CN106705974B

CN106705974B - Semantic role labeling and semantic extraction method for non-restricted path natural language

Info

Publication number: CN106705974B
Application number: CN201611264509.2A
Authority: CN
Inventors: 张珂; 陈奇
Original assignee: North China Electric Power University
Current assignee: North China Electric Power University
Priority date: 2016-12-30
Filing date: 2016-12-30
Publication date: 2020-05-12
Anticipated expiration: 2036-12-30
Also published as: CN106705974A

Abstract

A semantic role labeling and semantic extraction method of an unrestricted path natural language comprises the steps of firstly collecting Chinese path natural language corpora under an unrestricted condition and establishing a Chinese path natural language corpus; then, a semantic role labeling method based on language block analysis and dependency syntax analysis is used for realizing automatic labeling of path natural language corpora; and finally, according to the semantic role labeling result, dividing the path units in sequence, and extracting the navigation semantic information of the path units. The invention uses a semantic role labeling method based on language block analysis and dependency syntax analysis to label the path natural language semantic roles, and divides path units according to the extracted semantic role labeling result, and finally extracts the semantic information of each path unit. The method can accurately divide the path units according to the sequence and accurately extract the path semantic information, so that the smooth implementation of the robot road-asking navigation can be guided.

Description

Semantic role labeling and semantic extraction method for non-restricted path natural language

Technical Field

The invention relates to a semantic role labeling and semantic extraction method of natural language for robot navigation, belonging to the technical field of data processing.

Background

Natural language processing is an important branch in the field of artificial intelligence, which aims at studying methods for communication between computers and people using natural language. Natural language processing is not a simple study of the grammatical and syntactic relationships of natural language, but rather a study of systems that can effectively implement computer communications or human-computer interactions based on natural language, which is part of computer science. The use of natural language to realize human-computer interaction means that a computer can not only understand natural language but also express ideas, intentions, and the like using natural language. The former is natural language understanding, the latter is natural language generation.

People have long sought to use natural language to interact with computers, because it has both important theoretical and obvious practical significance. Firstly, people can use the computer by using the natural language which is used to, and do not need to spend a great deal of time to learn various complex computer languages; secondly, people can further understand the mechanism of human language ability and intelligence through the system.

Language is an important method for human communication, and if human-computer interaction can be performed through natural language, a robot can be controlled through the natural language. If the robot can be controlled in natural language, the robot can be controlled by ordinary people at will. The robot is controlled by the natural language more conveniently than other methods, and the method is more in line with the communication habit of human beings. Modern robot technology is rapidly developed under the promotion of sensor technology, computer technology and artificial intelligence, wherein the mobile robot has mobility and autonomy, and is widely applied to the fields of service, detection, logistics and the like. One of the core technologies of a mobile robot is a navigation technology, particularly an autonomous navigation technology. Among them, autonomous navigation using a natural language controlled robot is becoming a research focus. Researchers hope that the robot can be controlled by natural language to complete a navigation task in the future, and the navigation task is the basis of other complex tasks, so that navigation through natural language is the basis for realizing other navigation tasks, and has important significance for developing artificial intelligence.

In the field of path natural language processing, robot navigation based on english path natural language processing has been developed primarily, and researchers have focused on determining a robot navigation path through research on verbs in combination with corresponding environments. The research on the natural language of the chinese route is still immature, and there is a large gap compared with the natural language processing of the english route. Lezao legendary small dragon with horns et al employ a hierarchical finite state automata-based approach to analyze orientation relationships in natural languages. Zhang Xueliang et al proposed a path natural language processing method based on urban traffic by studying the grammar of Chinese path natural language. Liu Yu et al put forward an NLRP analysis method based on restricted Chinese on the basis of deep research of verbs in natural languages of the routes. A jiang civilization et al performed preliminary analysis of path natural language processing from two natural language processing methods based on statistics and rules. Lexinde et al propose a method for path natural language processing based on block analysis. After a small amount of linguistic data are collected under a specific environment, a plurality of main semantic components are provided, and strong connection between syntax and semantics of the Chinese path natural language is found. The semantic block extraction method comprises the steps of carrying out a series of processing on linguistic data such as word segmentation, noun entity recognition, block analysis and the like, and extracting semantic blocks in path natural languages by adopting a block analysis-based method. And finally, extracting the semantic information effective to the constructed path for each extracted semantic language block according to the corresponding slot. However, these chinese path natural language processing methods all have certain technical problems:

1. in the traditional Chinese path natural language processing method, only a small amount of linguistic data collected in a specific environment are used as research objects, and the actual natural language path description is diversified, so that a larger-scale path natural language corpus under an unrestricted condition needs to be established for path natural language processing research work;

2. under the condition of the unrestricted path natural language, the relationship between the syntactic relationship and the semantics of the path natural language is more complex, and the semantic role marking of the path natural language is difficult to accurately carry out only by the existing semantic role marking method of the path natural language, so that the accuracy of navigation semantic information extraction is influenced;

3. in order to perform robot route-asking navigation, accurate navigation semantic information needs to be obtained on the basis of semantic role labeling, but path natural language often comprises a plurality of ordered path units, so that the problem that how to divide the path units and extract semantic information from each path unit is needed to be solved at present.

Disclosure of Invention

The invention aims to provide a semantic role labeling and semantic extraction method of an unrestricted path natural language aiming at the defects of the prior art so as to improve the accuracy of the robot navigation semantic information extraction.

The problems of the invention are solved by the following technical scheme:

a semantic role labeling and semantic extraction method of an unrestricted path natural language comprises the steps of firstly collecting Chinese path natural language corpora under an unrestricted condition and establishing a Chinese path natural language corpus; then, a semantic role labeling method based on language block analysis and dependency syntax analysis is used for realizing automatic labeling of path natural language corpora; and finally, according to the semantic role labeling result, dividing the path units in sequence, and extracting the navigation semantic information of the path units.

The method for semantic role labeling and semantic extraction of the unrestricted path natural language comprises the following steps of:

a. performing word segmentation and part-of-speech tagging on the linguistic data in the linguistic data base by using an NLPIR Chinese word segmentation system;

b. block analysis of a path-natural language using Conditional Random Fields (CRF)

Firstly, carrying out language block classification on a path natural language corpus, wherein the language block classification comprises 7 language blocks containing semantics and 1 boundary language block, the 7 language blocks containing semantics are respectively direction conversion (DT), target finding according to Reference (RT), target advancing without reference (MT), direct target finding (ST), Preposition Reference (PR), Space Conversion (SC) and target advancing according to reference (FR), and carrying out manual language block labeling on an original language corpus according to the defined language block classification to obtain a language block analysis language corpus; secondly, carrying out language block labeling on each word by adopting an IOB2 labeling method, wherein each word has only three states, namely the word is at the beginning of the language block, the word is in the interior of the language block and the word does not belong to any language block, and the three states are respectively represented by 'B-X', 'I-X' and 'O', wherein 'X' represents the type of the language block; then, determining a language block analysis characteristic template according to the current word and the context environment thereof; finally, training a CRF model according to the customized feature template and the training corpus to obtain the probability distribution of the model, and obtaining the probability value of a word given a certain language block label symbol according to the probability distribution of the model and the context environment of the word, thereby realizing the automatic labeling of the language block;

c. dependency parsing for a path native language using CRF

Firstly, determining a dependency syntax labeling method, and performing dependency syntax labeling on a training corpus: the label of each word is defined as follows: [ +/- ] dPOS, where "+" indicates that the dominant word follows the dependent word, "-" indicates that the dominant word precedes the dependent word, POS indicates the part of speech of the dominant word, and "d" indicates the number of occurrences between the dependent word and the dominant word and the number of occurrences where the dominant word has the same part of speech; then, defining a dependency syntactic analysis characteristic template; finally, a CRF model is trained by using a training corpus according to the dependency syntax analysis feature template, so that dependency syntax automatic labeling is realized;

d. semantic role labeling of path natural language by using CRF (conditional random field) based on language block analysis and dependency syntax analysis results

Firstly, dividing semantic roles into a core semantic role and an additional semantic role; then, semantic role labeling is carried out on the training corpus according to the defined semantic role category; secondly, the results of the language block analysis and the dependency syntax analysis are used as features, and a semantic role labeling feature template is established; and finally, training a CRF (fuzzy F) model by using a training corpus according to the semantic role marking feature template to realize automatic semantic role marking.

The method for semantic role labeling and semantic extraction of the unrestricted path natural language comprises the following steps of dividing path units and extracting navigation semantic information of the path units:

⑴ inputting a complete semantic angle color block and its contained words, part of speech labels, language block labels and dependency syntax labels;

⑵ judging the type of each word block in sequence, if not, turning to step ⑶, if it is, the path unit extraction is finished;

⑶ judging semantic role labeling type, if A0, going to step ⑷, if V, going to step ⑸, if DIR, going to step ⑹, if LOC, going to step ⑺, if A1, going to step ⑼;

⑷ judging the part of speech of the words in A0, filling the words with n (noun) parts of speech into the motion body module in sequence;

⑸ judging the part of speech of the words in V, and filling the words with the part of speech V (verb) into the motion control module in sequence;

⑹ judging the part of speech of the words in the DIR, and filling the words with the part of speech f (direction words) into the direction control module;

⑺ judging the type of the words in LOC, if PR, filling the words into the motion reference module, if SC, going to step ⑻;

⑻ judging the number of words with part of speech n (noun) in SC, if the number of noun is 1, going to step ⑽, if the number of noun is more than 1, going to step ⑾;

⑼ judging the number of words with part of speech n (noun) in A1, if the number of nouns is 1, going to step ⑽, if the number of nouns is more than 1, going to step ⑾;

⑽ judging the dependency relationship of the noun, if the word depends on the verb (v), filling the word into the sports target, if the word depends on the non-verb, adding the depended word into the noun until the new noun depends on the verb, and filling the new noun into the sports target;

⑾ judging the dependency relationship among nouns, if they are parallel, taking them as a whole noun, then going to step ⑽, if one noun a depends on another noun b, filling the noun a into the motion reference module, the noun b is processed according to step ⑽;

⑿ judging whether the motion control module and the motion target module are not empty, if not, turning to step ⑴, if yes, the path unit extraction is finished.

The invention uses a semantic role labeling method based on language block analysis and dependency syntax analysis to label the path natural language semantic roles, and divides path units according to the extracted semantic role labeling result, and finally extracts the semantic information of each path unit. The method can accurately divide the path units according to the sequence and accurately extract the path semantic information, so that the smooth implementation of the robot road-asking navigation can be guided.

Drawings

FIG. 1 is a flow diagram of semantic role labeling;

fig. 2 is a flow chart of path unit division and semantic extraction.

Detailed Description

The invention will be further explained with reference to the drawings.

The invention provides a Chinese-oriented path natural language processing method, which comprises the following 3 steps:

1. collecting Chinese path natural language corpora under an unlimited condition, and establishing a Chinese path natural language corpus.

2. And realizing automatic labeling of path natural language corpora by using a semantic role labeling method based on language block analysis and dependency syntax analysis.

3. And according to the semantic role labeling result, dividing the path units in sequence, and extracting navigation semantic information of the path units.

Step 1, collecting Chinese path natural language linguistic data under the unrestricted condition, and establishing a Chinese path natural language corpus. In order to collect sufficient unrestricted path natural language corpora, the invention uses Webots for Nao to build 10 3D simulation environments, then records 10 videos according to the simulation environments and gives a top view of each environment, and gives a navigation task for the robot in each environment (namely giving an initial position and a target position of the robot). Under the condition of not informing task details, 100 volunteers with different ages and different professions freely select a path capable of completing a navigation task and give corresponding natural language path description by watching videos and top views, in order to ensure the generality of the corpus, the selected 100 volunteers are aged from 12 years to 60 years, and the education degree is from schoolchildren to major, and thus the volunteers are distributed all over the country. Volunteers freely select descriptive sentences in an unrestricted environment, and finally 1000 linguistic data are collected to form a Chinese path natural language corpus.

The 2 nd step is semantic role labeling based on language block analysis and dependency syntax analysis, and the specific flow is shown in fig. 1.

(1) And performing word segmentation and part-of-speech tagging on the linguistic data in the linguistic database by using an NLPIR Chinese word segmentation system.

(2) Conditional Random Fields (CRF) are used to perform block analysis on the path natural language. Firstly, the path natural language corpus is divided into language block categories, and the invention determines 7 language blocks containing semantics and 1 boundary language block, as shown in table 1. And manually labeling the language blocks of the original language database according to the defined language block types to obtain a language block analysis language database. Secondly, for the problem of language block analysis, carrying out language block labeling on each word, specifically adopting an IOB2 labeling method, wherein each word has only three states, namely the word is at the beginning of the language block, the word is in the interior of the language block and the word does not belong to any language block, and the three states are respectively represented by 'B-X', 'I-X' and 'O', wherein 'X' represents the type of the language block. Then, a speech block analysis feature template is determined from the current word and its context, and the speech block analysis feature template defined by the present invention is shown in table 2. And finally, training a CRF model by using a training corpus according to the feature sets, namely training to obtain an optimal parameter vector Lambda ═ Lambda according to a customized feature template and the training corpus₁,…λ_KAnd after parameter estimation of the CRF model is completed, obtaining probability distribution of the model, and obtaining a probability value of a word given with a certain word block label symbol according to the probability distribution of the model and the context environment of the word, thereby realizing automatic labeling of the word block.

TABLE 1 language block definitions

TABLE 2 feature templates for language block analysis

(3) And performing dependency syntax analysis on the path natural language by using the CRF. Firstly, determining a dependency syntax labeling method, and performing dependency syntax labeling on a training corpus. The label of each word is defined as follows: [ +/- ] dPOS, where "+" indicates that the dominant word follows the dependent word, "-" indicates that the dominant word precedes the dependent word, POS indicates the part of speech of the dominant word, and "d" indicates the number of words present between the dependent word and the dominant word that have the same part of speech. Then, the present invention defines a dependency parsing feature template, as shown in Table 3. And finally, training a CRF model by using a training corpus according to the feature sets, and realizing automatic labeling of dependency syntax.

TABLE 3 feature templates for dependency parsing

(4) And performing semantic role labeling on the path natural language by using the CRF based on the result of the language block analysis and the dependency syntax analysis. First, the present invention divides semantic roles into core semantic roles and additional semantic roles, Arg + number (abbreviated with a + number) representing core semantic roles (core identification), wherein Arg0 generally represents the action's event, Arg1 generally represents the action's event, Arg2-Arg4 have different semantic meanings according to the predicate verb, V is also a core semantic role representing a predicate, ArgM-represents an adjunct component, where x represents the function of the adjunct component, and the adjunct component is classified as shown in table 4. And then, semantic role labeling is carried out on the training corpus according to the defined semantic role category. Then, the result of the speech block analysis and the dependency syntax analysis is used as a feature to establish a semantic role labeling feature template, and the semantic role labeling feature template based on the speech block analysis and the dependency syntax analysis defined by the present invention is shown in table 5. And finally, training a CRF model by using a training corpus according to the feature sets to realize automatic semantic role labeling.

TABLE 4 adjunct ingredients

TABLE 5 semantic role tagging feature templates based on chunk analysis and dependency syntax analysis

And 3, dividing the path units according to the semantic role labeling result, and further extracting the navigation semantic information of each path unit. A path element consists of four parts: motion body, motion direction, motion control, motion target. The path unit division and semantic extraction flow is shown in fig. 2.

(1) And inputting a complete semantic angle color block and words, part of speech labels, language block labels and dependency syntax labels contained in the semantic angle color block.

(2) Judging the type of each word block in sequence. If not, turning to the step (3); if the path is 'B-PT', the path unit extraction is finished.

(3) And judging the semantic role marking type. If A0, go to step (4); if the voltage is V, turning to the step (5); if the DIR exists, turning to the step (6); if LOC, go to step (7); if A1, go to step (9).

(4) And judging the part of speech of the word in A0, and filling the word with the part of speech n (noun) into the motion body module in sequence.

(5) And judging the part of speech of the words in the V, and sequentially filling the words with the part of speech V (verbs) into the motion control module.

(6) And judging the part of speech of the words in the DIR, and sequentially filling the words with the part of speech f (directional words) into the direction control module.

(7) And judging the type of the language block of the words in the LOC. If PR, fill the word into the motion reference module. If the result is SC, go to step (8).

(8) The number of words with part of speech n (noun) in SC is determined. If the noun number is 1, go to step (10); if the number of nouns is greater than 1, go to step (11).

(9) The number of words with part of speech n (noun) in a1 is determined. If the noun number is 1, go to step (10); if the number of nouns is greater than 1, go to step (11).

(10) The dependency relationship of the noun is determined. If the word depends on the verb (v), filling the word into the moving target; if the word depends on the non-verb, the depended word is added into the noun until the new noun depends on the verb, and then the new noun depends on the verb and is filled into the motion target.

(11) And judging the dependency relationship among the nouns. If the nouns are in parallel relationship, then the nouns are taken as a whole noun, and the step (10) is carried out; if one of the nouns a depends on the other noun b, filling the noun a into the motion reference module, and processing the noun b according to the step (10).

(12) And judging whether the motion control module and the motion target module are not empty. If not, turning to the step (1); if yes, the path unit extraction is completed.

The invention provides a path natural language semantic role labeling method based on language block analysis and dependency syntax analysis, aiming at the characteristic that the relationship between the syntax relationship and the semantic of a path natural language is more complex under the condition of an unlimited path natural language. Firstly, according to the characteristics of a path natural language, a proper language block type and dependency syntax type definition method is provided, a language block analysis characteristic template and a syntax analysis characteristic template are designed, and CRF is utilized to realize high-accuracy language block automatic labeling and dependency syntax automatic labeling; secondly, in order to extract path semantic information facing to robot question-path navigation, a new semantic role category and a semantic role labeling characteristic template based on language block analysis and dependency syntax analysis are designed, and high-accuracy path natural language semantic role labeling is realized by adopting CRF.

In order to perform robot route-asking navigation, accurate navigation semantic information needs to be obtained on the basis of semantic role labeling, but a path natural language often comprises a plurality of ordered path units. The invention provides a path unit dividing and semantic extracting method based on semantic role marking information, which can accurately divide path units according to a sequence and simultaneously accurately extract path semantic information, thereby guiding a robot to smoothly implement road asking navigation.

Experimental analysis:

to illustrate the above advantages, experimental methods were employed for verification. Dividing all the corpora in the corpus according to the proportion of 4:1, wherein 80% of the corpora are training corpora, and 20% of the corpora are testing corpora. And then training a training corpus according to the defined characteristic template to obtain a corresponding training model, and testing the accuracy of the obtained training model by using a test corpus.

Experiment one: speech block analysis experiment

Table 6 shows the analysis results of the respective blocks. The accuracy P and the recall R, F1 of the speech block analysis are very high, which shows that the partition type determined by the invention is reasonable and is suitable for path natural language processing. The higher speech block analysis accuracy rate also meets the requirement of semantic role labeling on the basis of the subsequent speech block analysis.

TABLE 6 analysis of the blocks (%)

Experiment two: dependency syntactic analysis experiment

Table 7 shows the dependency parsing result. The automatically labeled mean value using dependency syntax represents the dependency syntax analysis result, which, although relatively low in accuracy, recall, and F1 values as compared to the chunk analysis, is still numerically high and is complementary to the chunk analysis from the semantic character label recognition perspective.

Table 7 dependency syntax analysis result (%)

Experiment three: semantic role labeling experiment based on language block analysis and dependency syntax analysis

Table 8 shows semantic role labeling results based on chunk analysis and dependency syntax analysis. The accuracy, the recall rate and the F1 value of the method are high, which indicates the effectiveness of the method; and for the robot navigation task, the path unit is mainly composed of the following 4 parts: the predicate, the direction word, the reference and the target are exactly in V, AM-DIR and A1 respectively, and the labeling effect of the predicate, the direction word, the reference and the target is better than the average value, which shows that the method of the invention is very suitable for extracting the path semantic information of the path natural language.

TABLE 8 semantic role labeling results based on chunk analysis and dependency syntax analysis (%)

Experiment four: the method of the invention is utilized to realize the path semantic information extraction process of the example sentence.

Example sentence: the robot walks forwards to the bed, leftwards to the chair, turns right, walks to the wall, leftwards to the doorway, passes through the door, enters the living room, walks forwards to the vase, rightwards to the space between the sofa and the table, turns left, and takes the pop can on the table.

Word segmentation and part of speech tagging results:

the robot/n forward/vi walk/v to/v bed/n,/wd go/p left/f walk/v to/v chair/n side/f,/wd go/d right/f turn/v,/wd straight/d walk/v to/v wall/n, go/p left/f walk/v to/v doorway/s,/wd go/pdiref pass/v door/n,/wd enter/v living room/n,/wd forward/vi walk/v to/v vase/n side/f,/wd go/p right/f walk/v to/v sofa/n and/cc table/n side/f, the/wd turns to/p left/f/v, the/wd brings/v table/n/f/ude 1 pop can/n. /wj

And (3) analyzing the language blocks:

[ robot/n Forward/vi walk/v to/v bed/n ] MT [,/wd ] PT [ go/p left/f ] DT [ go/v to/v chair/n side/f ] MT [,/wd ] PT [ re/d ] O [ go/p right/f turn/v ] DT [,/wd ] PT [ go/v to/v wall/n ] MT [,/wd ] PT [ go/p left/f ] DT [ go/v to/v doorway/s ] MT [ go/p right/f ] DT [,/wd ] PT [ pass/v door/n ] SC [,/wd ] PT [ enter/v living room/n ] SC [,/wd ] PT [ go/v side/n/f ] PT [ forward/vi ] O [ go/v to/v/n side/f ] MT [, the/wd PT [ go/p right/f ] DT [ go/v to/v sofa/n and/cc table/between/f MT [,/wd PT [ go/p left/f turn/v ] DT [,/wd PT [ na/v table/n on/f/ude 1 pop can/n ] RT [. /wj ] PT

Dependency parsing results:

[ robot/n ] +1V/SBV [ forward/vi ] +1V/ADV [ walk/V ]0/ROOT [ to/V ] -1V/CMP [ bed/n ] -1V/POB [, wd ] O [ go/P ] +1V/ADV [ left/F ] -1P/POB [ walk/V ] -2V/COO [ to/V ] -1V/CMP [ chair/n ] +1F/ATT side/F ] -1V/POB [,/wd ] O [ re/d ] +1V/ADV [ go/P ] +1/ADV [ right/F ] -1P/POB [ turn/V ] -2V/COO [,/wd ] O [ go/d ] +1V/ADV [ walk/V ] -1V/COO [ to/V ] -1V/CMP [ wall/n ] -1V/POB [,/wd ] O [ to/P ] +1V/ADV [ left/F ] -1P/POB [ walk/V ] -2V/COO [ to/V ] -1V/CMP [ door/s ] -1V/POB [,/wd ] O [ to/P ] +1V/ADV [ right/F ] -1P/POB [ cross/V ] -2V/COO [ door/n ] -1V/VOB [,/wd ] O [ enter/V ] -1V/COO [ living room/n ] -1V/VOB [,/wd ] O [ forward/vi ] +1V/ADV [ walk/V ] -1V/COO [ to/V ] -1V/CMP [ vase/n ] +1F [,/ ATT [ side/F ] -1V/POB [,/wd ] O [ side/P ] +1V/ADV [ right/F ] -1P/POB [ walk/V ] -2V/COO [ side/V ] -1V/CMP [ sofa/N ] +1F/ATT [ and/cc ] +1N/RAD [ table/N ] -1N/COO [ between/F ] -1V/POB [,/wd ] O [ side/P ] +1V/ADV [ left/F ] -1P/POB [ turn/V ] -2V/COO [,/wd ] O [ na/V ] -1V/COO [ table/N ] +1F/ATT [ up/F ] +1N/ATT [/ude 1] -1F/ATT [ can/N ] -1V/ VOB [. W j O

Semantic role labeling result:

[ robot/n ] A0[ Forward/vi ] ADV [ go/V to/V ] V [ bed/n ] A1[,/wd ] O [ go/p left/f ] AM-DIR [ go/V to/V ] V [ chair/n side/f ] A1[,/wd ] O [ re/d ] AM-ADV [ go/p right/f ] AM-DIR [ turn/V ] V [,/wd ] O [ straight/d ] AM-ADV [ go/V to/V ] V [ wall/n ] A1[,/wd ] O [ go/p left/f ] AM-DIR [ go/V to/V ] doorway V [ A1[,/wd ] O [ go/V/s ] A1[,/p right/f ] AM-DIR [ pass/V ] V [ door/n ] A1[, [ go/V ] door/n ] A1[, [ w ]0[ enter/V ] V [ living room/n ] A1[, [ w ] d ] [ forward/vi ] AM-ADV [ go/V to/V ] V [ vase/n side/f ] A1[, [ w ] O [ go/p right/f ] AM-DIR [ go/V to/V ] V [ sofa/n and/cc table/n between/f ] A1[, [ w ] O [ go/p left/f ] AM-DIR [ turn/V ] V [, [ w ] O [ V ] table/n/f/ude 1 can/n ] A1 [. W j O

The following 9 path units are obtained in sequence.

1. A motion body: robot

The direction is as follows: air (forward default)

And (3) motion control: walking machine

Moving the target: bed

2. The direction is as follows: left side of

And (3) motion control: walking machine

Moving the target: chair (Ref. TM. chair)

3. The direction is as follows: right side

And (3) motion control: walking machine

Moving the target: wall with a plurality of walls

4. The direction is as follows: left side of

And (3) motion control: walking machine

Moving the target: door opening

5. The direction is as follows: right side

And (3) motion control: through the hole

Moving the target: door with a door panel

6. The direction is as follows: air conditioner

And (3) motion control: enter into

Moving the target: parlor

7. The direction is as follows: air conditioner

And (3) motion control: walking machine

Moving the target: flower vase

8. The direction is as follows: right side

And (3) motion control: walking machine

Moving the target: between sofa and table 9, direction: left side of

And (3) motion control: take

Moving the target: a pop can on the table.

Claims

1. A semantic role labeling and semantic extraction method of an unrestricted path natural language is characterized in that the method comprises the steps of firstly collecting Chinese path natural language corpora under an unrestricted condition and establishing a Chinese path natural language corpus; then, a semantic role labeling method based on language block analysis and dependency syntax analysis is used for realizing automatic labeling of path natural language corpora; finally, according to the semantic role labeling result, dividing the path units according to the path natural language sequence, and extracting the navigation semantic information of the path units;

the automatic labeling method for the path natural language corpus comprises the following steps:

b. conditional random field CRF (conditional random field) for language block analysis of path natural language

Firstly, carrying out language block classification on a path natural language corpus, wherein the language block classification comprises 7 language blocks containing semantics and 1 boundary language block B-PT, the 7 language blocks containing semantics are respectively direction conversion DT, a reference searching target RT, a non-reference target advancing MT, a direct searching target ST, a preposition reference PR, a space conversion SC and a reference advancing FR, and carrying out manual language block labeling on an original language corpus according to the defined language block classification to obtain a language block analysis language corpus; secondly, carrying out language block labeling on each word by adopting an IOB2 labeling method, wherein each word has only three states, namely the word is at the beginning of the language block, the word is in the interior of the language block and the word does not belong to any language block, and the three states are respectively represented by 'B-X', 'I-X' and 'O', wherein 'X' represents the type of the language block; then, determining a language block analysis characteristic template according to the current word and the context environment thereof; finally, training a CRF model according to the customized feature template and the training corpus to obtain the probability distribution of the model, and obtaining the probability value of a word given a certain language block label symbol according to the probability distribution of the model and the context environment of the word, thereby realizing the automatic labeling of the language block;

c. dependency parsing for a path native language using CRF

Firstly, determining a dependency syntax labeling method, and performing dependency syntax labeling on a training corpus: the label of each word is defined as follows: [ +/- ] dPOS, where "+" denotes that the dominant word follows the dependent word, "-" denotes that the dominant word precedes the dependent word, POS denotes the part of speech of the dominant word, "d" denotes a number, which denotes the number of words that exist between the dependent word and the dominant word and that have the same part of speech as the dominant word; then, defining a dependency syntactic analysis characteristic template, wherein semantic role marking types comprise an action actor A0, an action victim A1, a predicate V, an adverb mark ADV, a location lattice LOC and a direction mark DIR; finally, a CRF model is trained by using a training corpus according to the dependency syntax analysis feature template, so that dependency syntax automatic labeling is realized;

Firstly, dividing semantic roles into a core semantic role and an additional semantic role; then, semantic role labeling is carried out on the training corpus according to the defined semantic role category; secondly, the results of the language block analysis and the dependency syntax analysis are used as features, and a semantic role labeling feature template is established; finally, training a CRF (model frequent learning) model by using a training corpus according to the semantic role marking feature template to realize automatic semantic role marking;

the steps of dividing the path unit and extracting the navigation semantic information of the path unit are as follows:

⑵ judging the type of each word block in sequence, if not, turning to step ⑶, if it is, then path unit extraction is completed;

⑷ judging the part of speech of the words in A0, and filling the words with n parts of speech into the motion main body module;

⑸ judging the part of speech of the V words, and filling the sequence of the V words into the motion control module;

⑹ judging the part of speech of the words in the DIR, and filling the words with the part of speech of f direction words into the direction control module;

⑻ judging the number of words with part of speech n in SC, if the number of nouns is 1, going to step ⑽, if the number of nouns is more than 1, going to step ⑾;

⑼ judging the number of the words with part of speech n in A1, if the number of the nouns is 1, going to step ⑽, if the number of the nouns is more than 1, going to step ⑾;

⑽ judging the dependency relationship of the noun, if the word depends on the verb v, filling the word into the sports target, if the word depends on the non-verb, adding the depended word into the noun until the new noun depends on the verb, and filling the new noun into the sports target;