TWI712949B - Method for calculating a semantic similarity - Google Patents

Abstract

A method for calculating a semantic similarity includes the steps of: inputting a sentence to be parsed; removing a pleonasm of the sentence according to a default pleonasm of a label; performing antonym identification on the sentence according to a default antonym of the label; substituting synonyms for words of the sentence according to a default synonym of the label; performing characteristic-word identification on the sentence according to a default characteristic word of the label to obtain a standard sentence; and calculating a similarity between the standard sentence and the label to output a response sentence semantically corresponding to the standard sentence.

Description

Semantic similarity calculation method

The present invention relates to a similarity calculation method, in particular to a semantic similarity calculation method.

With the rapid development of technology, the communication mode between humans and smart electronic devices has been carried out through the most natural and convenient voice. In recent years, robots with interactive appeals have been published one after another.

One of the more familiar human-computer interaction technologies is to first build a pre-built dialogue database on the robot for the words or questions that the user may express. When the robot receives the voice message, it will talk to the built-in dialogue database. In order to identify the semantic meaning of the voice message, the interactive conversation is conducted. However, if the function of two-way interactive communication is required, the equipment requires a huge amount of dialogue data. The dialogue database is built manually, except It takes a huge amount of time and labor costs, and it will increase the memory space for storing the database. If the established dialogue database is not continuously expanded and updated, the user will lose the freshness of the robot after a few times.

Another technology is deep learning through neural networks. In practice, neural networks are mostly implemented using supercomputers or single-chip systems. In the case of a single-chip system, the same set of circuits in the single-chip system can play different computing layers in a multilayer artificial neural network at different time points. The more layers of the neural network, the more complex the simulation can be. However, when the number of layers increases, the number of neurons required in the entire network will increase substantially, which will result in a huge hardware cost burden, and the cost of each computing layer Input data, learnable parameters and The amount of data of the calculation results is very considerable, which is by no means the burden of ordinary enterprises.

The above shortcomings all show the various problems derived from the use of conventional human-computer interaction technology. According to the current development of artificial intelligence, it is still not easy to achieve autonomous human-computer interaction. After all, language is a cultural product of humans through long-term learning and experience. Therefore, How to use the limited dialogue database and quickly capture the user’s semantics for analysis is an important topic.

In view of this, the purpose of the present invention is to provide a semantic similarity calculation method, which includes the following steps.

Input a sentence to be parsed, and process the sentence and the pre-set vocabulary of each marker word to remove the miscellaneous words, and extract the words in the sentence and the pre-set antonym of each marker word for antonym checking , The sentence is replaced with similar words preset for each marker word, and the sentence is checked with the characteristic words preset for each marker word to obtain a semantically parsed rule sentence. The similarity between the rule sentence and the marked word is calculated to output a response sentence corresponding to the semantic meaning of the rule sentence.

Another technical means of the present invention is that the sentence and the predetermined constant feature words of each tag word are first checked for feature words, and then the feature word check is performed with the variable feature words preset for each tag word, and the difference between the tag words The feature words include at least one constant feature word, at least one variable feature word, or a combination thereof, and each variable feature word has a plurality of associated feature words related to the variable feature word.

Another technical means of the present invention is that the sentence and the constant feature word preset for each marker word are first checked for feature words, and then the feature word check is performed with the variable feature words preset for each marker word, and the marker word The feature words include at least one constant feature word, at least one variable feature word, or a combination thereof, each variable feature word has a plurality of associated feature words related to the variable feature word, and the plural variable feature words are in an intersection relationship with each other.

Another technical means of the present invention lies in the above complex The characteristic words of the number variables have the sequence of order.

Another technical means of the present invention is that the above is to check the constant feature word first, and then check the variable feature word.

Another technical means of the present invention is that the above-mentioned constant feature word and the variable feature word that need to be matched with the signature word for feature checking are required to obtain the rule sentence.

Another technical means of the present invention is that the above-mentioned response sentence extracts the constant feature word corresponding to the marker word or the variable feature word is respectively provided with at least one constant response feature word, at least one variable response feature word, or a combination thereof, and The constant response characteristic word and the variable response characteristic word are set corresponding to the sequence of the constant characteristic word and the variable characteristic word.

Another technical means of the present invention is that the above-mentioned similarity calculation will perform a feature word check based on the sequence of the constant feature word and the variable feature word.

Another technical means of the present invention is to check the response sentence against a preset matching degree threshold after the similarity between the rule sentence and the marked word is calculated, and to retain those that are greater than the matching degree threshold Response sentence to output the response sentence.

Another technical means of the present invention is that when the sentence cannot be matched with any marked word in the rule base, that is, according to the word of the sentence and a marked word in an extensive rule base, a match is obtained. A broad response sentence based on the words of the sentence.

The beneficial effect of the present invention is that by setting at least one constant feature word, at least one variable feature word or a combination of the two in the tag word, and the variable feature word is further provided with an associated feature word related to the variable feature word itself, The feature word check is performed in the expression mode of the diversified user sentence, and the feature word is responded to by the corresponding set of variables. There can be a variety of different answers. In addition to reducing the time for manpower to set the marker word and the computing time of the computer, Can greatly enhance the flexibility of human-computer interaction to meet Use requirements in different fields and occasions.

91~97‧‧‧Step

Fig. 1 is a schematic flowchart illustrating a preferred embodiment of the semantic similarity calculation method of the present invention.

The features and technical content of the related patent applications of the present invention will be clearly presented in the following detailed description of the preferred embodiments with reference to the drawings.

Refer to FIG. 1, which is a preferred embodiment of the semantic similarity calculation method of the present invention. It is suitable for analyzing the semantics of the communication between the user and the robot and generating a corresponding response. The method includes the following steps.

First, proceed to step 91, input a sentence to be parsed, and process the sentence and the pre-defined miscellaneous words of each tag word to remove the miscellaneous words. The miscellaneous word removal refers to removing the superfluous words in the question sentence, and preset The miscellaneous words can be zero or plural, for example, ask, if, for example, like... and other spoken meaningless words, and the input sentence can be directly talked with the robot by the user, or captured The process of taking the speech and then converting the speech to text or text to speech is not the technical focus of the present invention, and will not be repeated here. Here, use the labeled word field "Do you like {xq0}{xq1}" in Table 1 below as the description of this embodiment, and the de-clutter field lists preset de-clutter words such as: relatively speaking, Compare, see, etc.

Then, proceed to step 92 to extract the words in the sentence and the preset antonyms of each tag word for antonym checking. The antonyms are based on the words opposite to each marker word, and the preset antonyms can be zero or plural. For example, the antonyms of willing in the marker words can be unwilling, unloving, Don’t want, don’t need, don’t, etc., if the default antonym of the tag word appears in this step, it means that it has a different meaning from the tag word and will be checked with other tag words. The preset tag words can be targeted at hospitals, schools, travel A dictionary of marked words used in a particular field or occasion, such as parks, department stores, etc., is used for customer service consultation on that occasion. The antonyms of the marker words in Table 1 below are preset to be stars.

Then, proceed to step 93 to replace the sentence with the preset similar words of each tag word. The similar words are set according to each tag word, and the preset similar words can be 0 or plural. For example, the column of similar words like favorites in the marked words of Table 1 above lists likes, favorites, preferences, etc., or the founding father and Sun Yat-sen, etc. In actual implementation, it is also possible to perform step 93 similar word replacement first and then perform step 92 antonym check, which should not be limited to this.

Then, proceed to step 94. The sentence and the preset constant feature words of each tag word are checked for feature words. As shown in Table 1 above, the constant feature word field is preset as you like. If "you like" appears, it means that the tag word If the meaning is the same, proceed to the next step, otherwise, check with other marked words.

Then, in step 95, the sentence and the variable characteristic words preset for each tag word are checked for characteristic words to obtain a semantically parsed rule sentence. The feature words of the marker words include at least one constant feature word, at least one variable feature word, or a combination thereof. In addition, each variable feature word has a plurality of associated feature words related to the variable feature word. Here, the characteristic word can be a noun, a verb, a gerund or an adjective. As shown in Table 1, the variable characteristic word column is preset with basketball, dancing, dancing, etc.

Further, the feature words of the plural variables are in an intersection relationship with each other, and the feature words of the plural variables have a sequence, such as {xq0}{xq1} in the marker words in Table 1 above. In the process, the constant feature word is checked first, and then the variable feature word is checked. Furthermore, the constant feature word and the variable feature word of the marker word for feature checking must be met at the same time to obtain the rule sentence.

Next, proceed to step 96 to calculate the similarity between the rule sentence and the marked word to output a response sentence corresponding to the semantics of the rule sentence. As shown in Table 1, the response sentence field is preset to: I prefer {xq1}Some, I like both of these sports! , {Xq0} and {xq1} are great exercises! . When {xq0} and {xq1} are basketball and football respectively, the response sentence is: I prefer football, and I like both of these sports! , Basketball and football are both very good sports!

In step 96, the response sentence extraction corresponding to the constant feature word of the tag word and the variable feature word are respectively provided with at least one constant response feature word, at least one variable response feature word, or a combination thereof, and the constant response feature word The variable response characteristic word is set corresponding to the sequence of the constant characteristic word and the variable characteristic word. Pay special attention to the similarity meter Calculate will check the characteristic words according to the sequence of the constant characteristic words and the variable characteristic words.

Finally, proceed to step 97. After the similarity between the rule sentence and the marked word is calculated, the response sentence is checked against a preset matching degree threshold, and the response sentence that is greater than the matching degree threshold is retained. The response sentence is output, otherwise, the response sentence is not output. Generally speaking, the similarity will be set at about 70~80%. Although the de-cluttering process is performed from the beginning, in order to avoid the useless clutter that has not been completely removed, the preset matching degree threshold is used. To ensure the correctness of the calculated results.

In particular, when the sentence cannot be matched with any of the preset marked words, that is, one of the antonym check, the feature word check, and the similarity calculation fails to match, that is, according to the feature word of the sentence Matching with a marked word database in an extensive rule database to obtain a broad response sentence based on the characteristic words of the sentence.

Suppose the semantic meaning is X1, X2, X3..., X1 has n expressions such as X11, X12, X1...X1n, and so on X2, X3, suppose the user’s question or sentence to be parsed is Y1, When calculating whether Y1 is expressing the semantic meaning of X1, the conventional practice is to calculate the similarity between Y1 and X11, X12, X13,...X1n respectively, so it is necessary to establish n pen label words (Label), which is X11, X12, X13,...X1n, this will consume construction time and labor costs. The present invention solves the complicated method of establishing n-stroke marker words, that is, one marker word represents X11, X12,...X1n and other n-stroke expressions, and X11 is used as marker word to represent X11-X1n, X21-X2n, X31 -X3n.... to Xmn and other m semantics, so that it can get mxn expressions, in addition to saving the time to build the rule base, it can also shorten the calculation time to improve the real-timeness of human-computer interaction.

First of all, take the example of antonym checking. The marked word is Do you like football? If the sentence is Do you like football, the similarity is extremely high during the similarity calculation. On the contrary, if the sentence is that you don’t like football. foot The similarity of the ball is also very high. Although the two sentences only have a different word "no", the semantics of the two sentences are completely different and opposite. Therefore, through the antonym check step, the sentence can be input at this step." Do you don’t like football” is set to match the result of the mismatch with the tag word “Do you like football”, and then compare it with other tag words. There is no need to perform subsequent replacements with the tag word “Do you like football?” Steps such as inspection or analysis to save analysis time and improve analysis accuracy.

Next, let’s take the example of the marked words including and not including variable characteristic words. The marked word X11 is "you{like}{足球}?". The words like and football in parentheses are constant characteristic words, and the language expressions are diverse. For example, X12 "Do you like football?" X13 "Do you like football?" X14 "Do you like football?" X15 "Honestly, do you like football games?" X16 "Do you not like football?", When the sentences X12~X16 are input separately, X13's Kan and X15's are honestly preset and meaningless words, which will be removed in this step 91, X12's love kick, X13's love to watch, X14's like or not , X15 likes to watch, and the marked word likes are the preset similar words. In this step 93, replace with like, X16 do you not like football and the marked word Do you like football are the preset similar words, and will also be in It is replaced in this step 93.

Then, perform the feature word check. In this step, the two feature words like like and football must exist at the same time, and the feature word “like” is checked first before the feature word “football” is checked. After that, X12 will you love to play football? Do you like football for similarity calculation; X13 will change from whether you like watching football to do you like football for similarity calculation; X14 will change from whether you like football to you like football for similarity calculation; X15 will be honestly , Do you like watching football matches become you like football for similarity calculation; X16 will be changed from Do you not like football to Do you like football for similarity calculation. After the processing flow of this design, X12~X16 can get a very high similarity value when calculating the similarity, so as to recognize it and the label word (X11) Have the same meaning.

In addition, take an example of a variable feature word included in the tag word. When the tag word is "Do you like {soccer/basketball/volleyball/yoga...}", the parenthesis is a variable feature word, and it is set as Related related feature words basketball, volleyball, yoga... and other m related feature words, and the variable response feature word of the response sentence corresponds to the variable feature word setting. The bracket is a variable response feature word. The response sentence can be designed to answer It’s "I love {playing football/basketball/volleyball/yoga...} the most", {football/basketball/volleyball/yoga...} is also one of my favorites, {football/basketball/volleyball/yoga...} is a bit difficult, I I don’t like waiting for multiple designs to answer.

Further, when the input sentence is "Do you like basketball?", when the feature word check is performed, it matches the associated feature word "Basketball" in the variable feature word, and the sentence is changed to "Do you like basketball", and then the similarity is calculated Very high similarity. In the end, the answer to the response sentence is "I love basketball the most", or "Basketball is also one of my favorite items", or "Basketball is a bit difficult, I don't like it very much", etc. The answer. And if the input sentence is "Do you like volleyball", the corresponding answer of the response sentence is "Volleyball is a bit difficult, I don't like it very much".

In summary, the semantic similarity calculation method of the present invention sets at least one constant feature word, at least one variable feature word, or a combination of the two in the tag word, and the variable feature word is further set with the variable feature word itself Related related feature words are checked by the expression of multiple sentences, and the feature words are responded to by corresponding variables. There can be a variety of different answers, in addition to reducing the time of manpower setting up the tag words and computer calculations In addition to time, it can greatly increase the flexibility of human-computer interaction to meet the needs of different fields and occasions, so it can indeed achieve the purpose of the invention.

However, the above are only preferred embodiments of the present invention, and should not be used to limit the scope of implementation of the present invention, that is, according to the present invention. The simple equivalent changes and modifications made to the scope of the patent application and the description of the invention are still within the scope of the patent for the present invention.

91~97‧‧‧Step

Claims (7)

A method for calculating semantic similarity includes the following steps: input a sentence to be parsed, and perform de-cluttering processing on the sentence and the pre-set vocabulary of each tag word; extract the words in the sentence and each tag word Perform antonym check for preset antonyms; perform similar word replacement between the sentence and the preset similar words of each tag word; perform feature word check for the sentence and the preset feature words of each tag word to obtain For a rule sentence after semantic analysis, when checking the characteristic words of the sentence, the sentence and the constant characteristic words preset for each marker word are first checked for characteristic words, and then with the variable characteristic words preset for each marker word Feature word check, and the feature word of the mark word includes at least one constant feature word, at least one variable feature word, or a combination thereof, and each variable feature word has a plurality of associated feature words related to the variable feature word, and The characteristic words of the plural variables are arranged in a sequence; and the similarity calculation is performed on the rule sentence and the marker word to output a response sentence corresponding to the semantic meaning of the rule sentence. According to the semantic similarity calculation method described in item 1 of the scope of patent application, Wherein, when the feature word check of the sentence is performed, the sentence and the constant feature word preset for each tag word are first checked for feature words, and then the feature word check is performed with the variable feature words preset for each tag word, and the The feature words of the marker words include at least one constant feature word, at least one variable feature word, or a combination thereof. Each variable feature word has a plurality of associated feature words related to the variable feature word, and the plural variable feature words are mutually exclusive Intersection relationship. According to the semantic similarity calculation method described in item 2 of the scope of patent application, when the feature word check of the sentence is performed, the constant feature word of the marker word for feature check and the variable feature word must be met at the same time to obtain The rule statement. According to the semantic similarity calculation method described in item 2 of the scope of patent application, in the similarity calculation, the response sentence extracts the constant feature word corresponding to the tag word or the variable feature word is respectively provided with at least one constant response feature Word, at least one variable response feature word, or a combination thereof, and the constant response feature word and the variable response feature word are set corresponding to the sequence of the constant feature word and the variable feature word. According to the semantic similarity calculation method described in item 2 of the scope of patent application, in the similarity calculation, the characteristic word check is performed according to the sequence of the constant characteristic word and the variable characteristic word. According to the semantic similarity calculation method described in item 1 of the scope of patent application, in the similarity calculation, after the similarity calculation between the rule sentence and the marker word, the response sentence is matched with a preset matching degree The threshold is checked, and the response sentences that are greater than the matching threshold are retained, To output the response sentence. According to the semantic similarity calculation method described in item 1 of the scope of patent application, when the sentence cannot be matched with any marked word in the rule base, it is based on the words of the sentence and the words in an extensive rule base. The marked words are matched to obtain a broad response sentence based on the words of the sentence.

Images