JPH0634207B2 - Topic prediction device - Google Patents

Description

The present invention relates to a system and apparatus for predicting the topic of a given sentence in natural language processing.

(Prior art and its problems) Conventional methods of natural language analysis have mainly focused on processing based on syntactic information of sentences with good characteristics, including machine translation (Reference: Japanese Machines through Semantic Structure) Translation system Nikkei Electronics December 17, 1984 issue). However, linguistic information exchanged between humans usually contains noises such as typographical errors, omissions, grammatical deviations, omissions, etc. If you want to process all parts of a sentence uniformly,
There is a drawback that analysis is impossible due to the noise part in the input sentence.

In order to deal with such noise, we realized on a computer how to grasp the contents of a sentence, and how to recognize important and unimportant parts of a sentence, as if humans were reading diagonally. , It is important to predict what the speaker intends and which is noise. An object of the present invention is to provide a method and apparatus for predicting the topic of a given sentence.

(Means for Solving Problems) A first invention of the present application analyzes an input sentence by using a grammatical rule of a target language, and determines a word or phrase that may define a topic of the input sentence. Extracting and predicting the topic of the sentence from the extracted words and phrases, and predicting the topic predicted from the most words and phrases in the input sentence among the predicted topics to the topic of the input sentence Is a topic prediction method.

A second invention of the present application is a topic dictionary in which candidates of topics predicted by each word are registered in advance, a morphological analysis unit that divides an input sentence into word strings, and an output of the morphological analysis unit. And a syntactic analysis unit that analyzes the grammatical structure of the input sentence, and an output of the syntactic analysis unit that extracts only the main constituent elements of the input sentence from the syntactic analysis result. Section, the topic candidate predicted for each word extracted by the main component extraction section with reference to the topic dictionary is extracted, and the number of times extracted for each extracted topic candidate is held,
The topic prediction device is characterized by comprising a topic candidate having the highest number of appearances, and a topic selection unit that outputs the topic candidate as a topic of the input sentence.

(Operation and Principle) The present invention has solved the problems of the prior art by the above means. Next, the operation of the present invention will be described using an example. As an example, consider an example of a posted message for buying and selling second-hand goods using mass media.

"In addition, I will add a charger.""For those who bought it, I will add a charger.""The charger is an extra." In the above example, the topic of the sentence is In addition to buying and selling items for buyers, it also means adding bonuses. Since it is premised that the content described here is limited to the topic of buying and selling second-hand goods,
From the appearance of words that are keywords such as “bonus” and “attach”, the content of the topic described in the above sentence can be predicted.

However, in the above-mentioned method of simply showing a certain topic from the input sentence when a specific pattern is found,
Even if it is known in advance that the target area described in the input sentence is small, it is often wrongly predicted. Here is an example.

If you simply evaluate "The delivery of goods and bonuses is by hand."
"Explanation of items for sale", "Bonus"->"Offer of bonus",
The topic to be predicted for the sentence of "passing"->"exchangemethod","handing"->"exchangemethod" diverges. In order to avoid this, the grammatical structure of the input sentence is analyzed and whether a certain word is simply a modifier,
Alternatively, it is only necessary to determine whether or not it is the central constituent element of the case that constitutes the input sentence, and to examine the topic prediction only for the word that is the central constituent element of the case that constitutes the sentence. Cases include nominative cases (persons who perform actions), instrumental cases (tools of actions), target cases (objects of actions), source cases (starting point, cause of actions),
There are goal cases (return points of actions), place cases (places of actions), time cases (time of actions), and the like.

Then, the phrase “delivery of goods and bonuses” is a nominative case, and its central component is “delivery”, and the part of “commodities and bonuses” modifies “delivery”. There is only. Therefore, if “commodity” and “bonus” are ignored, and topic prediction inspection is performed only on “delivery”, the topic prediction of the sentence is “delivery”-> “exchange method”, “hand delivery”. -> It is predicted to be one like "Exchange method".

In the present invention, in an application in which it is known in advance that the target area described in the input sentence is narrow in this way, the relationship between the word that is a keyword in the target area and the topic candidate in the target area predicted by the word is previously determined. If you prepare a dictionary and describe a sentence, analyze the given input sentence according to grammatical rules, extract the words that are the main constituent elements of the sentence as keywords, and predict from those keywords. In this method, topic candidates are counted and the topic candidate predicted from the most keywords is predicted as the topic of the sentence. For example, in, since both the “bonus” and “attach” predict the topic of “offer offer”, the topic of is “offer offer”. Similarly,
Then, the topic of "exchange method" will be predicted from both "delivery" and "handing".

(Example) Hereinafter, the detail of this invention is demonstrated according to an Example.

FIG. 1 is a block diagram showing an embodiment of the first invention of the present application. In FIG. 1, reference numeral 1 denotes a topic dictionary in which topic candidates predicted from each word are registered in advance in word units. Reference numeral 2 is a topic-indicator segmentation unit that grammatically analyzes the input sentence and extracts words that are the main constituent elements of the input sentence. 3 refers to the topic dictionary, does nothing for each of the words extracted by the topic indicator cutout unit 2 if there is no topic candidate to predict, and if there is, extracts it and for each extracted topic candidate, Keep the number of times it was taken out,
When the extraction of the topic candidates and the counting of the number of appearances of all the words extracted by the topic-indicating word cutout unit 2 are completed, the topic candidate having the highest number of appearances among the taken-up topic candidates is extracted from the input sentence. It is a topic selection unit that outputs as a topic.

First, the input sentence is input to the topic instruction word slicing unit 2 and is segmented into words according to the part of speech. For example, the input sentence "Additional charge to the charger." Is "extra", "ni",
It is divided into "charger" and "attach." Next, the structure of the input sentence is analyzed according to the grammar rules. "As a bonus"
It can be seen that is the source case, "charger" is the target case, and "attach" is the main verb. It can also be seen that the central components of each case are "bonus" and "charger".

Next, the central constituents of each case and the words of the main verb are cut out one by one and input to the topic selection unit 3. When the topic selection unit 3 receives a word, it refers to the topic dictionary 1 to end the processing of the word if the word is not registered as a keyword in the topic dictionary 1, and if it is registered, the topic predicted by the word. The candidate is taken out from the topic dictionary 1 and the topic item and the number of appearances are stored. This process is performed for all the central constituent elements and words of the central verb of the given input sentence, and the topic candidate with the highest frequency of occurrence is output as the topic of the given input sentence.

Next, an apparatus for specifically realizing this method will be described as an embodiment of the second invention of the present application. FIG. 2 is a diagram showing an embodiment of the second invention of the present application. In FIG. 2, 10 is a morphological analysis unit, 11 is a syntax analysis unit, 12 is a main component extraction unit, 13 is a topic dictionary, 14 is a keyword comparison unit, 15 is a topic candidate storage unit, and 16 is a topic prediction unit. .

Here, the explanation of FIG. 2 will be made with reference to the example sentence “delivery of goods and bonuses is hand delivery” which is an example sentence used in the section (action and principle).

The morpheme analysis unit 10 divides the input sentence into parts of speech.
Example sentences are "product""to""bonus""no""delivery"
It is divided into "ha,""handing," and "is." Various methods have already been proposed for morphological analysis processing, but in the present invention, one of them is not specified (reference: Japanese information processing Takahashi, Tanaka et al. Modern science company).

The output of the morphological analysis unit 10 becomes the input of the syntax analysis unit 11.
The parsing unit 11 parses an input sentence according to grammatical rules. As a result, "passing" is a nominative, and "passing"
It can be seen that "da" is the main verb and "commodity and bonus" is a modifier to the nominative. Various methods have already been proposed for parsing processing, but in the present invention, any one of them is not specified (reference: Japanese information processing Takahashi, Tanaka et al.
Modern science company).

The output of the syntax analysis unit 11 becomes the input of the main component extraction unit 12. When the main component extraction unit 12 receives the output of the syntax analysis unit 11, it outputs only the words that are the main components of the case in the input sentence and the main verb, and discards other modifiers. Play the role of. Therefore, the only remaining part of the example sentence is "delivery" and "handing".

The topic dictionary 13 has the structure shown in FIG. 3, and the topic candidate predicted from the word “delivery” is, for example, the “exchange method”.

When the keyword comparison unit 14 receives one word from the main component extraction unit 12, the keyword comparison unit 14 searches the topic dictionary 13 using that word as a key. If the word is not registered in the topic dictionary 13, the processing of the word is terminated and the main component extraction unit
Receive the next word from 12. If the word is registered in the topic dictionary 13, the topic candidates read from the topic dictionary 13 are output. In the example sentence, for example, when "delivery" is received, "exchange method" is output.

As shown in FIG. 4, the topic candidate storage unit 15 has an internal state of a topic candidate that has appeared and the number of appearances of the topic candidate. When the topic candidate storage unit 15 receives a topic candidate from the keyword comparison unit 14, it checks whether it is the first appearance,
If so, the topic candidate name and the number of appearances (that is, 1
Stored in the topic candidate storage unit 15. If it has already appeared, the number of appearances is incremented by 1 and the process ends. In the example sentence, since the topic candidates of “exchange method” are extracted from the two keywords “delivery” and “hand delivery”, the number of times is “2” as shown in FIG.

Depending on the input sentence, the number of appearances of two or more topic candidates may be the same. The present invention describes a method for predicting the topic of an input sentence, and does not specify how to use the prediction to proceed with the process of interpreting the input sentence, but the same possibility If there are multiple topic candidates that have the above, for the time being, select one of the topic candidates to proceed with the analysis, and if the analysis becomes impossible on the way, try another candidate Be done. PROLOG is suitable for such processing (reference: knowledge representation and Prolog / KR Hideyuki Nakajima Sangyo Tosho, 1985).

When the above process is completed for all the words in the input sentence, the topic prediction unit 16 reads the internal state of the topic candidate storage unit 15, selects the topic candidate name with the highest number of appearances, and outputs it. This topic candidate is the topic of the input sentence and is output to the outside.

(Effect of the Invention) According to the present invention, it is possible to determine the topic of a sentence in natural language processing, so that the content of the sentence can be analyzed based on prediction. In addition, even for a noisy input sentence, which was not good in the conventional method, it is possible to determine what is important and what is not important in the content of a certain sentence by using the present invention. It can analyze many natural language sentences. The method proposed in the present invention is
It is suitable for constructing a natural language dialogue system as an interface for a system, a database, or an expert system that extracts only the user's necessary information from a large amount of natural language sentences.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a diagram showing a configuration of an embodiment of a first invention of the present application,
FIG. 2 is a diagram showing the configuration of an embodiment of the second invention of the present application,
FIG. 3 is a conceptual diagram showing the configuration of the topic dictionary in the embodiment of FIG. 2, and FIG. 4 is a conceptual diagram showing the configuration of the topic candidate storage unit in the embodiment of FIG. 1 ... Topic dictionary, 2 ... Topic index extraction unit, 3 ... Topic selection unit, 10 ... Morphological analysis unit, 11 ... Syntax analysis unit, 12 ... Main component extraction unit, 13 ... Topic dictionary, 14 ... Keyword comparison unit,
15 ... Topic candidate storage section, 16 ... Topic prediction section.

Claims (1)

[Claims] 1. A topic dictionary in which topic candidates predicted by each word are registered in advance on a word-by-word basis, a morphological analysis unit that divides an input sentence into word strings, and an output of the morphological analysis unit is received. A syntactic analysis unit that analyzes the grammatical structure of the input sentence; a main constituent element extraction unit that receives the output of the syntactic analysis unit and extracts only words that are the main constituent elements of the input sentence from the syntactic analysis result; The topic candidate predicted for each word extracted by the main component extraction unit with reference to the topic dictionary is extracted, and the number of times extracted for each of the extracted topic candidates is held, and the number of appearances is highest. And a topic selection unit that outputs the topic candidates that were the most common as the topic of the input sentence.