WO2003012681A1

WO2003012681A1 - Method and system for automatically enhancing semantic resources with a real-time question-answer electronic system

Info

Publication number: WO2003012681A1
Application number: PCT/FR2002/002513
Authority: WO
Inventors: Olivier Collin; Florence Duclaye
Original assignee: France Telecom
Priority date: 2001-07-19
Filing date: 2002-07-15
Publication date: 2003-02-13
Also published as: FR2827685A1; EP1407389A1; FR2827685B1

Abstract

The invention concerns a method for enhancing semantic resources. The invention is characterised in that it consists in using a real-time question/answer electronic system (1) to obtain automatic enhancement of said resources (2).

Description

METHOD AND SYSTEM FOR AUTOMATICALLY ENRICHING SEMANTIC RESOURCES FROM A REAL-TIME QUESTION-ANSWER ELECTRONIC SYSTEM.

The invention relates to a system for the automatic enrichment of semantic resources from a real-time electronic question-answer system.

By electronic question-answer system is meant a system according to which a user asks a question in natural language from a computer tool such as a microcomputer (PC) connected through a communication network to a system comprising a database server in order to obtain information constituting the answer to his question.

Most often the questions can only be presented in an extremely simplified form, that is to say in the form of keywords. Indeed, the oldest systems are not able to understand a sentence and can only work with keywords.

The most recent systems incorporate a linguistic analysis to translate the question into a specialized language adapted to the search for information by a search engine.

Unfortunately current systems are limited if they do not have significant linguistic resources. Indeed, the system may be quite capable of finding the answer to the question, "Who killed Kennedy?" And not at all capable of answering the question. "Who is Kennedy's murderer?" does not have a database of synonyms words or paraphrases. The present invention overcomes this problem.

The invention provides a method for automatic enrichment of semantic resources from a real-time electronic question-and-answer system. The real-time electronic question and answer system is the tool used in accordance with the invention to automatically enrich the semantic resources of a semantic network. This enrichment system can therefore be used to enrich the semantic resources of a real-time electronic question-answer system and more generally of a semantic network.

According to the invention, the system is capable of subjecting the question posed to a linguistic analysis operation in order to construct a pattern for finding the answer which is presented in the simplest form: argumentl - relation - argument2. This pattern is then sought and extracted (by direct matching) from the documents reported by the search engine or argument 1 is the response sought.

According to the invention, the system integrates an operation of re-formulation of the questions in order to search for the same answer by means of several formulations which are se cantically equivalent. The re-formulation of the questions increases the quantity of documents that can provide an answer.

According to the invention, the re-formulation of the questions is done by learning and makes it possible to automatically obtain a set of formulations semantically equivalent to a first formulation of a relationship between two arguments.

The invention more particularly relates to a method for enriching semantic resources mainly characterized in that it consists in using an electronic real time system of question answers to obtain an automatic enrichment of said resources. According to another characteristic, the method comprises the following phases:

-an initialization phase including:

• the establishment of a search pattern in the form of a relationship between two or more arguments based on the keywords of the initial question received, the research relating to one of the boss's arguments being known,

• the search for at least one answer from a search system coupled to one or more databases of texts formulated in natural language through one or more communication networks,

• the extraction of one or more responses from documents provided by the research,

- a learning phase comprising: • a search for one or more re-formulations of the relationship formulated in the pattern of the initial question, based on the arguments known by this question and the answer or answers obtained,

• the storage of these re-formulations on a database coupled to one of the communication networks.

According to another characteristic, the search for a re-formulation includes:

- returning known arguments to determine at least one other relationship between these arguments, - searching for documents containing said known arguments,

- extracting sentences containing known arguments from these documents, - a linguistic analysis of the sentences to extract one or more relations of the paraphrase type from the relation present in the pattern of the initial question, - the storage of relations of the paraphrase type from the relation of the initial question.

The subject of the invention is also a system of enrichment of semantic resources, mainly characterized in that it is coupled to an electronic real time system of question answers to obtain an automatic enrichment of said resources.

According to another characteristic, the system includes: - means for implementing an initialization phase of these means comprising:

• means of establishing a search pattern in the form of a relationship between two or more arguments from the keywords of the initial question received, the search relating to one of the pattern of the other's arguments being known ,

• means for searching for at least one response, coupled to one or more databases through one or more communication networks, • means for extracting one or more responses from documents provided by the means of research, means of implementing a learning phase comprising: • means of researching one or more reformulations of the relation formulated in the pattern of the initial question from the arguments known by this question and from the or answers obtained, • means for storing these re-formulations on a database coupled to one of the communication networks.

According to another characteristic, the means of searching for a re-formulation include:

- means of returning the arguments known by the original question,

- means of searching for documents containing these arguments, - means of extracting sentences containing known arguments, from these documents,

- means of linguistic analysis of sentences to extract one or more relations of the paraphrase type from the relation present in the pattern of the initial question,

- means for storing relations of the paraphrase type of the relation of the initial question.

The reference means, the means of extracting sentences containing the arguments known to these documents and the means of linguistic analysis of the sentences are produced by modules of computer programs or processes capable of implementing the stated functions and the means of search are carried out by a program called search engine.

Other particularities and advantages will appear clearly on reading the description which is made and which is given by way of nonlimiting example and with reference to the drawings in which: - Figure 1 represents the functional diagram of a system of automatic enrichment of semantic resources according to the invention,

FIGS. 2A, 2B represent the detailed functional diagrams of the system of FIG. 1, - Figure 3 shows the device for implementing the functional blocks of Figures 2A and 2B according to one embodiment.

According to the invention, an electronic system for real-time information search is used. This system is coupled to the system of enrichment of semantic resources through a communication network. It can be an intra-net network giving access to one or more databases which can provide answers to questions asked by a user. It can also be the Internet.

The search system, whatever it is, is qualified as a basic system, it being understood that it can rely only on a linguistic analysis operation making it possible to extract at least one answer to a question asked in natural language by a user using to do this keywords.

It is further considered according to the invention that the system is capable of constructing the question in the form of a search pattern: argument 1, relation, argument 2.

For example to the question: "Who killed Kennedy? ", The simplest boss in the" murderous "form killed Kennedy.

This pattern then allows thanks to the search engine of the system to extract the response "X" in the phase found in one or more documents found at the end of the search "X killed Kennedy". This basic search system carries the reference

I in FIG. 1. It is coupled through a communication network R to one or more databases 3.

It can of course be the WEB in the case where the network is Internet. The system also includes a device for automatic enrichment of linguistic resources 2. This device carries out learning operations for re-formulations to automatically obtain a set of semantic formulations equivalent to a first formulation of a relationship between two arguments. .

The overall process is inspired by the so-called “bootstrapping” methods which make it possible to model classes of similar examples from a limited set of examples.

Learning follows the initialization of an acquisition (block 1, figure 1).

As we saw in the initialization stage of an acquisition, the main terms of the question are extracted to build a query for the "standard" search engine taking the form of a series of keywords or a single text string: Question: Who killed Kennedy? Keywords: "killed, Kennedy".

The question is submitted to a linguistic analysis allowing to build a pattern of research for the answer; the pattern which is then sought makes it possible to extract the response. When the answer is known, we again use the data from the databases or the WEB to find relations linking the answer (X) and argument 2 (Kennedy) by a query containing these two arguments: X, Kennedy.

On the other hand, we use a proximity operator to control the distance between the arguments and to get rid of insertions providing syntactic complexity. This operator allows control of distance between arguments in number of words (1, 2,

Solution sentences: X, Kennedy's murderer ... Kennedy's murderer, X ...

A simple syntactic treatment makes it possible to extract the synonymous relationships and more generally the paraphrases: "murderer", "assassin", "criminal"

In a dual way, the formulations obtained from initial arguments of which the semantic category is known, allow to categorize new arguments:

"Murderer" kills "name" "murderer" killed "name" "murderer" kills "name".

Using the query: "murderer", we can extract arguments whose labels are "murderer" and "name" and thus generate a list of murderers and murdered people.

We see that, step by step, the system establishes several relationships which globally form a more abstract relationship generalizable by the same predicate of the type:

[murderer] [predicate_generate] [name]

The treatment of the question also makes it possible to obtain semantic labels: Who killed President Kennedy? allows to extract the “president” label by a simple syntactic treatment. Thus, the overall process consists in acquiring new data and new relationships by building on the previously acquired data and relationships. For a given statement, it is a question of exploring a new category of relations, starting from a known category by limiting the variable element.

First consequence: this technique makes it possible to gradually enrich all of the known relationships of a natural language processing system (for example a semantic network).

This system allows the automatic creation of a semantic network. This enrichment can then be used for all applications using such a system. In particular, the system is directly usable for the extraction of information involving the acquired relationships. For example, to extract from a text all of the assassins and murdered people, we can use the relationships learned as "murderer of" and "killed". Second consequence: the system allows in the same way to acquire formulations in another language by using multilingual data from the WEB. In response to the query [X, Kennedy], we get: X assassinated Kennedy, X killed Kennedy. This constitutes a learning system for translation and which makes it possible to provide a translation into another language.

The different functional blocks of FIGS. 2A and 2B make it possible to illustrate in detail the operations implemented during the initialization phase I and during the learning phase II operated respectively by the search system 1 and by the device enrichment 2 illustrated in Figure 1. The initial phase is carried out by a search system which according to a preferred embodiment implements operations 10 to 17 of FIG. 2A.

In the example taken, the question asked relates triplets of the type Argumentl, Relation, Argument2 (step 10). A linguistic analysis (step 11) makes it possible to obtain the extraction pattern of Argumentl: semantic type of Argl, Structure Relation-Arg2 (step 15). The search engine searches Arg2 close to Relation on the databases to which it has access or across the Web. According to the example taken:.

Relationship = murderer of

Argument2 = Kennedy We thus obtain HTML documents (step 13),

An extraction of the sentences containing Relation and Arg2 is implemented by the basic system (step 15).

A second linguistic analysis makes it possible to search in these documents for the structure of the Argumentl-Relation-Argument2 type (step 16). Knowing Argument 2 and Relation (from the question) the analyzer extracts Argument 1. Step 16 thus makes it possible to extract the answer Argument 1. For each question asked the system proceeds to this first phase to find at least one answer . The system continues with the next phase II which allows it to carry out a learning process. Knowing Argument 1 and Argument 2 (step 20; Arg 1 and Arg 2 are known after the initialization step), the search engine performs a search using the keywords Argument 1 close to Argument 2 (step 21) . The search engine finds documents in response and provides them in HTML format (step 22). These documents are filtered to obtain a well-formed text format and language paragraphs.

In step 23, a filter is used to extract sentences containing Argument 1 and Argument 2. A linguistic analysis is performed on these sentences (step 24).

This analysis extracts N instances of

"Relation" between Argument 1 and Argument 2 (step 25).

At the end of each learning phase, the system obtains synonyms of the desired relationship and more generally paraphrases and stores these synonyms in a database. The system therefore makes it possible to enrich the resources for research on a question asked.

Indeed at the time of the second linguistic analysis the system brings the stored synonyms.

It can of course be provided that the learning phase is automatically inhibited when for a given question, already asked a large number of times, a sufficient number of synonym or paraphrase is recorded. For this a threshold for this number will be fixed. It then suffices that the program implementing the steps which have just been described provides for a comparison test between the threshold value and the number of synonyms recorded. We will detail in the following the different modules to implement the steps that have just been detailed. These modules are computer processes or programs capable of implementing the stated functions. Element 1 for searching for the first answer (FIG. 1) comprises a module 100 for formatting and calling a module for linguistic analysis 110 and a search engine 120. The module 100 receives on an input to the question Q in ASCII coming from an html interface; this module formats this question according to a format predetermined in practice in a file usable by the linguistic analysis module 11.

This module 100 delivers to an output c the question formatted in the desired format. This module 100 receives at another input b the data extracted by the analysis module on the typology of the question. This data includes keywords and other characteristic data. The module 100 makes it possible to deliver to another output e keywords. The other characteristic data are in the form of computer files and are stored in the data memory of the linguistic module 110.

The keywords are transmitted to the search engine 12 so that it can provide the intermediate response as an output. This response is a candidate sentence containing the keywords. The 2000 response processing set includes:

a module 200 for receiving on an input has the intermediate response, formatting it according to a predetermined format, delivering to an output c the formatted response, and supplying another output with the final response in a format suitable for the transmission over the network; a linguistic analysis module 110, in practice, this is the same module as that which extracts the keywords and other data. This module 110 is capable of receiving at an input the formatted intermediate response, of verifying this response from the stored characteristic data other than the keywords, extracting the final response after this verification, and output the final response.

The data, other than the keywords, are stored in a data memory 1200 by the operating program of the analysis module 110. FIG. 1 shows a unit 1100 representing the operating program memory of the module, which comprises a program allowing to pilot all the operations of this module and the launching of the execution of a linguistic analysis program.

The learning phase is also executed by this device and enriches the data memory.

Claims

1. A method of enriching semantic resources characterized in that it consists in using a real time electronic system of question answers to obtain an automatic enrichment of said resources and in that it comprises:

-an initialization phase including:

- a learning phase comprising:

• a search for one or more re-formulations of the relation formulated in the pattern of the initial question, from the arguments known by this question and the answer or answers obtained,

2. Automatic enrichment method according to claim 1, characterized in that the search for a re-formulation comprises:

- returning known arguments to determine at least one other relationship between these arguments, - the search for documents providing at least one answer to this new question, the extraction of sentences containing said known arguments, from these documents, - a linguistic analysis of sentences to extract one or more relations of the paraphrase type from the relation present in the boss of the original question,

- the storage of the paraphrase type relations of the relation of the initial question.

3. Semantic resource enrichment system, characterized in that it is coupled to a real time electronic question and answer system to obtain an automatic enrichment of said resources and in that it comprises: means for implementing an initialization phase of these means comprising:

Means of searching for at least one response, coupled to one or more databases of texts formulated in natural language through one or more communication networks,

Means for extracting one or more responses from documents provided by the research means, means for implementing a learning phase comprising:

• means of finding one or more reformulations of the relationship formulated in the pattern of the initial question from the arguments known by this question and the answer (s) obtained,

• means for storing these re-formulations on a database coupled to one of the communication networks.

4. Automatic enrichment system according to claim 3, characterized in that the means for seeking a re-formulation comprise: - means for returning the arguments known by the initial question,

- means of searching for documents containing these arguments,

- means of extracting sentences containing known arguments, from these documents,

- means of linguistic analysis of the sentences to extract one or more relations of the paraphrase type from the relation present in the pattern of the initial question, - means of storage of the relations of the paraphrase type from the relation of the initial question.

5. Automatic enrichment system according to claim 4, characterized in that the reference means, the means for extracting sentences containing the known arguments, from these documents and the means for linguistic analysis of the sentences are produced by modules computer programs or processes capable of implementing the stated functions and in that the search means are carried out by a program called the search engine.