based on news sequences, β () is a first affine result based on related article sequences, γ () is a second affine result based on related article sequences, all W and b are learnable parameters, which indicate a dot product operation from element to element; the structure of gated G2 for all related articles is the same as gated G1 for spurious news, and
Figure FDA0002896892880000032
and
Figure FDA0002896892880000033
output features corresponding to gated G1 and G2, respectively, are the core sequence pieces in the false newsThe sequence of the overall views in the passage and all the related articles.
6. The evidence-based inference network false news interpretability detection method of claim 1, wherein in step 3, the whole view sequence is used by cross attention unit
Figure FDA0002896892880000034
Semantics of local single related article
Figure FDA0002896892880000035
Full interaction and fusion; the cross attention network is:
Figure FDA0002896892880000036
wherein the content of the first and second substances,
Figure FDA0002896892880000037
representing the conflict semantics captured for the ith related article.
7. The evidence-based reasoning network false news interpretability detection method of claim 1, wherein the specific method of the step 4 is as follows:
using two BilSTM modules to separately encode core sequence segments of false news
Figure FDA0002896892880000038
And conflicting semantics for each related article
Figure FDA0002896892880000041
And as its context representation by means of the hidden layer output of BilSTM, respectively
Figure FDA0002896892880000042
And
Figure FDA0002896892880000043
matching the false news and the salient features of each related article by means of an attention mechanism, wherein for the jth word in the ith related article, the attention mechanism is as follows:
Figure FDA0002896892880000044
Figure FDA0002896892880000045
wherein the content of the first and second substances,
Figure FDA0002896892880000046
is a conflict of the ith article
Figure FDA0002896892880000047
The number j of the word (a),
Figure FDA0002896892880000048
and
Figure FDA0002896892880000049
respectively measuring the original normalized relevance of the conflicted jth word to the whole false news sequence;
by interacting the entire sequence of false news with the conflicting sequence of related articles, a false news guide conflict representation can be obtained
Figure FDA00028968928800000410
Figure FDA00028968928800000411
A summation calculation between the elements is constructed to integrate the entire sequence of false news-guide conflicts and coding conflicts:
Figure FDA00028968928800000412
wherein the content of the first and second substances,
Figure FDA00028968928800000413
is a conflicting composite representation of the ith related article,
Figure FDA00028968928800000414
is a summation calculation operation between elements;
will be provided with
Figure FDA00028968928800000415
And
Figure FDA00028968928800000416
splicing is carried out, and then the low-dimensional prediction vector is obtained by inputting the low-dimensional prediction vector into only one full-connection layer and is used as a consistency representation between the core segment of the false news and the conflict of the related articles
Figure FDA00028968928800000417
Figure FDA00028968928800000418
For different individual related articles, respective consistent prediction vectors are obtained in a similar operation, i.e.
Figure FDA00028968928800000419
Integrating consistency prediction vectors of related articles through splicing operation, and predicting probability distribution p of the articles through the following formula:
Figure FDA0002896892880000051
wherein, WpAnd bpAre trainable parameters.
The model is trained by minimization of the cross entropy error:
loss＝-∑ylogp (17)
where y is based on the truth label.
8. A false news interpretability detection system for inferring a network based on evidence, comprising:
the input coding module is used for inputting a false news content sequence, a splicing sequence of all related articles and a plurality of different single related article sequences, and respectively capturing the internal context hidden layer sequence representations of the false news content sequence, the splicing sequence of all related articles and the different single related article sequences through the BilSTM;
the mutual interaction sharing module is used for enabling the false news content sequence to interact with all the related article characteristics, capturing two types of sequence characteristics, focusing core sequence segments concerned by users in the false news content, and learning the overall viewpoint sequences of all the users in the related articles;
the fine-grained conflict mining module is used for mutually interacting the whole viewpoint sequence and the sequence of each related article and exploring the potential fine-grained sequence conflict in each related article;
an evidence-based consistency module for consistency modeling of the core sequence segments against potential sequence conflicts surrounding the core sequence segments of the fake news by the spleen and stomach for interpretable fake news detection.
CN202110045012.6A 2021-01-13 2021-01-13 False news interpretability detection system and method based on evidence inference network Active CN112988959B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110045012.6A CN112988959B (en) 2021-01-13 2021-01-13 False news interpretability detection system and method based on evidence inference network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110045012.6A CN112988959B (en) 2021-01-13 2021-01-13 False news interpretability detection system and method based on evidence inference network

Publications (2)

Publication Number Publication Date
CN112988959A true CN112988959A (en) 2021-06-18
CN112988959B CN112988959B (en) 2023-07-14

Family

ID=76345298

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110045012.6A Active CN112988959B (en) 2021-01-13 2021-01-13 False news interpretability detection system and method based on evidence inference network

Country Status (1)

Country Link
CN (1) CN112988959B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113849599A (en) * 2021-09-03 2021-12-28 北京中科睿鉴科技有限公司 Joint false news detection method based on mode information and fact information
CN116579337A (en) * 2023-07-07 2023-08-11 南开大学 False news detection method integrating evidence credibility

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020061578A1 (en) * 2018-09-21 2020-03-26 Arizona Board Of Regents On Behalf Of Arizona State University Method and apparatus for collecting, detecting and visualizing fake news
KR20200064943A (en) * 2018-11-29 2020-06-08 고려대학교 산학협력단 Fake news detection server and method based on korean grammar transformation
CN111581980A (en) * 2020-05-06 2020-08-25 西安交通大学 False news detection system and method based on decision tree and common attention cooperation
CN111581979A (en) * 2020-05-06 2020-08-25 西安交通大学 False news detection system and method based on evidence perception layered interactive attention network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020061578A1 (en) * 2018-09-21 2020-03-26 Arizona Board Of Regents On Behalf Of Arizona State University Method and apparatus for collecting, detecting and visualizing fake news
KR20200064943A (en) * 2018-11-29 2020-06-08 고려대학교 산학협력단 Fake news detection server and method based on korean grammar transformation
CN111581980A (en) * 2020-05-06 2020-08-25 西安交通大学 False news detection system and method based on decision tree and common attention cooperation
CN111581979A (en) * 2020-05-06 2020-08-25 西安交通大学 False news detection system and method based on evidence perception layered interactive attention network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
BERNHARD LUTZ等: "Identifying Linguistic Cues of Fake News Associated with Cognitive and Affective Processing: Evidence from NeuroIS", 《NEUROIS 2020: INFORMATION SYSTEMS AND NEUROSCIENCE》 *
於张闲等: "基于深度学习的虚假健康信息识别", 《软件导刊》 *
楼靓: "社交网络虚假新闻识别方法", 《浙江交通职业技术学院学报》 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113849599A (en) * 2021-09-03 2021-12-28 北京中科睿鉴科技有限公司 Joint false news detection method based on mode information and fact information
CN116579337A (en) * 2023-07-07 2023-08-11 南开大学 False news detection method integrating evidence credibility
CN116579337B (en) * 2023-07-07 2023-10-10 南开大学 False news detection method integrating evidence credibility

Also Published As

Publication number Publication date
CN112988959B (en) 2023-07-14

Similar Documents

Publication Publication Date Title
Yang et al. Tree-augmented cross-modal encoding for complex-query video retrieval
Guo et al. Deep multimodal representation learning: A survey
Abbasiantaeb et al. Text‐based question answering from information retrieval and deep neural network perspectives: A survey
Mudgal et al. Deep learning for entity matching: A design space exploration
Lopez et al. Deep Learning applied to NLP
Niu et al. Multi-modal multi-scale deep learning for large-scale image annotation
Hassan et al. Sentiment analysis on bangla and romanized bangla text using deep recurrent models
Lin et al. Moment retrieval via cross-modal interaction networks with query reconstruction
Du et al. Gated attention fusion network for multimodal sentiment classification
Chen et al. CAAN: Context-aware attention network for visual question answering
Mohamad Nezami et al. Face-cap: Image captioning using facial expression analysis
Galanopoulos et al. Attention mechanisms, signal encodings and fusion strategies for improved ad-hoc video search with dual encoding networks
CN111581981A (en) Evaluation object strengthening and constraint label embedding based aspect category detection system and method
CN112988959A (en) False news interpretability detection system and method based on evidence inference network
CN116561305A (en) False news detection method based on multiple modes and transformers
Liu et al. Multimodal emotion recognition based on cascaded multichannel and hierarchical fusion
Tajrobehkar et al. Align R-CNN: A pairwise head network for visual relationship detection
Gouthaman et al. Linguistically-aware attention for reducing the semantic gap in vision-language tasks
CN114662586A (en) Method for detecting false information based on common attention multi-mode fusion mechanism
Deng et al. Multimodal affective computing with dense fusion transformer for inter-and intra-modality interactions
Zhong et al. A tree-based structure-aware transformer decoder for image-to-markup generation
Hou et al. Semantic alignment network for multi-modal emotion recognition
Ly et al. An end-to-end multi-task learning model for image-based table recognition
Liu et al. FeatInter: exploring fine-grained object features for video-text retrieval
Zhou et al. Multi-modal multi-hop interaction network for dialogue response generation

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