KR20220114157A - Commonsense question answer reasoning method and apparatus - Google Patents

Abstract

According to the present invention, a commonsense question-answer reasoning method and device are disclosed. According to the present invention, provided is an explainable commonsense question-answer reasoning completion device. The device comprises a memory connected to a processor. The memory stores program instructions executable by the processor to extract a plurality of entities from a plurality of combinations for each of a question and a plurality of answers; extract one or more paths between the extracted entities by referring to a knowledge graph for each of the combinations; convert the one or more paths into sentences through data augmentation; generate a first embedding vector by combining the sentence and the question and generating a second embedding vector for each of the combinations; input the first embedding vector and the second embedding vector to a multi-head attention module to calculate an attention score for each of the combinations; and infer one of the answers corresponding to the question as a correct answer through the attention score. Accordingly, the present invention can grasp a relationship between a question and a correct answer.

Description

Commonsense question answer reasoning method and apparatus

The present invention relates to a common-sense question-answer reasoning method and apparatus.

Commonsense Questions Answering is one of the big challenges facing artificial intelligence, which aims to accurately answer questions for which the answer is not explicitly expressed.

Commonsense question answering is a system that automatically selects one correct answer to a question given a question and candidate answers. CommonsenseQA dataset is used to answer common sense questions with n questions, and each question consists of 5 candidate answers with 1 correct answer and 4 incorrect answers.

When answering the Commonsense question, external common sense knowledge or facts are required.

In this case, knowledge and facts about the answers to these questions can be found in the knowledge graph.

The knowledge graph has factual information in the form of triples (s, r, t), but an incomplete knowledge graph has a problem in that it is difficult to use in an AI-based system.

In the prior art, a method of extracting path (path) information from the knowledge graph and using it to answer the Commonsense Question may be provided, but there is a problem in that it is difficult to find a meaningful answer from the extracted path information due to the incompleteness of the knowledge graph. .

In the prior art, the number of paths is increased to solve this problem, but the generated paths have a problem in that they are not well trained in the neural network model.

Republic of Korea Patent Publication No. 10-2019-0133931

In order to solve the problems of the prior art, the present invention intends to propose a common-sense question answer reasoning method and apparatus capable of grasping a relationship between a question and a correct answer.

In order to achieve the above object, according to an embodiment of the present invention, a memory coupled to the processor is included, wherein the memory extracts a plurality of entities from a plurality of combinations for each of a question and a plurality of answers. and extracting one or more paths between the extracted plurality of entities by referring to the knowledge graph for each of the plurality of combinations, converting the one or more paths into a sentence through data augmentation, and converting the sentence and the question to generate a first embedding vector, generate a second embedding vector for each of the plurality of combinations, and input the first embedding vector and the second embedding vector to a multi-head attention module to calculate the attention score of each of a plurality of combinations, and to infer one of the plurality of answers corresponding to the question as the correct answer through the attention score. An inference completion device is provided.

The program instructions may tokenize the question to extract a plurality of entities, and check whether words corresponding to the plurality of entities extracted from the question exist in the knowledge graph in the knowledge graph.

The program instructions may select a path less than a preset length among paths connecting the entity extracted from the question and the entity extracted from the answer, and convert the selected path into a sentence.

The program instructions translate the one or more paths made of a first language into a second language through back-translation, and translate the translated second language back into the first language to convert the one or more paths into a sentence. can

The program instructions generate the first embedding vector and the second embedding vector by using a RoBERTa model, and the first input data for generating the first embedding vector is a start token <s>, the question, different The second input data for generating the second embedding vector is composed of a token <sep> for separating segments, the sentence, and an end token </s>. It may consist of a token <sep>, one of the plurality of answers, and an end token </s>.

The multi-head attention module includes a plurality of learning models corresponding to the plurality of combinations, and the plurality of learning models includes a second embedding vector for each of the plurality of combinations and a plurality of combinations, respectively. A plurality of attention scores may be calculated by receiving one or more paths extracted from the second embedding vector generated from the transformation sentence as input.

According to another aspect of the present invention, there is provided a method for inferring a common-sense question answer in a device including a processor and a memory, extracting one or more paths between the extracted plurality of entities by referring to a knowledge graph for each of the plurality of combinations. step; converting the one or more paths into sentences through data augmentation; generating a first embedding vector by combining the sentence and the question; generating a second embedding vector for each of the plurality of combinations; calculating an attention score of each of a plurality of combinations by inputting the first embedding vector and the second embedding vector to a multi-head attention module; and inferring one of a plurality of answers corresponding to the question as a correct answer through the attention score.

According to another aspect of the present invention, there is provided a computer readable program for performing the above method.

According to the present invention, it is possible to solve the problem that it is difficult to find a meaningful path using an incomplete knowledge graph through a multi-head attention mechanism by utilizing data augmentation and question answer embedding values.

1 is a diagram illustrating the configuration of a common-sense question answer reasoning apparatus according to a preferred embodiment of the present invention.
2 is a diagram for explaining a common-sense question answer reasoning process according to the present embodiment.
3 is a diagram for explaining a process of embedding a path according to the present embodiment.
4 is a diagram for explaining an embedding process for a combination of a question and an answer according to the present embodiment.

Since the present invention can have various changes and can have various embodiments, specific embodiments are illustrated in the drawings and described in detail.

However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all modifications, equivalents and substitutes included in the spirit and scope of the present invention. In describing each figure, like reference numerals have been used for like elements.

1 is a diagram illustrating the configuration of a common-sense question answer reasoning apparatus according to a preferred embodiment of the present invention.

As shown in FIG. 1 , the device according to the present embodiment may include a processor 100 and a memory 102 .

The processor 100 may include a central processing unit (CPU) capable of executing a computer program or other virtual machines.

Memory 102 may include a non-volatile storage device such as a fixed hard drive or a removable storage device. The removable storage device may include a compact flash unit, a USB memory stick, and the like. Memory 102 may also include volatile memory, such as various random access memories.

The memory 102 stores program instructions executable by the processor 100 .

The program instructions according to the present embodiment extract a plurality of entities from a plurality of combinations for each of a question and a plurality of answers, and refer to a knowledge graph for each of the plurality of combinations to determine one or more paths between the plurality of extracted entities. extracting, converting the one or more paths into sentences through data augmentation, combining the sentences and the questions to generate a first embedding vector, and generating a second embedding vector for each of the plurality of combinations and inputting the first embedding vector and the second embedding vector into a multi-head attention module to calculate an attention score of each of a plurality of combinations, and use the attention score to calculate a plurality of values corresponding to the question. One of the answers is inferred as the correct answer.

In general, Commonsense question answering is to select one of five answers as the correct answer, so path extraction, sentence transformation, embedding, and attention score calculation can be performed for each of five combinations of one question and five answers.

2 is a diagram for explaining a common-sense question answer reasoning process according to the present embodiment.

2 shows an inference process for the question of Where do adult use glue sticks? and one answer of Office.

According to the present embodiment, an inference process for each of a single question and a plurality of answers is performed, and finally, one of the plurality of answers is determined as the correct answer.

Referring to FIG. 2 , a plurality of entities are extracted for questions and answers, and it is checked whether the extracted entities exist in the knowledge graph.

The apparatus according to the present embodiment extracts a plurality of entities by tokenizing a question, and the plurality of entities may be a word corresponding to a noun among several words included in the question and a word including an answer.

This can be defined as a concept matching process and checks whether words that can become entities exist in the knowledge graph using the n-grams technique.

For example, when there is the sentence “What is likely the result of a small episode of falling?”, the plurality of entities may be {result, small, episode, fall}. Conventionally, in the case of “falling asleep”, the knowledge graph There are all three cases that can be matched in {fall, fall_asleep, asleep}.

Semantically, “fall” and “fall_asleep” have different meanings, so they are not useful. In this embodiment, an entity composed of a relatively long combination of words is used to extract entities that do not have different meanings.

Then, one or more paths for a plurality of entities are extracted with reference to the knowledge graph.

According to the present embodiment, one or more paths are extracted as described above in order to avoid a problem in which it is difficult to extract information from a question using an incomplete knowledge graph.

를 가지고 지식 그래프에서 두 엔티티 사이에 존재하는 하나 이상의 경로(path)를 추출한다. 경로의 길이는 짧은 것부터 긴 것까지 다양하게 존재할 수 있다. Entities extracted from questions and answers

Extracts one or more paths that exist between two entities in the knowledge graph. The length of the path may vary from short to long.

According to the present embodiment, a relatively short path (eg, 5 or less) is selected in order to select a meaningful path among a plurality of paths.

: . In the knowledge graph, the paths from the entity extracted from the question to the entity extracted from the answer can be expressed as follows.

: .

Since it is inefficient to use all of the many paths, a score per path is calculated using the random walk probability, and a path to be used for inference is selected through this.

2 exemplarily shows that two paths "adult-->capableof-->work-->atlocation-->office" and "glue_stick-->atlocation-->office" are extracted.

In this embodiment, in order to use the path for learning, data augmentation is used to transform the path into a meaningful sentence.

During data augmentation, the path is converted into a sentence through the back-translation method.

In this method, the route is translated into a specific language and then translated back to the original language.

If a path such as people->capableOf->taste_food<-capableOf<-tongues is extracted, the path is first translated into French through the back-translation method, and then translated back into English, such as "people can taste food with a tongue" create a sentence

The input path and the converted sentence are different, but the meaning is the same even after reverse translation.

According to this embodiment, one or more paths are converted into sentences, and the transformed final sentences and questions are embedded using RoBERTa (transformer-based language model).

This is to generate an embedding vector corresponding to the transformed sentence by embedding the transformed sentence in a multidimensional space.

It can be used by fine-tuning the next sentence prediction task among the RoBERTa models.

3 is a diagram for explaining a process of embedding a path according to the present embodiment.

Referring to FIG. 3 , the input data (first input data) to the RoBERTa model is as follows, a start token <s>, a question (q), a token <sep> for separating different segments, and a transformation that is a data augmentation result value It consists of sentence(s) and end token </s>.

Input Data:

Also, according to the present embodiment, embedding is performed not only on the converted sentence but also on the combination of the question and the answer.

4 is a diagram for explaining an embedding process for a combination of a question and an answer according to the present embodiment.

4 , the following input data (second input data) is used, and the second input data is a start token <s>, a question (q), a token for separating different segments <sep>, and an answer ( Answer choice) and </s> tokens meaning the last.

According to the present embodiment, a semantic difference between an embedding vector generated through first input data including sentences and questions transformed using an attention mechanism and an embedding vector generated through second input data including questions and answers Compare similarities.

Through semantic similarity comparison, it is possible to determine whether the relationship between a question and multiple answers is significant.

5 is a diagram for explaining a process of determining a correct answer to a question according to the present embodiment.

Referring to FIG. 5 , a plurality of learned models corresponding to a plurality of combinations (Question+choice1 to Question+choice5) of one question and a plurality of answers, respectively, are provided, and each model is an embedding vector for each of the plurality of combinations. and one or more paths extracted from each of a plurality of combinations, an embedding vector generated from a transformation sentence is input.

Each model outputs attention scores (score1 to score5) for semantic similarity through comparison of embedding vectors, and determines the highest value as the correct answer to the question through softmax.

The above-described embodiments of the present invention have been disclosed for the purpose of illustration, and various modifications, changes, and additions will be possible within the spirit and scope of the present invention by those skilled in the art having ordinary knowledge of the present invention, and such modifications, changes and additions should be regarded as belonging to the following claims.

Claims (10)

As a common-sense question-answer reasoning completion device,
processor; and
a memory coupled to the processor;
The memory is
extracting a plurality of entities from a plurality of combinations for each of a question and a plurality of answers,
extracting one or more paths between the plurality of extracted entities with reference to the knowledge graph for each of the plurality of combinations,
Transform the one or more paths into sentences through data augmentation,
combining the sentence and the question to generate a first embedding vector,
generating a second embedding vector for each of the plurality of combinations;
inputting the first embedding vector and the second embedding vector to a multi-head attention module to calculate an attention score of each of a plurality of combinations;
to infer one of the plurality of answers corresponding to the question as the correct answer through the attention score;
A descriptive common-sense question-answer reasoning completion device that stores program instructions executable by the processor. According to claim 1,
The program instructions are
A common-sense question answer reasoning completion apparatus for extracting a plurality of entities by tokenizing the question, and checking whether words corresponding to the plurality of entities extracted from the question exist in the knowledge graph in the knowledge graph. According to claim 1,
The program instructions are
A common-sense question answer inference completion device for selecting a path less than a preset length among paths connecting the entity extracted from the question and the entity extracted from the answer, and converting the selected path into a sentence. According to claim 1,
The program instructions are
Common-sense question answer reasoning for translating the one or more paths made of a first language into a second language through back-translation, and translating the translated second language back to the first language to convert the one or more paths into a sentence complete device. According to claim 1,
The program instructions are
generating the first embedding vector and the second embedding vector using the RoBERTa model;
The first input data for generating the first embedding vector consists of a start token <s>, the question, a token separating different segments <sep>, the sentence and an end token </s>,
The second input data for generating the second embedding vector is composed of a start token <s>, the question, a token separating different segments <sep>, one of the plurality of answers, and an end token </s> Common-sense question-answer reasoning complete device. According to claim 1,
The multi-head attention module includes a plurality of learning models corresponding to the plurality of combinations,
The plurality of learning models receive a second embedding vector for each of the plurality of combinations and one or more paths extracted from each of the plurality of combinations, a second embedding vector generated from a transformation sentence, and a common sense to calculate a plurality of attention scores A complete device for question-answer reasoning. A method for inferring common-sense question answers in a device comprising a processor and memory, the method comprising:
extracting one or more paths between the plurality of extracted entities with reference to a knowledge graph for each of the plurality of combinations;
converting the one or more paths into sentences through data augmentation;
generating a first embedding vector by combining the sentence and the question;
generating a second embedding vector for each of the plurality of combinations;
calculating an attention score of each of a plurality of combinations by inputting the first embedding vector and the second embedding vector to a multi-head attention module; and
and inferring one of a plurality of answers corresponding to the question as a correct answer through the attention score. 8. The method of claim 7,
generating the first embedding vector and the second embedding vector using the RoBERTa model;
The first input data for generating the first embedding vector consists of a start token <s>, the question, a token separating different segments <sep>, the sentence and an end token </s>,
The second input data for generating the second embedding vector is composed of a start token <s>, the question, a token separating different segments <sep>, one of the plurality of answers, and an end token </s> A common-sense question-answer reasoning method. 8. The method of claim 7,
The multi-head attention module includes a plurality of learning models corresponding to the plurality of combinations,
The plurality of learning models receive a second embedding vector for each of the plurality of combinations and one or more paths extracted from each of the plurality of combinations, a second embedding vector generated from a transformation sentence, and a common sense to calculate a plurality of attention scores A method of reasoning about answering a question. A computer readable program for performing the method according to claim 7.

Images