A fine-grained opinion mining method for comprehensive and integrated discussion scenarios

Abstract

The present invention provides a fine-grained opinion mining method for a comprehensive integrated discussion scenario. Through a conditional random field model, real-time evaluation objects and real-time opinion words corresponding to the real-time evaluation objects in real-time discussion texts in a comprehensive integrated discussion scenario can be extracted; through a domain sentiment knowledge base, the sentiment tendency of real-time opinion words can be determined. In an embodiment of the present invention, fine-grained hierarchical opinion mining can be performed on real-time discussion texts in a comprehensive integrated discussion scenario, and then the opinions of participants on various aspects of the target field can be mined, so that participants can grasp the changes and trends of discussion opinions in a timely manner, and improve the efficiency of conference discussions.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to a fine-grained opinion mining method for comprehensive integrated discussion scenarios. Background Art

[0002] The essence of the comprehensive integration seminar system is to guide people to integrate relevant experience, theory, knowledge, information and data to the maximum extent when dealing with open complex systems through human-machine combination and group discussion, and to achieve the emergence of group wisdom and obtain a better understanding of complex systems through mutual stimulation among group members and collective processing of the above resources. This clearly elevates the individual wisdom in the comprehensive integration method to group wisdom, and significantly enhances the operability of the methodology of open complex systems.

[0003] In the scenario of comprehensive integrated discussion, participants can conduct qualitative discussion on a certain issue in text form online. However, the single discussion method makes the discussion inefficient and the opinions scattered, which is not conducive to grasping the changes in the discussion opinion trend. Since participants usually express their opinions on a certain aspect of the target field during the discussion, timely mining the evaluation objects and opinion tendencies of the discussion text will help grasp the discussion topic and predict the direction.

[0004] Traditional sentiment analysis technology is mainly carried out at the document and sentence level, and its main task is to identify whether the sentiment contained in the document and sentence is positive or negative. In the scenario of comprehensive integrated discussion, since participants often express their opinions on a certain aspect of the target field, a positive opinion on a certain aspect of the target field does not mean that he affirms all aspects of the target field, and vice versa. Therefore, it is necessary to mine opinions at the fine-grained level of the specific evaluation objects in the target field, and then mine the opinions of the participants on various aspects of the target field.

[0005] At present, there is no relevant research on fine-grained opinion mining in comprehensive and integrated discussion scenarios. Therefore, there is an urgent need to provide a fine-grained opinion mining method for comprehensive and integrated discussion scenarios. Summary of the invention

[0006] The present invention provides a fine-grained opinion mining method for comprehensive integrated discussion scenarios to address the defects in the prior art.

[0007] The present invention provides a fine-grained opinion mining method for a comprehensive integrated discussion scenario, comprising:

[0008] Obtain real-time discussion text in comprehensive and integrated discussion scenarios;

[0009] Based on the conditional random field model, extracting the real-time evaluation object in the real-time discussion text and the real-time opinion words corresponding to the real-time evaluation object;

[0010] Determining the emotional tendency of the real-time opinion word based on the real-time evaluation object, the real-time opinion word, and the domain emotion knowledge base;

[0011] Among them, the feature weights used in the conditional random field model and the domain sentiment knowledge base are determined based on the historical discussion text in the comprehensive integrated discussion scenario, and the historical discussion text is the discussion text in the comprehensive integrated discussion scenario obtained before the real-time discussion text; the domain sentiment knowledge base stores triplets of evaluation objects, opinion words and sentiment tendencies.

[0012] According to a fine-grained opinion mining method for a comprehensive integrated discussion scenario provided by the present invention, the domain sentiment knowledge base is constructed based on the following method:

[0013] Based on the conditional random field model, extract the historical evaluation objects and the historical viewpoint words corresponding to the historical evaluation objects in the historical discussion text, and construct an initial domain knowledge base based on the historical evaluation objects and the historical viewpoint words;

[0014] The initial domain knowledge base is expanded based on a general semantic dictionary to construct the domain sentiment knowledge base.

[0015] According to a fine-grained opinion mining method for comprehensive integrated discussion scenarios provided by the present invention, the initial domain knowledge base is expanded based on a universal semantic dictionary to construct the domain sentiment knowledge base, specifically including:

[0016] For any historical viewpoint word in the initial domain knowledge base, if the any historical viewpoint word exists in the general semantic dictionary, based on the sentiment tendency of the any historical viewpoint word marked in the general semantic dictionary, construct a triple of the any historical viewpoint word, the historical evaluation object corresponding to the any historical viewpoint word, and the sentiment tendency of the any historical viewpoint word;

[0017] If any of the historical viewpoint words does not exist in the universal semantic dictionary, then based on the co-occurrence frequency between the any of the historical viewpoint words and the target viewpoint words with known sentiment orientation during the search, the sentiment orientation of the any of the historical viewpoint words is determined, and a triple of the any of the historical viewpoint words, the historical evaluation object corresponding to the any of the historical viewpoint words, and the sentiment orientation of the any of the historical viewpoint words is constructed;

[0018] Searching for synonyms and antonyms of any of the historical viewpoint words from the general semantic dictionary, and based on the triples, determining a triple of the synonym, the historical evaluation object corresponding to any of the historical viewpoint words, and the sentiment tendency of the synonym, and determining a triple of the antonym, the historical evaluation object corresponding to any of the historical viewpoint words, and the sentiment tendency of the antonym;

[0019] Based on each triple, a sentiment knowledge base of the domain is constructed.

[0020] According to a fine-grained opinion mining method for a comprehensive integrated discussion scenario provided by the present invention, the historical evaluation objects in the historical discussion text and the historical opinion words corresponding to the historical evaluation objects are extracted based on a conditional random field model, and the method also includes:

[0021] Performing preprocessing operations on the historical research text;

[0022] The preprocessing operation includes a word segmentation operation, a stop word removal operation and a useless symbol removal operation.

[0023] According to a fine-grained opinion mining method for a comprehensive integrated discussion scenario provided by the present invention, the emotional tendency of the real-time opinion word is determined based on the real-time evaluation object, the real-time opinion word and the domain emotional knowledge base, specifically including:

[0024] If it is determined that both the real-time evaluation object and the real-time opinion word exist in the domain emotion knowledge base, then based on the triples corresponding to the real-time evaluation object and the real-time opinion word stored in the domain emotion knowledge base, the sentiment tendency of the real-time opinion word is determined;

[0025] If it is determined that the real-time evaluation object is not in the domain emotion knowledge base and the real-time opinion word exists in the domain emotion knowledge base, then based on the opinion word set associated with the real-time opinion word stored in the domain emotion knowledge base, determining the emotional tendency of the real-time opinion word;

[0026] If it is determined that the real-time opinion word is not in the domain emotion knowledge base and the real-time evaluation object exists in the domain emotion knowledge base, then based on the opinion word set associated with the real-time evaluation object stored in the domain emotion knowledge base, determining the emotional tendency of the real-time opinion word;

[0027] If it is determined that the real-time evaluation object and the real-time opinion word are not in the domain emotion knowledge base, the emotion tendency of the real-time opinion word is determined based on the co-occurrence frequency between the real-time opinion word and the target opinion word with known emotion tendency during the search.

[0028] According to a fine-grained opinion mining method for a comprehensive integrated discussion scenario provided by the present invention, the real-time evaluation object in the real-time discussion text and the real-time opinion words corresponding to the real-time evaluation object are extracted based on a conditional random field model, specifically including:

[0029] Based on the conditional random field model, the real-time evaluation object and the real-time opinion in the real-time discussion text are simultaneously extracted through pre-defined word features, part-of-speech features and syntactic features.

[0030] According to a fine-grained opinion mining method for comprehensive integrated discussion scenarios provided by the present invention, the feature weights used in the conditional random field model are estimated from the historical discussion text based on maximum likelihood estimation and a penalty function.

[0031] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the fine-grained opinion mining method for comprehensive integrated discussion scenarios as described in any one of the above are implemented.

[0032] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the fine-grained opinion mining method for comprehensive integrated discussion scenarios as described in any one of the above are implemented.

[0033] The present invention also provides a computer program product, including a computer program, characterized in that when the computer program is executed by a processor, the computer program implements the steps of any of the above-mentioned fine-grained opinion mining methods for comprehensive integrated discussion scenarios.

[0034] The fine-grained opinion mining method for comprehensive integrated discussion scenarios provided by the present invention can extract the real-time evaluation objects and the real-time opinion words corresponding to the real-time evaluation objects in the real-time discussion text in the comprehensive integrated discussion scenario through the conditional random field model; the emotional tendency of the real-time opinion words can be determined through the domain sentiment knowledge base. In the embodiment of the present invention, the real-time discussion text in the comprehensive integrated discussion scenario can be mined at a fine-grained level, and then the opinions of the participants on various aspects of the target field can be mined, so that the participants can grasp the changes and trends of the discussion opinions in a timely manner, and improve the efficiency of the conference discussion. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 It is a flow chart of a fine-grained opinion mining method for comprehensive integrated discussion scenarios provided by the present invention;

[0037] Figure 2 It is a structural schematic diagram of a fine-grained opinion mining system for comprehensive integrated discussion scenarios provided by the present invention;

[0038] Figure 3 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] At present, traditional sentiment analysis technology is mainly carried out at the document and sentence level, and its main task is to identify whether the sentiment contained in the document and sentence is positive or negative. In the comprehensive integrated discussion scenario, since participants often express their opinions on a certain aspect of the field (i.e., the evaluation object), a positive opinion on a certain evaluation object does not mean that he affirms all aspects of the field, and vice versa. Therefore, it is necessary to mine opinions at the fine-grained level of specific evaluation objects to mine the opinions of participants on each evaluation object.

[0041] At present, there is no research on fine-grained opinion mining for comprehensive integrated discussion scenarios, but there are some research works on fine-grained opinion mining for Chinese short texts such as product reviews and Weibo, mainly in the following categories:

[0042] Method 1: For online comment data, by constructing a set of extraction rules and pruning filtering strategies corresponding to explicit feature opinions, explicit features and their corresponding opinion words are directly obtained.

[0043] Method 2: For Weibo data, we use dependency relationships to extract several grammatical rules, and then combine them with the sentiment dictionary. Focusing on sentiment words and evaluation aspects, we perform joint extraction of evaluation aspects and evaluation words, and finally perform fuzzy inductive matching to form a clearer viewpoint.

[0044] Method 3: For online review data, based on the improvement of the traditional text topic model (Latent Dirichlet Allocation, LDA), the domain layer is introduced, and a domain-discriminant LDA is proposed to mine opinions on online reviews.

[0045] Method 4: For online comment data, a deep neural network with a sequence mechanism combined with an attention mechanism is used to extract and mine opinions, and analyze emotional attitudes.

[0046] Among the above methods, rule-based methods are more suitable for semi-structured data, but not for unstructured data in comprehensive integrated discussion scenarios; topic models are more suitable for processing long texts, while comprehensive integrated discussion scenarios are mostly short texts, resulting in sparse data that is not suitable for topic models; deep learning-based methods require the support of large sample data, while discussion texts are usually real-time small sample data. Therefore, the above methods cannot meet the actual needs of comprehensive integrated discussion scenarios, and it is necessary to study effective machine learning methods that are suitable for comprehensive integrated discussion scenarios.

[0047] In addition, during the discussion, participants are under great time pressure and have no time to search and record relevant speeches, or the frequent search affects the progress of the discussion, hindering the generation and improvement of knowledge during the discussion. Therefore, it is urgent to conduct fine-grained opinion mining on the discussion text to assist the discussion process to proceed efficiently.

[0048] On this basis, an embodiment of the present invention provides a fine-grained opinion mining method for a comprehensive integrated discussion scenario, including:

[0049] S1, obtain the real-time discussion text in the comprehensive integrated discussion scenario;

[0050] S2, extracting the real-time evaluation objects in the real-time discussion text and the real-time opinion words corresponding to the real-time evaluation objects based on the conditional random field model;

[0051] S3, determining the emotional tendency of the real-time opinion word based on the real-time evaluation object, the real-time opinion word and the domain emotion knowledge base;

[0052] Among them, the feature weights used in the conditional random field model and the domain sentiment knowledge base are determined based on the historical discussion text in the comprehensive integrated discussion scenario, and the historical discussion text is the discussion text in the comprehensive integrated discussion scenario obtained before the real-time discussion text; the domain sentiment knowledge base stores triplets of evaluation objects, opinion words and sentiment tendencies.

[0053] Specifically, the fine-grained opinion mining method for comprehensive integrated discussion scenarios provided in the embodiments of the present invention is executed by a server, which can be a local server or a cloud server. The local server can specifically be a computer, etc., which is not specifically limited in the embodiments of the present invention.

[0054] First, execute step S1 to obtain the real-time discussion text in the comprehensive integrated discussion scenario. The comprehensive integrated discussion scenario refers to the discussion scenario in the large-scale discussion hall system. In this comprehensive integrated discussion scenario, through the combination of man and machine and group discussion, the relevant experience, theory, knowledge, information and data are integrated to the maximum extent, and through the mutual stimulation between group members and the collective processing of the above resources, the emergence of group wisdom is realized, and a better understanding of complex systems is obtained. The real-time discussion text refers to the discussion text obtained in real time. The real-time discussion text can be the discussion content of the participants, which can be a sentence or a sentence. The discussion content can include the participants' views on a certain aspect of the field of comprehensive integrated discussion.

[0055] Then, step S2 is performed to extract the real-time evaluation object and the real-time opinion words corresponding to the real-time evaluation object in the real-time discussion text through the conditional random field model. The conditional random field model is used to extract the real-time evaluation object and the real-time opinion words corresponding to the real-time evaluation object in the real-time discussion text based on the conditional random field (CRF). When the conditional random field model realizes the extraction function, the feature weights used can be determined by the historical discussion text in the comprehensive integrated discussion scenario. The historical discussion text refers to the discussion text obtained before the real-time discussion text in the neutralization and integrated discussion scenario. The process of determining the feature weight can be regarded as the process of training the conditional random field model through the historical discussion text, and the result of the training is to determine the feature weight. The real-time evaluation object refers to the specific aspects of the field of comprehensive integrated discussion involved in the real-time discussion text, and the real-time opinion words refer to the expression of opinions on the real-time evaluation object in the real-time discussion text. The number of real-time evaluation objects in the real-time discussion text can be one or more, and the number of real-time opinion words corresponding to each real-time evaluation object can also be one or more, which is not specifically limited in the embodiment of the present invention.

[0056] Finally, step S3 is executed to determine the emotional tendency of the real-time opinion words according to the real-time evaluation object, the real-time opinion words and the domain emotion knowledge base. Among them, the domain emotion knowledge base stores triplets of evaluation objects, opinion words and emotional tendencies, and the triplets are used to characterize the correspondence between the evaluation objects, opinion words and emotional tendencies. Emotional tendency refers to emotional polarity, that is, whether the opinion words are positive or negative when used to evaluate the corresponding evaluation objects. Through the real-time evaluation objects and real-time opinion words, combined with the triplets stored in the domain emotion knowledge base, the emotional tendency of the real-time opinion words when used to evaluate the corresponding real-time evaluation objects can be determined. The triplets in the domain emotion knowledge base can be determined by the historical discussion texts in the comprehensive integration discussion scenario.

[0057] The fine-grained opinion mining method for comprehensive integrated discussion scenarios provided in the embodiments of the present invention can extract the real-time evaluation objects and the real-time opinion words corresponding to the real-time evaluation objects in the real-time discussion text in the comprehensive integrated discussion scenario through the conditional random field model; the emotional tendency of the real-time opinion words can be determined through the domain sentiment knowledge base. In the embodiments of the present invention, the real-time discussion text in the comprehensive integrated discussion scenario can be mined at a fine-grained level, and then the opinions of the participants on various aspects of the target field can be mined, so that the participants can grasp the changes and trends of the discussion opinions in a timely manner, and improve the efficiency of the meeting discussion.

[0058] On the basis of the above-mentioned embodiments, in the fine-grained opinion mining method for comprehensive integrated discussion scenarios provided in the embodiments of the present invention, the domain sentiment knowledge base is constructed based on the following method:

[0059] Based on the conditional random field model, extract the historical evaluation objects and the historical viewpoint words corresponding to the historical evaluation objects in the historical discussion text, and construct an initial domain knowledge base based on the historical evaluation objects and the historical viewpoint words;

[0060] The initial domain knowledge base is expanded based on a general semantic dictionary to construct the domain sentiment knowledge base.

[0061] Specifically, in the embodiment of the present invention, when determining the domain sentiment knowledge base, firstly, the historical evaluation objects in the historical discussion text and the historical opinion words corresponding to the historical evaluation objects are extracted according to the conditional random field model. The extraction method is consistent with the extraction method of the real-time evaluation objects and the real-time opinion words corresponding to the real-time evaluation objects in the real-time discussion text, and will not be repeated here in the embodiment of the present invention.

[0062] Then, an initial domain knowledge base is constructed based on the historical evaluation objects and historical opinion words, that is, a pair of historical evaluation objects and historical opinion words is generated, and the pair is stored in the initial domain knowledge base. Before storing the pair in the initial domain knowledge base, noise can also be filtered out based on the word frequency of the historical opinion words corresponding to each historical evaluation object, that is, the historical opinion words with a word frequency less than or equal to a preset word frequency threshold are deleted, and then the pair obtained after filtering out the noise is stored in the initial domain knowledge base. Among them, the preset word frequency threshold can be set as needed, for example, it can be set to 0.36.

[0063] Finally, the initial domain knowledge base is expanded according to the universal semantic dictionary (HOWNET) to construct a domain sentiment knowledge base. The expansion operation can include two aspects: on the one hand, the sentiment tendency of historical opinion words in evaluating the corresponding historical evaluation objects is introduced into the initial domain knowledge base, so that the bigrams become triplets; on the other hand, the synonyms and antonyms of historical opinion words are determined, and then the triplets are expanded by combining the synonyms and antonyms to finally obtain the domain sentiment knowledge base.

[0064] In an embodiment of the present invention, historical evaluation objects and historical opinion words corresponding to the historical evaluation objects in the historical discussion text are extracted based on a conditional random field model, and an initial domain knowledge base is constructed based on the historical evaluation objects and the historical opinion words; and then the initial domain knowledge base is expanded in combination with a general semantic dictionary to construct a domain emotion knowledge base, which can enrich the domain emotion knowledge base and facilitate the judgment of the emotional tendency of the real-time opinion words in the real-time discussion text when evaluating the real-time evaluation objects.

[0065] On the basis of the above-mentioned embodiment, the fine-grained opinion mining method for comprehensive integrated discussion scenarios provided in the embodiment of the present invention, the initial domain knowledge base is expanded based on the universal semantic dictionary to construct the domain sentiment knowledge base, specifically including:

[0066] For any historical viewpoint word in the initial domain knowledge base, if the any historical viewpoint word exists in the general semantic dictionary, based on the sentiment tendency of the any historical viewpoint word marked in the general semantic dictionary, construct a triple of the any historical viewpoint word, the historical evaluation object corresponding to the any historical viewpoint word, and the sentiment tendency of the any historical viewpoint word;

[0067] If any of the historical viewpoint words does not exist in the universal semantic dictionary, then based on the co-occurrence frequency between the any of the historical viewpoint words and the target viewpoint words with known sentiment orientation during the search, the sentiment orientation of the any of the historical viewpoint words is determined, and a triple of the any of the historical viewpoint words, the historical evaluation object corresponding to the any of the historical viewpoint words, and the sentiment orientation of the any of the historical viewpoint words is constructed;

[0068] Searching for synonyms and antonyms of any of the historical viewpoint words from the general semantic dictionary, and based on the triples, determining a triple of the synonym, the historical evaluation object corresponding to any of the historical viewpoint words, and the sentiment tendency of the synonym, and determining a triple of the antonym, the historical evaluation object corresponding to any of the historical viewpoint words, and the sentiment tendency of the antonym;

[0069] Based on each triple, a sentiment knowledge base of the domain is constructed.

[0070] Specifically, in an embodiment of the present invention, when the initial domain knowledge base is expanded through a universal semantic dictionary, it is possible to determine whether any historical opinion word O1 in the initial domain knowledge base exists in the universal semantic dictionary. If the historical opinion word O1 exists in the universal semantic dictionary, then based on the sentiment tendency of the historical opinion word O1 marked in the universal semantic dictionary, a triple of the historical opinion word O1, the historical evaluation object F1 corresponding to the historical opinion word O1, and the sentiment tendency of the historical opinion word O1 is constructed.

[0071] If the historical opinion word O1 does not exist in the general semantic dictionary, the sentiment tendency of the historical opinion word O1 is determined based on the co-occurrence frequency between the historical opinion word O1 and the target opinion word O2 with known sentiment tendency during the search. The co-occurrence frequency is the probability of co-occurrence. In the embodiment of the present invention, a point-wise mutual information (PMI) algorithm based on the co-occurrence frequency can be used to determine the sentiment tendency of the historical opinion word O1. The PMI value between the historical opinion word O1 and each target opinion word O2 can be determined by the following formula:

[0072]

[0073] Among them, p(O1, O2) is the co-occurrence frequency of O1 and O2 when O1 and O2 are searched in the search engine at the same time, p(O1) is the frequency of O1 appearing when O1 is searched in the search engine, and p(O2) is the frequency of O2 appearing when O2 is searched in the search engine.

[0074] Then, the target opinion word corresponding to the maximum PMI value is selected, and the sentiment tendency corresponding to the target opinion word is used as the sentiment tendency of the historical opinion word O1. A triple is constructed between the historical opinion word O1, the historical evaluation object F1 corresponding to the historical opinion word O1, and the sentiment tendency of the historical opinion word O1.

[0075] The synonyms and antonyms of the historical viewpoint word O1 are searched from the general semantic dictionary, and based on the triples, the triples of the synonyms, the historical evaluation object F1, and the sentiment tendency of the synonyms are determined, and the triples of the antonyms, the historical evaluation object F1, and the sentiment tendency of the antonyms are determined. Among them, the sentiment tendency of the synonyms can be the same as the sentiment tendency of the historical viewpoint word O1, and the sentiment tendency of the antonyms can be opposite to the sentiment tendency of the historical viewpoint word O1. Finally, based on each triple, the domain sentiment knowledge base can be constructed.

[0076] In an embodiment of the present invention, by expanding the initial domain knowledge base through a general semantic dictionary, not only can the sentiment tendency of the historical opinion words be determined, but also the triples can be expanded through the synonyms and antonyms of the historical opinion words, thereby constructing a domain sentiment knowledge base, which can enrich the content of the domain sentiment knowledge base.

[0077] On the basis of the above-mentioned embodiment, the fine-grained opinion mining method for comprehensive integrated discussion scenarios provided in the embodiment of the present invention, based on the conditional random field model, extracts the historical evaluation objects in the historical discussion text and the historical opinion words corresponding to the historical evaluation objects, and also includes:

[0078] Performing preprocessing operations on the historical research text;

[0079] The preprocessing operation includes a word segmentation operation, a stop word removal operation and a useless symbol removal operation.

[0080] Specifically, in the embodiment of the present invention, before extracting historical evaluation objects and historical viewpoint words from the historical research text, the historical research text can be preprocessed, and the preprocessing operation can include word segmentation, stop word removal, and useless symbol removal on the historical research text, and then the preprocessed historical research text is extracted, which can ensure extraction efficiency and extraction quality. The preprocessing operation can be implemented using the open source Chinese word segmentation software ICTCLAS, which is not specifically limited in the embodiment of the present invention.

[0081] On the basis of the above-mentioned embodiment, the fine-grained opinion mining method for comprehensive integrated discussion scenarios provided in the embodiment of the present invention, based on the real-time evaluation object, the real-time opinion word and the domain sentiment knowledge base, determines the sentiment tendency of the real-time opinion word, specifically including:

[0082] If it is determined that both the real-time evaluation object and the real-time opinion word exist in the domain emotion knowledge base, then based on the triples corresponding to the real-time evaluation object and the real-time opinion word stored in the domain emotion knowledge base, the sentiment tendency of the real-time opinion word is determined;

[0083] If it is determined that the real-time evaluation object is not in the domain emotion knowledge base and the real-time opinion word exists in the domain emotion knowledge base, then based on the opinion word set associated with the real-time opinion word stored in the domain emotion knowledge base, determining the emotional tendency of the real-time opinion word;

[0084] If it is determined that the real-time opinion word is not in the domain emotion knowledge base and the real-time evaluation object exists in the domain emotion knowledge base, then based on the opinion word set associated with the real-time evaluation object stored in the domain emotion knowledge base, determining the emotional tendency of the real-time opinion word;

[0085] If it is determined that the real-time evaluation object and the real-time opinion word are not in the domain emotion knowledge base, the emotion tendency of the real-time opinion word is determined based on the co-occurrence frequency between the real-time opinion word and the target opinion word with known emotion tendency during the search.

[0086] Specifically, in the embodiment of the present invention, when determining the emotional tendency of the real-time opinion word, it is first determined whether the real-time evaluation object F0 and the real-time opinion word O0 both exist in the domain emotion knowledge base. If both exist in the domain emotion knowledge base, the emotional tendency of the real-time opinion word O0 can be determined based on the triple corresponding to the real-time evaluation object F0 and the real-time opinion word O0 stored in the domain emotion knowledge base. At this time, the emotional tendency in the triple is the emotional tendency of the real-time opinion word O0.

[0087] If the real-time evaluation object F0 and the real-time opinion word O0 both exist in the domain sentiment knowledge base, but there is no triple corresponding to the real-time evaluation object F0 and the real-time opinion word O0 in the domain sentiment knowledge base, then the opinion word set O-set in the triple of the real-time evaluation object F0 in the domain sentiment knowledge base can be determined first, and then the sentiment tendency of the real-time opinion word O0 can be determined based on the sentiment tendency relationship between each opinion word in the opinion word set O-set and the real-time opinion word O0.

[0088] If the real-time evaluation object F0 is not in the domain sentiment knowledge base, and the real-time opinion word O0 is in the domain sentiment knowledge base, the sentiment tendency of the real-time opinion word O0 can be determined based on the opinion word set O-set1 associated with the real-time opinion word O0 stored in the domain sentiment knowledge base. The opinion word set O-set1 refers to a set of opinion words associated with the real-time opinion word O0, and the opinion words associated with the real-time opinion word O0 may include synonyms and antonyms of the real-time opinion word O0. Then, based on the sentiment tendency of the opinion words associated with the real-time opinion word O0, the sentiment tendency of the real-time opinion word O0 can be determined.

[0089] If the real-time opinion word O0 is not in the domain sentiment knowledge base and the real-time evaluation object F0 is in the domain sentiment knowledge base, the sentiment tendency of the real-time opinion word O0 can be determined based on the opinion word set O-set2 associated with the real-time evaluation object F0 stored in the domain sentiment knowledge base. The opinion word set O-set2 refers to a set of opinion words associated with the real-time evaluation object F0. The opinion words associated with the real-time evaluation object F0 can be opinion words in the triple that includes the real-time evaluation object F0. Then, based on the sentiment tendency of the opinion words associated with the real-time evaluation object F0, the sentiment tendency of the real-time opinion word O0 can be determined.

[0090] If the real-time opinion word F0 and the real-time opinion word O0 are not in the domain sentiment knowledge base, the sentiment tendency of the real-time opinion word O0 can be determined according to the co-occurrence frequency between the real-time opinion word F0 and the target opinion word O2 with known sentiment tendency during the search, combined with the PMI algorithm based on co-occurrence frequency.

[0091] In the embodiment of the present invention, the emotional tendency of the real-time opinion words is determined by using a domain emotional knowledge base, so that the determination result can be more accurate and reliable.

[0092] On the basis of the above-mentioned embodiment, the fine-grained opinion mining method for comprehensive integrated discussion scenarios provided in the embodiment of the present invention, based on the conditional random field model, extracts the real-time evaluation objects in the real-time discussion text and the real-time opinion words corresponding to the real-time evaluation objects, specifically including:

[0093] Based on the conditional random field model, the real-time evaluation object and the real-time opinion in the real-time discussion text are simultaneously extracted through pre-defined word features, part-of-speech features and syntactic features.

[0094] Specifically, in an embodiment of the present invention, when extracting real-time evaluation objects and real-time opinion words corresponding to the real-time evaluation objects in a real-time discussion text, a conditional random field model can be used to achieve this through predefined word features, part-of-speech features, and syntactic features.

[0095] For the current word in the real-time discussion text, the [-2,2] distance window of the current word can be used as the word feature of the current word. The unit of the distance window is word, that is, the distance window includes 2 words on the left and 2 words on the right of the current word, plus the current word, a total of 5 words. The current word can be any word in the real-time discussion text.

[0096] Since the part of speech of the evaluation object is often a noun or a noun phrase, and the opinion word is often an adjective or an adverb, the part of speech of the current word in the [-2,2] distance window is used as the part of speech feature of the current word. It is defined as a syntactic feature whether the current word belongs to a noun phrase (i.e., noun + noun), a verb phrase (i.e., verb + noun), or an adjective phrase (i.e., adjective + noun + adjective).

[0097] Then the real-time discussion text is annotated by word features, part-of-speech features and syntactic features to obtain the annotated discussion text.

[0098] Assume that X and Y represent the joint distribution random variables of the observation sequence to be labeled and its corresponding label sequence, respectively. The conditional random field (X, Y) is an undirected graph model with the observation sequence X as the global condition. Define G = (V, E) as an undirected graph, where E represents the corresponding relationship between X and Y, and Y = {Y v |v∈V}, that is, each node v in V corresponds to a random variable Y v The components of the labeled sequence are as follows. If each random variable Y v For G to obey the Markov property, then (X,Y) constitutes a conditional random field.

[0099] In the comprehensive integrated discussion scenario, the real-time discussion text is used as the observation sequence X = {x1,…,x n}, where x i represents the i-th word in the real-time discussion text. The goal is to find a label sequence Y={y1,…,y n The tags Fe, Op, and Bg represent the evaluation object, opinion word, and background word, respectively, and the problem of extracting the evaluation object and opinion word is transformed into a sequence labeling problem.

[0100] For a given observation sequence X, a statistical model of the conditional probability p(Y|X) of the labeled sequence Y can be established. Solving the sequence labeling task is to find Y so that P(Y|X) is maximized.

[0101] For a random variable Y{u|u≠v,{u,v∈V}},Y v The conditional probability is:

[0102] p(Y v |X,Y u , u≠v, {u, v∈V})=p(Y v |X,Y u , (u, v)∈E)

[0103] The probability of defining a labeled sequence Y given an observation sequence X is the normalized product of potential functions, where each factor of the potential function has the following form:

[0104]

[0105] Wherein, j represents the identifier of the transfer characteristic function, i.e., the jth transfer characteristic function, and its value is (1~|E|); k represents the identifier of the state characteristic function, i.e., the kth state characteristic function, and its value is (1~|V|). j (Y i-1 , Y i , X,i) is the transition feature function for the observation sequence X and the position i and i-1 mark, s k (Y i , X,i) is the state characteristic function of the mark and observation sequence X at position i, and the parameter λ j and μ k Represents the weight of the feature function.

[0106] Order k (Y i , X, i) = s k (Y i-1 , Y i , X, i), then where f j (Y i-1 , Y i , X, i) represents a transfer characteristic function t j (Y i-1 , Y i , X, i) or a state characteristic function s k (Y i , X, i).

[0107] In the embodiment of the present invention, predefined word features, part-of-speech features and syntactic features are introduced to simultaneously extract real-time evaluation objects and real-time opinions in the real-time discussion text, which can make the extraction result more accurate.

[0108] On the basis of the above-mentioned embodiments, in the embodiments of the present invention, a fine-grained opinion mining method for comprehensive integrated discussion scenarios is provided, wherein the feature weights used in the conditional random field model are estimated from the historical discussion text based on maximum likelihood estimation and a penalty function.

[0109] Specifically, when determining the feature weights used in the conditional random field model, each historical discussion text can be taken as a given observation sequence, then X = X1, X2, ..., X n , and its corresponding tag sequence Y = Y1, Y2, ..., Y n The conditional probability is:

[0110]

[0111] Where n is the number of historical research texts, Z(X) is the normalization factor, and:

[0112]

[0113] In the embodiment of the present invention, the feature weight λ of the conditional random field can be estimated from the observation sequence composed of each historical discussion text using maximum likelihood estimation.

[0114] The observation sequence composed of each historical research text is used as the training set T = { <X k , Y k >}, k ranges from 1 to n. Maximum likelihood estimation assumes that P(Y|X, λ) is a function of λ, and the λ that maximizes P(Y|X, λ) is the estimated value.

[0115]

[0116] Since the likelihood function L Δ It is a convex function, and the point where the derivative is equal to zero is the maximum point. Taking the derivative of λ, the partial derivative formula is as follows:

[0117]

[0118] in, is λ j Feature expectations in conditional random field model distributions. It can be solved using the forward-backward algorithm in dynamic programming.

[0119] Considering that if the maximum likelihood estimation is used directly, the over-learning problem may occur, a penalty function is introduced in the embodiment of the present invention to solve the over-learning problem that may occur. Then the original problem becomes:

[0120]

[0121] Among them, const is a constant.

[0122] The derivative becomes:

[0123]

[0124] The estimation of λ may use an iterative scaling method, such as GIS, IIS, L-BFGS algorithm, etc., to optimize the objective function, which is not specifically limited in the embodiment of the present invention.

[0125] In summary, the fine-grained opinion mining method for comprehensive integrated discussion scenarios provided in the embodiments of the present invention can provide participants with real-time observation of the comment objects of the discussion, as well as the polarity and direction of opinions when discussing specific issues in a comprehensive integrated discussion room environment, thereby providing participants with good interaction conditions and greatly improving the efficiency of meeting discussions and decision-making.

[0126] like Figure 2 As shown, based on the above embodiments, an embodiment of the present invention provides a fine-grained opinion mining system for comprehensive integrated discussion scenarios, including:

[0127] An acquisition module 21 is used to acquire real-time discussion text in a comprehensive integrated discussion scenario;

[0128] An extraction module 22, for extracting the real-time evaluation object and the real-time opinion words corresponding to the real-time evaluation object in the real-time discussion text based on a conditional random field model;

[0129] A determination module 23, configured to determine the emotional tendency of the real-time opinion word based on the real-time evaluation object, the real-time opinion word and the domain emotion knowledge base;

[0130] Among them, the feature weights used in the conditional random field model and the domain sentiment knowledge base are determined based on the historical discussion text in the comprehensive integrated discussion scenario, and the historical discussion text is the discussion text in the comprehensive integrated discussion scenario obtained before the real-time discussion text; the domain sentiment knowledge base stores triplets of evaluation objects, opinion words and sentiment tendencies.

[0131] On the basis of the above-mentioned embodiments, the fine-grained opinion mining system for comprehensive integrated discussion scenarios provided in the embodiments of the present invention further includes a knowledge base construction module for:

[0132] Based on the conditional random field model, extract the historical evaluation objects and the historical viewpoint words corresponding to the historical evaluation objects in the historical discussion text, and construct an initial domain knowledge base based on the historical evaluation objects and the historical viewpoint words;

[0133] The initial domain knowledge base is expanded based on a general semantic dictionary to construct the domain sentiment knowledge base.

[0134] On the basis of the above-mentioned embodiments, in the fine-grained opinion mining system for comprehensive integrated discussion scenarios provided in the embodiments of the present invention, the knowledge base construction module is specifically used for:

[0135] For any historical viewpoint word in the initial domain knowledge base, if the any historical viewpoint word exists in the general semantic dictionary, based on the sentiment tendency of the any historical viewpoint word marked in the general semantic dictionary, construct a triple of the any historical viewpoint word, the historical evaluation object corresponding to the any historical viewpoint word, and the sentiment tendency of the any historical viewpoint word;

[0136] If any of the historical viewpoint words does not exist in the universal semantic dictionary, then based on the co-occurrence frequency between the any of the historical viewpoint words and the target viewpoint words with known sentiment orientation during the search, the sentiment orientation of the any of the historical viewpoint words is determined, and a triple of the any of the historical viewpoint words, the historical evaluation object corresponding to the any of the historical viewpoint words, and the sentiment orientation of the any of the historical viewpoint words is constructed;

[0137] Searching for synonyms and antonyms of any of the historical viewpoint words from the general semantic dictionary, and based on the triples, determining a triple of the synonym, the historical evaluation object corresponding to any of the historical viewpoint words, and the sentiment tendency of the synonym, and determining a triple of the antonym, the historical evaluation object corresponding to any of the historical viewpoint words, and the sentiment tendency of the antonym;

[0138] Based on each triple, a sentiment knowledge base of the domain is constructed.

[0139] On the basis of the above-mentioned embodiment, the fine-grained opinion mining system for comprehensive integrated discussion scenarios provided in the embodiment of the present invention further includes a preprocessing module for:

[0140] Performing preprocessing operations on the historical research text;

[0141] The preprocessing operation includes a word segmentation operation, a stop word removal operation and a useless symbol removal operation.

[0142] On the basis of the above-mentioned embodiments, in the fine-grained opinion mining system for comprehensive integrated discussion scenarios provided in the embodiments of the present invention, the determination module is specifically used for:

[0143] If it is determined that both the real-time evaluation object and the real-time opinion word exist in the domain emotion knowledge base, then based on the triples corresponding to the real-time evaluation object and the real-time opinion word stored in the domain emotion knowledge base, the sentiment tendency of the real-time opinion word is determined;

[0144] If it is determined that the real-time evaluation object is not in the domain emotion knowledge base and the real-time opinion word exists in the domain emotion knowledge base, then based on the opinion word set associated with the real-time opinion word stored in the domain emotion knowledge base, determining the emotional tendency of the real-time opinion word;

[0145] If it is determined that the real-time opinion word is not in the domain emotion knowledge base and the real-time evaluation object exists in the domain emotion knowledge base, then based on the opinion word set associated with the real-time evaluation object stored in the domain emotion knowledge base, determining the emotional tendency of the real-time opinion word;

[0146] If it is determined that the real-time evaluation object and the real-time opinion word are not in the domain emotion knowledge base, the emotion tendency of the real-time opinion word is determined based on the co-occurrence frequency between the real-time opinion word and the target opinion word with known emotion tendency during the search.

[0147] On the basis of the above-mentioned embodiments, in the fine-grained opinion mining system for comprehensive integrated discussion scenarios provided in the embodiments of the present invention, the extraction module is specifically used for:

[0148] Based on the conditional random field model, the real-time evaluation object and the real-time opinion in the real-time discussion text are simultaneously extracted through pre-defined word features, part-of-speech features and syntactic features.

[0149] On the basis of the above-mentioned embodiments, in the fine-grained opinion mining system for comprehensive integrated discussion scenarios provided in the embodiments of the present invention, the feature weights used in the conditional random field model are estimated from the historical discussion text based on maximum likelihood estimation and penalty function.

[0150] Specifically, the functions of each module in the fine-grained opinion mining system for comprehensive integrated discussion scenarios provided in the embodiments of the present invention correspond one-to-one to the operational procedures of each step in the above-mentioned method embodiments, and the effects achieved are also consistent. Please refer to the above-mentioned embodiments for details, and no further details will be given in the embodiments of the present invention.

[0151] Figure 3 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 3As shown, the electronic device may include: a processor (processor) 310, a communication interface (Communications Interface) 320, a memory (memory) 330 and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340. The processor 310 can call the logic instructions in the memory 330 to execute the fine-grained opinion mining method for the comprehensive integrated discussion scenario provided by the above embodiments, the method comprising: obtaining a real-time discussion text in a comprehensive integrated discussion scenario; based on a conditional random field model, extracting real-time evaluation objects in the real-time discussion text and real-time opinion words corresponding to the real-time evaluation objects; based on the real-time evaluation objects, the real-time opinion words and the domain sentiment knowledge base, determining the emotional tendency of the real-time opinion words; wherein the feature weights used in the conditional random field model and the domain sentiment knowledge base are determined based on the historical discussion text in the comprehensive integrated discussion scenario, and the historical discussion text is the discussion text in the comprehensive integrated discussion scenario obtained before the real-time discussion text; the domain sentiment knowledge base stores triplets of evaluation objects, opinion words and emotional tendencies.

[0152] In addition, the logic instructions in the above-mentioned memory 330 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0153] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the fine-grained opinion mining method for the comprehensive integrated discussion scenario provided by the above-mentioned embodiments, the method including: obtaining a real-time discussion text in a comprehensive integrated discussion scenario; based on a conditional random field model, extracting real-time evaluation objects and real-time opinion words corresponding to the real-time evaluation objects in the real-time discussion text; based on the real-time evaluation objects, the real-time opinion words and a domain sentiment knowledge base, determining the emotional tendency of the real-time opinion words; wherein the feature weights used in the conditional random field model and the domain sentiment knowledge base are both determined based on historical discussion texts in the comprehensive integrated discussion scenario, and the historical discussion texts are discussion texts in the comprehensive integrated discussion scenario obtained before the real-time discussion texts; the domain sentiment knowledge base stores triples of evaluation objects, opinion words and emotional tendencies.

[0154] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the fine-grained opinion mining method for the comprehensive integrated discussion scenario provided in the above-mentioned embodiments, the method comprising: obtaining a real-time discussion text in a comprehensive integrated discussion scenario; extracting real-time evaluation objects and real-time opinion words corresponding to the real-time evaluation objects in the real-time discussion text based on a conditional random field model; determining the emotional tendency of the real-time opinion words based on the real-time evaluation objects, the real-time opinion words and a domain sentiment knowledge base; wherein the feature weights used in the conditional random field model and the domain sentiment knowledge base are both determined based on historical discussion texts in the comprehensive integrated discussion scenario, and the historical discussion texts are discussion texts in the comprehensive integrated discussion scenario obtained before the real-time discussion texts; the domain sentiment knowledge base stores triplets of evaluation objects, opinion words and emotional tendencies.

[0155] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0156] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fine-grained opinion mining method for comprehensive integrated discussion scenarios, characterized in that: include: Obtain real-time discussion text in comprehensive and integrated discussion scenarios; Based on the conditional random field model, extracting the real-time evaluation object in the real-time discussion text and the real-time opinion words corresponding to the real-time evaluation object; Determining the emotional tendency of the real-time opinion word based on the real-time evaluation object, the real-time opinion word, and the domain emotion knowledge base; The feature weights used in the conditional random field model and the domain sentiment knowledge base are determined based on the historical discussion text in the comprehensive integrated discussion scenario, and the historical discussion text is the discussion text in the comprehensive integrated discussion scenario obtained before the real-time discussion text; the domain sentiment knowledge base stores triples of evaluation objects, opinion words, and sentiment tendencies; The determining the emotional tendency of the real-time opinion word based on the real-time evaluation object, the real-time opinion word and the domain emotion knowledge base specifically includes: If it is determined that both the real-time evaluation object and the real-time opinion word exist in the domain emotion knowledge base, then based on the triples corresponding to the real-time evaluation object and the real-time opinion word stored in the domain emotion knowledge base, the sentiment tendency of the real-time opinion word is determined; If it is determined that the real-time evaluation object is not in the domain emotion knowledge base and the real-time opinion word exists in the domain emotion knowledge base, then based on the opinion word set associated with the real-time opinion word stored in the domain emotion knowledge base, determining the emotional tendency of the real-time opinion word; If it is determined that the real-time opinion word is not in the domain emotion knowledge base and the real-time evaluation object exists in the domain emotion knowledge base, then based on the opinion word set associated with the real-time evaluation object stored in the domain emotion knowledge base, determining the emotional tendency of the real-time opinion word; If it is determined that the real-time evaluation object and the real-time opinion word are not in the domain emotion knowledge base, the emotion tendency of the real-time opinion word is determined based on the co-occurrence frequency between the real-time opinion word and the target opinion word with known emotion tendency during the search.

2. According to claim 1, the fine-grained opinion mining method for comprehensive integrated discussion scenarios is characterized in that: The domain sentiment knowledge base is constructed based on the following method: Based on the conditional random field model, extract the historical evaluation objects and the historical viewpoint words corresponding to the historical evaluation objects in the historical discussion text, and construct an initial domain knowledge base based on the historical evaluation objects and the historical viewpoint words; The initial domain knowledge base is expanded based on a general semantic dictionary to construct the domain sentiment knowledge base.

3. The fine-grained opinion mining method for comprehensive integrated discussion scenarios according to claim 2 is characterized in that: The expanding the initial domain knowledge base based on the universal semantic dictionary to construct the domain sentiment knowledge base specifically includes: For any historical viewpoint word in the initial domain knowledge base, if the any historical viewpoint word exists in the general semantic dictionary, based on the sentiment tendency of the any historical viewpoint word marked in the general semantic dictionary, construct a triple of the any historical viewpoint word, the historical evaluation object corresponding to the any historical viewpoint word, and the sentiment tendency of the any historical viewpoint word; If any of the historical viewpoint words does not exist in the universal semantic dictionary, then based on the co-occurrence frequency between the any of the historical viewpoint words and the target viewpoint words with known sentiment orientation during the search, the sentiment orientation of the any of the historical viewpoint words is determined, and a triple of the any of the historical viewpoint words, the historical evaluation object corresponding to the any of the historical viewpoint words, and the sentiment orientation of the any of the historical viewpoint words is constructed; Searching for synonyms and antonyms of any of the historical viewpoint words from the general semantic dictionary, and based on the triples, determining a triple of the synonym, the historical evaluation object corresponding to any of the historical viewpoint words, and the sentiment tendency of the synonym, and determining a triple of the antonym, the historical evaluation object corresponding to any of the historical viewpoint words, and the sentiment tendency of the antonym; Based on each triple, a sentiment knowledge base of the domain is constructed.

4. The fine-grained opinion mining method for comprehensive integrated discussion scenarios according to claim 2 is characterized in that: The extracting of the historical evaluation objects and the historical viewpoint words corresponding to the historical evaluation objects in the historical discussion text based on the conditional random field model also includes: Performing preprocessing operations on the historical research text; The preprocessing operation includes a word segmentation operation, a stop word removal operation and a useless symbol removal operation.

5. The fine-grained opinion mining method for comprehensive integrated discussion scenarios according to any one of claims 1 to 4, characterized in that: The extracting of the real-time evaluation object and the real-time opinion words corresponding to the real-time evaluation object in the real-time discussion text based on the conditional random field model specifically includes: Based on the conditional random field model, the real-time evaluation object and the real-time opinion in the real-time discussion text are extracted simultaneously through pre-defined word features, part-of-speech features and syntactic features.

6. The fine-grained opinion mining method for comprehensive integrated discussion scenarios according to any one of claims 1 to 4, characterized in that: The feature weights used in the conditional random field model are estimated from the historical discussion text based on maximum likelihood estimation and a penalty function.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the steps of the fine-grained opinion mining method for comprehensive integrated discussion scenarios as described in any one of claims 1 to 6.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the fine-grained opinion mining method for comprehensive integrated discussion scenarios as described in any one of claims 1 to 6 are implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the fine-grained opinion mining method for comprehensive integrated discussion scenarios as described in any one of claims 1 to 6 are implemented.

Images