An Event Opinion Mining Method Based on Argument Structure

Abstract

The present invention discloses an event opinion mining method based on argument structure. The steps include: 1) For a given event e and a document d, extract all opinion fragments related to the given event e from the document d; 2) For each of the opinion fragments, extract an event sub-structure expressing an opinion from the event e as the opinion object corresponding to the opinion fragment; 3) According to the processing result of step 2), obtain the event opinion set T of the event e = {…, (o k , a k ), …|e, d}; where o k is the k-th opinion fragment in the document d; a k is the opinion object corresponding to the opinion o k . The present invention is not limited to opinions with emotions and is not limited to obtaining opinions on the event itself. It can obtain richer opinion information, including opinions on the event itself, event participants, event sub-events, and related events.

Description

Technical Field

[0001] The present invention relates to the task of opinion mining, in particular to event-centered opinion mining, and belongs to the technical field of natural language processing. Background Art

[0002] Opinion mining is a key task in natural language processing, aiming to identify people's opinions and emotions towards entities, events, and their substructures from given texts. For example, given the sentence "The speed of XX mobile phone is very fast", an opinion mining system should be able to identify that "very fast" is the author's opinion on the "speed" of "XX mobile phone".

[0003] In recent years, due to the difficulty in dividing the structure of events, people have focused on entity-based opinion mining, which is defined as extracting (attribute, opinion, type, sentiment) quadruples from given texts. For example, in the sentence "The speed of XX mobile phone is very fast", an entity-centered opinion mining system will extract the opinion quadruple ("speed", "very fast", performance, positive). The entity-centered opinion mining task has developed relatively maturely from the extraction of single elements of opinions or attributes to the extraction of quadruples.

[0004] However, the task definition and model of the entity-centered opinion mining task cannot be directly applied to the event-centered opinion mining task. First, entity-centered opinions mainly focus on the sentiment polarity of opinion holders, while event-centered opinions are more concerned with non-emotional predictions, judgments, or suggestions, etc. Second, event-centered opinions have a more complex structure. Given an event, people can express their views on the event itself, sub-events, related events, and related entities. Third, event-centered opinions have unique expression characteristics. The target of event-centered opinions usually does not directly appear in the opinion text, while the target of entity-centered opinions is usually explicit. Event-centered opinions usually appear in long news and articles, which are mixed with a large amount of factual information. In contrast, entity-centered opinions mainly appear in short and concentrated comments or tasks. Therefore, it is necessary to design a theoretical framework for event-centered opinion mining so that opinions on events can be mined at a fine-grained level. Summary of the Invention

[0005] To overcome the deficiency that the existing opinion mining tasks cannot perform fine-grained mining of event-centered opinions, the present invention proposes an event opinion mining method based on argument structure, which includes: 1. Opinion definition based on representational classification theory; 2. Event structure division based on argument structure; 3. A new event-centered opinion mining task based on the above opinion definition and event structure; 4. A two-stage task framework based on the above task definition. The opinion definition based on representational classification theory can break through the limitation of traditional opinion mining tasks that only study opinions expressing emotions, and explore more valuable opinion categories for events in a deeper way, including five categories: judgment, attitude, belief, emotion, and suggestion. The event structure division based on argument structure enables the opinions on events to be mined in a fine-grained manner.

[0006] The technical solution of the present invention is as follows:

[0007] An event opinion mining method based on argument structure, the steps of which include:

[0008] 1) For a given event e and document d, extract all opinion fragments related to the given event e from the document d;

[0009] 2) For each of the said opinion fragments, extract the event sub-structure expressing the opinion from the event e as the opinion object corresponding to the said opinion fragment;

[0010] 3) According to the processing result of step 2), obtain the event opinion set T of event e = {...,(o k ,a k ),...|e,d}; where, o k is the kth opinion fragment in the document d; a k is the opinion object corresponding to the opinion o k .

[0011] Furthermore, the opinion objects of the said opinion fragments include the event itself, event sub-events, and event participants.

[0012] Furthermore, the said event itself means that the opinion fragment directly expresses an opinion on the entire event; the event sub-event means that the opinion fragment expresses an opinion on a sub-event or related event of the event; the participant means that the opinion fragment directly expresses an opinion on the entity involved in the event.

[0013] Furthermore, the method for extracting all opinion fragments related to the given event e from the document d is:

[0014] 1) Concatenate each sentence in the given event e and the document d using the general concatenation symbol of BERT to construct the input: [CLS] event phrase [SEP] document sentence [SEP];

[0015] 2) Input the spliced text into the Transformer encoder;

[0016] 3) Input the representation corresponding to the start symbol [CLS] into the Softmax layer for classification. If the output result is 1, then the corresponding sentence is an opinion sentence related to event e; if the output result is 0, it is not an opinion sentence related to event e;

[0017] 4) Connect the consecutive opinion sentences in document d to obtain the opinion fragment.

[0018] Further, the method for obtaining the opinion object of the opinion fragment is as follows:

[0019] 1) Segment event e to obtain a segmentation set {w1, w2, …, w m}; where event e is a text phrase containing m words, and w m is the mth word in event e; document d contains n sentences, and s n is the nth sentence in document d;

[0020] 2) Randomly combine consecutive segments in the event phrase to obtain multiple sub - phrases; use each of the sub - phrases as a candidate opinion object;

[0021] 3) Concatenate each candidate opinion object with the opinion fragment using the concatenation symbol of the Bidirectional Encoder Representations from Transformers (BERT) technology to construct the input: [CLS] candidate opinion object [SEP] opinion fragment [SEP];

[0022] 4) Input the concatenated text into the Bidirectional Encoder Representations from Transformers (BERT) technology to obtain the matching degree between the candidate opinion object and the corresponding opinion fragment, and select the candidate opinion object with the highest matching degree as the opinion object corresponding to the opinion fragment.

[0023] Further, use the "jieba" Chinese word segmentation component to segment event e to obtain a segmentation set {w1, w2, …, w m}.

[0024] Further, the categories of the opinion fragment include: judgment, attitude, belief, emotion, and suggestion.

[0025] A server, characterized in that it includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing the steps in the above - mentioned method.

[0026] A computer - readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the steps of the above - mentioned method.

[0027] Based on the difference from the entity - centered perspective, the present invention establishes a complete theoretical framework for the event - centered opinion mining task:

[0028] 1. Opinion definition based on representational classification theory

[0029] For the event opinion mining task, what users really want to understand is not only the simple positive or negative sentiment tendency, but also the specific views and judgments on the event. Compared with the entity - centered opinion definition, we need a more generalized interpretation of the event - centered opinion. Therefore, the present invention proposes an opinion definition based on representational classification theory.

[0030] The present invention first starts from the perspective of the difference between opinion and fact, and defines the event - centered opinion as a declarative sentence expressing the views of a person or organization on an event or related topic, and needs to meet the following specific characteristics: 1) Based on assumptions, it cannot be verified as correct or wrong by current evidence including practice, eyewitnesses, or documentary records; 2) It is not objectively existent and will change according to the change of the opinion holder.

[0031] The present invention also, based on representational classification theory, defines the event - centered opinion into the following five categories: 1) Judgment, such as inference, explanation, prediction of the future development of the event; 2) Attitude, such as the stance on a controversial event and the evaluation of people, places, things, and objects; 3) Belief, such as the existence of God; 4) Emotion, such as happiness and fear; 5) Suggestion, such as the guidance of the author to the reader's actions.

[0032] 2. Event structure division based on argument structure

[0033] The event structure division based on argument structure enables the opinion mining system to conduct opinion mining on the event itself, event sub - events, event - related events, and event participants, thus making fine - grained opinion mining for events possible. Specifically, the present invention divides the event structure into the following three categories: 1) The event itself, which means that the opinion fragment directly expresses views on the entire event; 2) Sub - events, which means that the opinion fragment expresses views on the sub - events or related events of the event; 3) Participants, which means that the opinion fragment directly expresses views on the entities involved in the event.

[0034] 3. A new event - centered opinion mining task based on the above - mentioned opinion definition and event structure

[0035] Based on the above - mentioned opinion definition and event structure division, the present invention proposes and defines a new event - centered opinion mining task. Specifically, given an event e = {w1, w2, …, w m} and a document d = {s1, s2, …, s n}, the task objective is to extract all opinion fragments related to a given event from the document, along with their types and corresponding specific opinion objects, that is, the triple set T = {...,(o k ,t k ,a k ),...|e,d}. Among them, the event e is a text phrase consisting of m words; the document d is an article containing n sentences; o k = {s i ,s i+1 ,…,s j |s ∈ d} is the k-th opinion fragment in the document; t k is the opinion type of the opinion o k , which is one of judgment, attitude, belief, emotion, and suggestion; a k = {w t ,w t+1 ,…,w l |w ∈ e} is the event sub-structure corresponding to the expressed opinion, that is, the opinion object of the opinion o k .

[0036] Taking the event "The exchange postponed the listing of Group A" and the document "The listing of Group A has been put on hold, triggering speculation and discussions among some people at home and abroad. Postponing the listing of Group A is to maintain the long-term healthy development of the capital market. The top priority for Group A is to earnestly rectify in accordance with the requirements of the regulatory authorities. Group A should set an example in all aspects such as operating legally, preventing risks, and fulfilling social responsibilities." as an example. The event sub-structure includes the event itself "The exchange postponed the listing of Group A", the sub-event "The listing of Group A", and the participants "The exchange" and "Group A". This document includes two opinion fragments related to the event. The first opinion fragment "Postponing the listing of Group A is to maintain the long-term healthy development of the capital market." is an opinion of the judgment type regarding the event sub-structure "The exchange postponed the listing of Group A", and the second opinion fragment "The top priority for Group A is to earnestly rectify in accordance with the requirements of the regulatory authorities. Group A should set an example in all aspects such as operating legally, preventing risks, and fulfilling social responsibilities." is an opinion of the suggestion type regarding the event sub-structure "Group A".

[0037] Based on the above task definition, the present invention proposes a three-stage framework for solving the event-centered opinion mining task:

[0038] 1. Event-centered opinion fragment extraction. Given an event and a relevant document, extract all opinion fragments related to the event from the document.

[0039] 2. Opinion category classification. Given the already extracted opinion fragments, judge the types of the opinion fragments.

[0040] 3. Opinion object extraction. Based on the already extracted opinion fragments and the given event phrases, extract the event sub-structures expressing opinions from the event phrases for each opinion fragment as its opinion object.

[0041] Based on the above two-stage framework, the present invention implements a baseline system, specifically including:

[0042] 1. Event-centered opinion fragment extraction

[0043] a) Concatenate the given event and each sentence in the document using the general concatenation symbol of the Transformer-based bidirectional encoder representation technique BERT to construct the input: [CLS] Event phrase [SEP] Document sentence [SEP].

[0044] b) Input the concatenated text into the Transformer encoder to obtain the vector representation corresponding to each word in the input.

[0045] c) Take the representation corresponding to the start symbol [CLS] and input it into the Softmax layer for classification. If the output result is 1, then this sentence is an opinion sentence related to the event; if the output result is 0, this sentence is not an opinion sentence related to the event.

[0046] d) Connect the consecutive opinion sentences in the document to obtain an opinion fragment.

[0047] 2. Opinion category classification

[0048] a) For each opinion sentence in the opinion fragment, take the representation corresponding to [CLS] in step c) of step 1 and input it into the Softmax layer for classification. The output results 0, 1, 2, 3, 4 represent five types of opinions respectively, judgment, attitude, belief, emotion, and suggestion.

[0049] b) Take the sentence type that appears the most in the opinion fragment as the type of the corresponding opinion fragment.

[0050] 3. Opinion object extraction

[0051] a) Segment the event phrase using the "Jieba" Chinese word segmentation component.

[0052] b) Randomly combine consecutive words in the event phrase to obtain all sub-phrases of the event phrase as candidate opinion objects.

[0053] c) Concatenate each candidate opinion object with the opinion fragment using the general concatenation symbol of BERT to construct the input: [CLS] Candidate opinion object [SEP] Opinion fragment [SEP].

[0054] d) Input the concatenated text into BERT to obtain the matching degree between the candidate opinion objects and opinion segments, and select the candidate opinion object with the highest matching degree as the final opinion object.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] 1. The opinion definition based on the representational classification theory enables the opinion mining system to identify more valuable opinion categories for events, such as judgments, attitudes, suggestions, and beliefs, not limited to sentiment-based opinions.

[0057] 2. The event structure division based on the argument structure enables the opinion mining system to perform fine-grained opinion mining for a given event, not limited to obtaining opinions on the event itself, but also able to obtain richer opinion information, including opinions on the event itself, event participants, event sub-events, and related events. Description of the Drawings

[0058] Figure 1 It is a three-stage model architecture for the event-centered opinion mining task. Detailed Embodiments

[0059] The present invention will be further described in detail below with reference to the drawings. The examples given are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0060] Taking the given event "The search volume for Xinjiang on the travel platform increased nearly threefold in one day" and the following document as an example to mine the opinions related to this event.

[0061] Scenario:

[0062] Event: The travel platform responded that the search volume for Xinjiang increased nearly threefold in one day

[0063] Document: (S1) Data provided by the online travel platform to reporters on the 25th showed that the search popularity of "Where to go in Xinjiang in April" increased by 275% on this platform in the past 24 hours. (S2) Xinjiang is vast and rich in tourism resources and has always been a popular choice for self-driving tours. (S3) The relevant person in charge said that with the continuous stability of the domestic epidemic prevention and control situation, the characteristics of Xinjiang being vast and sparsely populated can meet the higher requirements of tourists for hygiene quality and privacy, so it has become one of the most popular tourist destinations this year. (S4) In order to cope with the arrival of a large number of tourists, the product and service quality of scenic spots should be improved first. (S5) Secondly, there are still relatively few flights from Urumqi to Keketuohai at present, and it is hoped that the infrastructure construction in the fields of aviation, railways, and highways in Xinjiang can continue to be accelerated in the future to promote the rapid development of tourism in Xinjiang.

[0064] Implementation:

[0065] (1) Based on the view definition, extract all view segments related to events in the document. For example, in the given document, O1 = {S2}, O2 = {S3}, and O3 = {S4, S5} are three views of attitude, judgment, and suggestion types respectively.

[0066] (2) Based on the given event and the above-extracted view segments, match the corresponding view objects to the view segments. For example, for the above-given event and document, the event sub-structures included in the event "The travel platform responds that the search volume in Xinjiang has increased nearly threefold in a day" in addition to the event itself also include the sub-event "The search volume in Xinjiang has increased nearly threefold in a day" and the event participants "travel platform", "Xinjiang". Finally, three pairs of (view, event sub-structure) pairs can be extracted from the given event and document, {(O1, Xinjiang), (O2, The search volume in Xinjiang has increased nearly threefold in a day), (O3, The search volume in Xinjiang has increased nearly threefold in a day)}. Among them, O1 is the author's attitude towards "Xinjiang", O2 is the judgment of the "relevant person in charge" regarding "The search volume in Xinjiang has increased nearly threefold in a day", and O3 is the suggestion of the "relevant person in charge" regarding "The search volume in Xinjiang has increased nearly threefold in a day".

[0067] Although specific embodiments of the present invention are disclosed for illustrative purposes, which are intended to help understand the content of the present invention and implement it accordingly, those skilled in the art can understand that: without departing from the spirit and scope of the present invention and the appended claims, various substitutions, changes, and modifications are possible. Therefore, the present invention should not be limited to the content disclosed in the best embodiments, and the scope of protection required by the present invention shall be defined by the scope defined in the claims.

Claims

1. An event opinion mining method based on argument structure, the steps of which include: 1) For a given event e and document d, extract all opinion segments related to the given event e from the document d; 2) For each of the above-mentioned view fragments, extract the event sub-structure expressing the view from event e as the view object corresponding to the view fragment; wherein, the method for obtaining the view object of the view fragment is as follows: 21) Segment event e to obtain a word segmentation set {w1, w2, …, w m}; wherein, event e is a text phrase containing m words, and w m is the m-th word in event e; document d contains n sentences, and s n is the n-th sentence in document d; 22) Randomly combine consecutive word segments in the event phrase to obtain multiple sub-phrases; take each of the sub-phrases as a candidate view object; 23) Concatenate each candidate view object with one of the above-mentioned view fragments using the concatenation symbol of the Bidirectional Encoder Representations from Transformers (BERT) technique to construct the input: [CLS] candidate view object [SEP] view fragment [SEP]; 24) Input the concatenated text into the Bidirectional Encoder Representations from Transformers (BERT) technique to obtain the matching degree between the candidate view object and the corresponding view fragment, and select the candidate view object with the highest matching degree as the view object corresponding to the view fragment; 3) According to the processing result of step 2), obtain the event view set T of event e = {…, (o k , a k ), …|e, d}; where o k is the k-th view segment in document d; a k is the view o k corresponding view object.

2. The method according to claim 1, characterized in that, The opinion objects of the opinion segments include the event itself, event sub-events, and event participants.

3. The method according to claim 2, characterized in that, The event itself means that the opinion segment directly expresses an opinion on the entire event; the event sub-event means that the opinion segment expresses an opinion on a sub-event or related event of the event; the participant means that the opinion segment directly expresses an opinion on the entity involved in the event.

4. The method according to claim 1 or 2 or 3, characterized in that, The method for extracting all opinion segments related to the given event e from the document d is as follows: 1) Concatenate each sentence in the given event e and document d using the general concatenation symbol of BERT to construct the input: [CLS] event phrase [SEP] document sentence [SEP]; 2) Input the concatenated text into the Transformer encoder; 3) Input the representation corresponding to the start symbol [CLS] into the Softmax layer for classification. If the output result is 1, then the corresponding sentence is an opinion sentence related to event e; if the output result is 0, it is not an opinion sentence related to event e; 4) Connect the consecutive opinion sentences in the document d to obtain the opinion segment.

5. The method according to claim 1, characterized in that, Use the "Jieba" Chinese word segmentation component to segment the event e, obtaining the word segmentation set {w1, w2, …, w m}.

6. The method according to claim 1 or 2 or 3, characterized in that, The categories of the opinion segments include: judgment, attitude, belief, emotion, and suggestion.

7. A server, characterized in that, It includes a memory and a processor. The memory stores a computer program, and the computer program is configured to be executed by the processor. The computer program includes instructions for executing the steps in any one of the methods according to claims 1 to 6.

8. A computer-readable storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by the processor, the steps of any one of the methods according to claims 1 to 6 are implemented.

Images