JP3566977B2 - Natural language processing apparatus and method - Google Patents

Abstract

PURPOSE:To enable natural and smooth interaction by implementing an interruption from the processor at need during a process for natural language input from a user, and starting interaction for pointing out ambiguity and an error and solving a problem and also giving response indicating the understanding of a complicated input. CONSTITUTION:The natural language processor consists of a language analysis part 1, an area knowledge storage part 2, a response contents determination part 3, a language generation part 4, and a talk management part 5; and the talk management part 5 controls the language analysis part 1, response contents determination part 3, and language generation part 4 according to the contents of analyzing operation state information, response determining operation state information, and generating operation state information outputted from the language analysis part 1, response contents determination part 3, an language generation part 4.

Description

となり、「挟み」を出力する場合には、差分のマイクロ・フィーチャーを明示して「固定する挟み」、「鋏」の場合には、「刃物の鋏」または「切断する鋏」といった、同音異義語との違いを最小限の付加情報によって、明確な言い換え候補の生成が可能となる。
【０２１９】
変更例３
本実施例は、音声対話装置の音声出力部分と捉えることが可能であり、図２６に示すように音声対話装置の音声出力部分を本実施例に置き換えることが可能である。
【０２２０】
このことによって、音声対話装置で、利用者からの音声入力の認識において、認識率が同程度の複数の認識候補が存在して、対話処理部において、利用者に対して問い直しを行なうことを決定し、実際に問い直しを行なう際、認識候補中で互いに類音関係にある候補について、言い換え生成部、言い換え処理部での処理を加えることで、類音関係にある認識候補を類音関係のない別の語句に言い換えして、質問することが可能である。
【０２２１】
動作例：百科事典データベースのインタフェース手段としての音声対話装置
利用者：「センザイについて知りたい」
「センザイ」の認識候補として、同音かつ同アクセントの「洗剤」と「前栽」と、発音距離が１の「戦災」が挙がる。装置は、これらの「洗剤」「前栽」「戦災」の内、利用者はどれを指しているのか問い直しを行なうことを決定する。
【０２２２】
そして、「洗剤」「前栽」「戦災」は類音関係なので、そのまま問い直しては利用者に誤解を与えやすいので、言い換え候補として各々「洗浄剤」「庭」「戦争災害」を生成選択し、これらの内で利用者はどれを指しているのかを問い直す。
【０２２３】
問い直し手段として、認識候補の内、一番確率が高いものを選択して問い直すとすると、
装置：「洗浄剤ですか？」
あるいは、各認識候補を並立させて問い直すとすると、
装置：「洗浄剤ですか、庭ですか、戦争災害ですか？」
という装置の問い直し出力が生成される。
【０２２４】
【発明の効果】
第１の発明の自然言語処理装置によれば、利用者からの自然言語入力の処理中に、必要に応じて処理装置からの割り込みを行なうことができる。これにより、自然で円滑で柔軟性を持った対話を行なうことができる。
【０２２５】
第２の発明の自然言語処理装置によれば、同音異義語あるいは聞き誤られやすい発音の単語を持つ単語は、誤認されない形への言い換え処理を施されることで、自然な音声出力を得ることができる等の実用上多大な効果が奏せられる。
【図面の簡単な説明】
【図１】第１の実施例の自然言語処理装置のブロック図である。
【図２】言語解析部１のブロック図である。
【図３】解析動作状況情報の例である。
【図４】領域知識記憶部２の内容の例である。
【図５】言語生成部４のブロック図である。
【図６】談話管理部５のブロック図である。
【図７】談話制御規則記憶部５ｂの内容の例である。
【図８】自然言語処理装置での対話処理例である。
【図９】例文の係り受けに関する曖昧性をもつ解析結果候補の例である。
【図１０】第２の実施例の規則音声合成装置１００のブロック図である。
【図１１】類音リスト１０９の例である。
【図１２】一般アクセントの表記例である。
【図１３】アクセント辞書１１０の例である。
【図１４】共起辞書１１１の例である。
【図１５】シソーラス１１２の例である。
【図１６】シソーラス１１２の他の例である
【図１７】辞書１１３の例である。
【図１８】類音検索部１０２と類音判定部１０３での類音の検索と判定の例である。
【図１９】類音検索部１０２の処理フローである。
【図２０】類音判定部１０３の処理フローである。
【図２１】言い換え生成部１０４の処理フローである。
【図２２】言い換え判定部１０５の処理フローである。
【図２３】言い換え対象リスト１１４の例である。
【図２４】言い換え候補リスト１１５の例である。
【図２５】言い換え候補リスト１１５からの削除の例である。
【図２６】音声対話装置の例のブロック図である。
【符号の説明】
１ 言語解析部
２ 領域知識記憶部
３ 応答内容決定部
４ 言語生成部
５ 談話管理部
１０１ 構文解析部
１０２ 類音検索部
１０３ 類音判定部
１０４ 言い換え候補生成部
１０５ 言い換え候補判定部
１０６ 音韻処理部
１０７ 合成器
１０８ 語調変化選択部
１０９ 類音リスト
１１０ アクセント辞書
１１１ 共起辞書
１１２ シソーラス
１１３ 辞書
１１４ 言い換え対象リスト
１１５ 言い換え候補リスト[0001]
[Industrial applications]
The present invention relates to a natural language processing device that interacts with a user in a natural language.
[0002]
[Prior art]
2. Description of the Related Art In recent years, with the development and spread of computer devices, opportunities for general users other than experts to use the computer devices have been increasing.
[0003]
However, the operation of the computer device requires specialized knowledge, and it is difficult for a general user to freely use the computer device.
[0004]
Under such a background, research and development of a natural language interactive device for using a computer device have been actively performed by using a natural language such as Japanese or English which a user usually uses.
[0005]
In such a natural language dialogue device, the contents of a user's request are understood through language analysis and semantic interpretation, a problem is solved for the request, a response is determined, and the response is used by generating a language. To the person.
[0006]
Also, in a natural language dialogue apparatus using speech, speech recognition is performed at the beginning of the above-described processing, and speech synthesis processing is performed at the last stage of the processing.
[0007]
However, the conventional natural language dialogue device has a problem that natural dialogue cannot be performed because the user and the dialogue device are assumed to speak alternately.
[0008]
For example, even if an error or inconsistency exists in part of the user's utterance, or if the user makes an utterance that exceeds the analysis capability of the dialogue device, the dialogue device recognizes these problems during the utterance. It was impossible to point out to the user.
[0009]
Therefore, there is a problem that the user cannot notice an error until the time when all the utterances are completed, and the utterance made becomes useless, resulting in an increase in the burden on the user.
[0010]
Conventionally, in Japanese Patent Application Laid-Open No. 3-248268, a "voice dialogue processing method" detects a signal level of an input voice, and when the signal level is continuously lower than a threshold for a predetermined time or when a vowel of the input voice is continuous for a predetermined time. In some cases, an interactive device sends a message. However, since this is a sending process of an answer only by (sound) signal processing, an appropriate interruption is not performed in consideration of the content of the utterance of the user.
[0011]
As described above, in the conventional natural language dialogue device, sufficient consideration has not been given to simultaneous utterances such as guesswork and involvement, and therefore, there has been a problem that natural and efficient dialogue cannot be performed. .
[0012]
In recent years, along with the improvement of the processing capability of the computer device, it has become important for humans to improve the operational convenience of the computer device, that is, to improve the user interface.
[0013]
Various means such as a keyboard, a mouse, and an electronic pen have been proposed and put to practical use as means of the user interface. Among them, the natural language means using voice is natural, easy to use, and very efficient for human beings. For this reason, there have been proposed a number of commands to a computer device by voice input and a computer output method by voice.
[0014]
The interactive device is a computer device that understands a user's input in a natural language, processes the input, and outputs a result in a natural language as a front end of a computer device such as a database. A control mechanism for smoothing the flow is attached.
[0015]
A speech dialogue device is a dialogue device that uses voices for input means and output means, and enhances the efficiency of information transmission, such as omission and pronoun reference, which are frequently used in speech languages, while accurately reducing ambiguity. It must have a function that can be processed.
[0016]
As a problem peculiar to information transmission by voice, there is a problem such as misunderstanding of a homonymous word or a word having a similar pronunciation.
[0017]
In recent spoken dialogue apparatuses, in understanding input, context information is used as a constraint to select a relevant candidate or to ask a user again. In addition, when the reliability of the recognition candidates is low, for example, a method of re-inquiring the location or requesting re-input of the entire text is adopted.
[0018]
However, there is no solution to how to output homonymous words or easily pronounced pronunciation words in the output of the re-translated question or in the general output so as not to cause misunderstanding.
[0019]
[Problems to be solved by the invention]
As described above, in the conventional natural language dialogue device, since the dialogue device and the user are assumed to alternately utter, the dialogue device cannot start a response until the user's utterance is completely finished.
[0020]
For example, even if an error or inconsistency exists in part of the user's utterance, or if the user makes an utterance that exceeds the analysis capability of the dialogue device, the dialogue device recognizes these problems during the utterance. Since it is impossible to point out to the user, the utterance made by the user is wasted, resulting in an increase in the burden on the user, and as a result, a natural, smooth, and flexible dialogue cannot be performed. was there.
[0021]
However, in human-to-human dialogue, it is important to give the natural language dialogue device the ability to handle simultaneous utterances, given that smooth and efficient dialogue is performed by simultaneous utterances such as hail and interruption. .
[0022]
The first invention has been made in consideration of such circumstances, and realizes appropriate interruption and presentation of a message from a processing device in a dialogue between a natural language processing device and a user. The present invention provides a natural language processing apparatus capable of performing a smooth and flexible dialog.
[0023]
Further, as described above, conventionally, in the case of inputting / outputting a speech having a homonym or a word that is apt to be misunderstood, particularly in output, a word having a homonym or a word that is apt to be misunderstood is not misunderstood. Alternatively, there has been a problem that the processing cannot be performed in other words in a form that hardly occurs.
[0024]
The second invention has been made in view of such circumstances, and whether or not a paraphrase is changed to a form that is not misunderstood or hardly perceived in accordance with the situation, and in which case, which deformation means is used. It is an object of the present invention to provide a natural language processing apparatus which enables switching between using a sentence and using the contents of a target sentence in an accurate and efficient manner.
[0025]
[Means for Solving the Problems]
The natural language processing apparatus according to the first invention sequentially analyzes a natural language input from a user to output syntactic information, prosody information, or analysis result information of semantic information, and to change the state of the analysis operation. Language analysis means for outputting the analysis operation status information to be represented, and response content information determined based on the content of the analysis result information from the language analysis means,The state of the decision operation until the analysis result information is received and the response content information is determinedResponse content determining means for outputting the response determination operation status information to be represented, and for the response content information from the response content determining means, converting the content into a natural language output and sequentially presenting it to the user, A language generation unit that outputs generation operation status information indicating a status of a generation operation of a natural language output, and according to at least one of the analysis operation status information, the response determination operation status information, and the generation operation status information, A discourse management means for controlling the language generation means to perform simultaneous utterances such as an answer to the user's natural language input, or an interruption, wherein the language analysis means is generated in the process of analyzing the natural language input. Analysis operation status information including information on ambiguity is output in accordance with the number of candidates having ambiguity in the analysis result, and the discourse management unit outputs the analysis operation status. At the time of receiving the broadcast, by outputting a natural language for realizing the interrupt consisting interjection through the language generation means, responsive to nod to the utterance of the user, or is obtained by the so interrupts. In addition, language analysis that outputs analysis result information of syntactic information, prosodic information or semantic information by sequentially analyzing natural language input from a user, and outputs analysis operation status information indicating the status of analysis operation Means and response content information for the analysis result information from the language analysis means are determined according to the content thereof,The state of the decision operation until the analysis result information is received and the response content information is determinedResponse content determining means for outputting the response determination operation status information to be represented, and for the response content information from the response content determining means, converting the content into a natural language output and sequentially presenting it to the user, A language generation unit that outputs generation operation status information indicating a status of a generation operation of a natural language output, and according to at least one of the analysis operation status information, the response determination operation status information, and the generation operation status information, Discourse management means for controlling the language generation means to make simultaneous utterances such as a response to user's natural language input or interruption, and area knowledge storage for holding area knowledge information which is knowledge related to the content of the dialogue The language analysis means sequentially compares the obtained analysis result information with the area knowledge information recorded in the area knowledge storage means. When a contradiction occurs between the two, the discourse management means outputs the analyzed operation status information including the contradictory content. In this case, a natural language output that realizes the above is output to interrupt the user's utterance.
[0026]
According to a second aspect of the present invention, there is provided a natural language processing apparatus comprising: input means for inputting natural language information; and a similar sound for detecting, from the natural language information, a similar sound phrase having a phonetically similar or similar pronunciation. Detecting means;The paraphrase candidate of this detected similar phrase isA paraphrase candidate generating means for generating;In the case where the detected similar sound phrases are plural, the paraphrase candidates having a similar sound relation to each other among the paraphrase candidates of the respective similar sound phrases are deleted, and the non-deleted paraphrase candidates are referred to as the respective similar sound phrases. Paraphrase to replace eachAnd candidate determination means.
[0027]
[Operation]
An operation state of the natural language processing device of the first invention will be described.
[0028]
The language analysis means receives natural language input such as voice or character information from a user and sequentially analyzes the received portions to output syntactic information, prosodic information or semantic information analysis result information, and perform an analysis operation. Outputs analysis operation status information indicating the status of.
[0029]
The response content determination means receives the analysis result information from the language analysis means, determines response content information based on the content, and outputs response determination operation status information indicating the status of the response determination operation.
[0030]
The language generation means receives the response content information from the response content determination means, converts the content into a natural language output consisting of voice or natural language, sequentially presents the output to the user, and generates a generation operation indicating the status of the generation operation. Output status information.
[0031]
The discourse management means, according to at least one of the analysis operation status information, the response operation status information, and the generation operation status information output from the language analysis unit, the response content determination unit, or the language generation unit, At least one operation of the generating means is a simultaneous utterance such as a response to a user's natural language input or an interrupt.
[0032]
A case where the natural language processing apparatus of the first invention is further embodied will be described using an actual example.
[0033]
{Circle around (1)} The linguistic analysis means outputs analytic operation status information including information on ambiguity according to the number of ambiguity candidates in the analysis result generated in the process of analyzing the natural language input, and the discourse management means Upon receiving the analysis operation status information, the language generating means outputs a natural language output for realizing an interrupt made up of an interjection or the like, so that a response to the user's utterance or an interrupt is generated. If you set it to:
[0034]
A natural language input such as "Beautiful black eyes with long long hair ..." is made by the user, and in the process of interpreting this input, a large number of qualified phrase candidates to which the modifier "black" is related are generated. When the number of ambiguities exceeds the processing capability of the processing device, the user's utterance is interrupted at a necessary point in time without waiting for the user's utterance to end using interjections, etc. Initiate a disambiguation dialog following an event or interrupt. In addition, a natural language input such as "long black eyes" is input from the user, and in the process of analyzing the input, there are many candidates for the modified phrase to which the modifier "black" is related. When the number of ambiguities was generated and the number of ambiguities approached the processing capacity of the processing unit, the processing unit could not understand all of the user's input, but for the time being, hold on to this point and talk to the user. When the ambiguity is resolved, for example, by outputting an interjection such as "Umm" to indicate that you want to proceed, or after the user has input from the user, for example, "Women are ..." Then, the user is presented with an interjection such as "Eh", which expresses that the processing device has understood the user's input.
[0035]
As a result, the burden on the user caused by continuing useless utterances can be eliminated.
[0036]
{Circle around (2)} having area knowledge storage means for holding area knowledge information which is knowledge related to the content of the dialogue, wherein the language analysis means stores the obtained analysis result information and the area recorded in the area knowledge storage means; The comparison with the knowledge information is sequentially performed, and when inconsistency occurs between the two, the analysis output information including the inconsistent content is output as the analysis operation status information.When the discourse management unit receives the analysis operation status information, The following is a case where a natural language output for realizing an interrupt made up of an interjection or the like is output through the language generating means to interrupt the user's utterance.
[0037]
The user is inputting "Show the number of employees working at the branch of Company A in the United States", and according to the knowledge possessed by the processing equipment, there is no branch of Company A in the United States When there is a problem in dialog communication such as whether there is an error in the premise of the user's utterance or whether there is an error in the recognition result of the processing device for the user's input, Before the user's utterance ends, the user's utterance is interrupted using interjections, if necessary.
[0038]
As a result, the burden on the user caused by continuing useless utterances can be eliminated.
[0039]
{Circle around (3)} The discourse management means, when receiving analysis operation status information indicating that the user has accepted the interrupt from the language analysis means, analyzes the natural language input from the user who caused the interrupt. When the dialogue for eliminating the ambiguity of the language and the inconsistency in the meaning content of the user's natural language input is performed as follows.
[0040]
Subsequent to the interruption of (2), a dialog for activating an error in the premise of the user's utterance or an error in the recognition result of the processing device in response to the user's input is activated and solved.
[0041]
Thereby, a misunderstanding between the user and the processing device can be eliminated, and a flexible and smooth conversation can be performed.
[0042]
{Circle around (4)} When the language analysis means interrupts the user by interjections or the like under the control of the discourse management means, the user interrupts the utterance or generates an interjection or the like during understanding. It detects whether or not the user has accepted the interrupt of the processing device, and outputs analysis operation status information indicating the result, and the discourse management means performs a dialog for ambiguity resolution and inconsistency resolution in accordance with the analysis operation status information. If the system is controlled to start, it will be as follows.
[0043]
If an interrupt is made by an interjection such as "Ao", for example, if the user responds to "Yes" or interrupts the utterance that was being made, and receives an interrupt from the processing device, the interrupt is issued. If the processing is continued, for example, if the utterance that the user was performing is continued, the interrupt processing is stopped, and the processing after the interrupt is performed based on the status of the user's response to the interrupt from the processing device. Control.
[0044]
Thus, the dialog for ambiguity resolution and inconsistency resolution can be activated only when the user interrupts the utterance or utters a verbal or the like during the understanding to accept the interruption of the processing device.
[0045]
{Circle around (5)} Interruptable position feature storage means which stores in advance information about the end position of the utterance in the natural language dialogue, the syntactic feature of the partner utterance immediately before the interruptible position, or the prosodic feature of the partner utterance. The discourse management means, when an interrupt is required, compares the analysis result information obtained from the language analysis means with the contents of the interruptable position feature storage means to detect an interruptible position The following is a case where control is performed such that an interrupt is made by an interjection or the like through the language generating means at the time.
[0046]
When the processing device interrupts the user's utterance, the utterance performed by the user and the prosodic features such as the honorific expression at the end of the sentence, such as “-is,” and the pause (silence interval) For example, an interrupt is performed at an appropriate position by performing an interrupt based on comparison of information on syntactic or prosodic features at an interruptable position prepared in advance.
[0047]
This makes it possible to determine the end position of the utterance or the position where interruption is possible in the natural language dialogue, and to apply an interruption at an appropriate timing.
[0048]
{Circle around (6)} A dialog history storage means for holding a history of dialogues with the user is provided. From the contents of the dialogue history storage means, the position at which the utterance change occurs or the silent section position is determined as the utterance end position or interrupt. The interruptable position extracting means for extracting as a possible position, extracting a syntactic characteristic or a prosodic characteristic immediately before the position, and additionally recording the characteristic in the interruptable position characteristic storage means. If you set it to:
[0049]
By analyzing the history of the dialogue between the processing device and the user, for example, expressions such as “How about? ↑ (rise of intonation: increase of fundamental frequency of voice)” or “is it ぉImmediately after the user has made such an expression, alternation of the utterance, the fact that the executed interrupt tends to be accepted, and the like are extracted, and information on the interruptable position is extracted.
[0050]
As a result, practically significant effects can be obtained, such as a flexible interrupt that can be performed according to the utterance habit of each user.
[0051]
{Circle around (7)} The interruptable position feature storage unit holds information on the reliability of the interruptable position information, and the language analysis unit performs an interrupt based on the information recorded in the interruptable position feature storage unit. It is determined whether or not the user has accepted the interrupt by interrupting the utterance or generating a spelling during the response to the interrupt, outputting analysis operation status information including the result, and the discourse management means The case where the reliability of the interruptable position information stored in the interruptable position feature storage means is adjusted according to the analysis operation status information is as follows.
[0052]
When the user interrupts the user's utterance from the processing device, the reliability of the information of syntactic or prosodic features at the used interruptible position is adjusted depending on whether the interrupt was accepted by the user. I do.
[0053]
As a result, practically significant effects can be obtained, such as a flexible interrupt that can be performed according to the utterance habit of each user.
[0054]
A natural language processing device according to the second invention will be described.
[0055]
A similar sound detecting means detects a similar sound phrase which is a word having the same pronunciation or similar pronunciation phonetically.
[0056]
The paraphrase candidate generating means generates, as paraphrase candidates, words having the same pronunciation or similar pronunciation phonetically as the similar sound words detected by the similar sound detection means.
[0057]
The paraphrase candidate determining unit determines a target paraphrase candidate from the paraphrase candidates generated by the paraphrase candidate generation unit, and replaces the similar sound phrase with the paraphrase candidate according to the substitution rule.
[0058]
This makes it possible to obtain a clear and efficient natural language output by performing a paraphrase process on a word having a homonym or a word having a pronunciation that is likely to be misunderstood so as not to be mistaken.
[0059]
【Example】
First embodiment
Hereinafter, an embodiment of the natural language processing apparatus of the first invention will be described with reference to FIGS.
[0060]
FIG. 1 shows the configuration of the natural language processing apparatus according to the present embodiment, which comprises a language analysis unit 1, an area knowledge storage unit 2, a response content determination unit 3, a language generation unit 4, and a discourse management unit 5.
[0061]
(1) Language analysis unit 1
The language analysis unit 1 receives natural language input information from a user by voice or the like, performs speech recognition, language analysis, semantic interpretation, and the like, and outputs analysis information representing the analysis result of the natural language input information. And outputs analysis operation status information indicating the operation status of the language analysis unit 1.
[0062]
In addition, the language analysis unit 1 analyzes not only a normal response from the user but also an analysis of simultaneous utterances such as a hail or an interruption from the user during the natural language output of the interactive system.
[0063]
FIG. 2 shows an example of the configuration of the language analysis unit 1.
[0064]
Voice input from the user is taken into the voice recognition unit 12 through the microphone 10.
[0065]
The speech recognition unit 12 performs speech recognition with reference to knowledge for speech recognition such as the phoneme dictionary 14. The analysis result including the intermediate result is output from the language analysis unit 1 as analysis result information, and information such as success or failure of the analysis operation is output from the language analysis unit 1 as analysis operation status information.
[0066]
The syntactic analysis unit 16 analyzes the output of the speech recognition unit 12 with reference to the language dictionary 18, the grammar rules 20, and the like. The syntactic analysis result and the status of the syntactic analysis operation are similarly output from the language analysis unit 1 as analysis result information and analysis operation status information, respectively.
[0067]
The semantic interpretation unit 22 analyzes the output of the syntactic analysis unit 16 with reference to the interpretation rules 24 and the like. Similarly, the semantic interpretation result and the state of the semantic interpretation operation are output from the language analysis unit 1 as analysis result information and analysis operation state information, respectively.
[0068]
3A to 3E show examples of the analysis operation status information output by the language analysis unit 1.
[0069]
(A) is output when all of the results of speech recognition, syntactic analysis, semantic interpretation and analysis of the natural language input from the user in the language analysis unit 1 and the consistency between the analysis result and the contents of the domain knowledge storage unit 2 are successful. An example is shown.
[0070]
(B) indicates that, in the syntactic analysis, the number of analysis result candidates having dependency ambiguity has become larger than a preset threshold T1.
[0071]
(C) indicates that, in response to the interruption to the utterance of the user performed by the processing device, a response indicating the acceptance is made.
[0072]
(D) indicates that the user responded to the question asked by the processing device to the user.
[0073]
(E) is an example of information used when this embodiment is extended, and details will be described later.
[0074]
(2) Area knowledge storage unit 2
The domain knowledge storage unit 2 records domain knowledge information referred to by the response content determination unit 3, the discourse management unit 5, and the like.
[0075]
FIG. 4 shows an example of the contents of the area knowledge storage unit 2.
[0076]
In each entry of the area knowledge storage unit 2, the knowledge content information A represents the area knowledge held by the processing device, and B is the storage address information.
[0077]
For example, in the entry of the storage address P11 in FIG. 4, the content of the knowledge content information A is “(part number word processor A JWP-XX)”, which means that the “part number” of “word processor A” is “JWP-XX”. Represents the domain knowledge that
[0078]
(3) Response content determination unit 3
The response content determination unit 3 determines the content of a response to the user based on the content of the analysis information output from the language analysis unit 1 and the domain knowledge information obtained from the domain knowledge storage unit 2, and determines the content of the response. In addition to outputting response content information representing the response, the response determination operation status information representing the operation status of the response content determination unit 3 is also output.
[0079]
(4) Language generation unit 4
Based on the response content information output from the response content determination unit 3, the language generation unit 4 generates a response to the user in a natural language by language generation and speech synthesis with reference to grammar rules, language dictionaries, phonological rules, and the like. In addition to the output, it also outputs generation operation status information indicating the operation status of the language generation unit 4.
[0080]
The language generating unit 4 presents a message to the user, interrupts the user, and the like. In this process, in the generation of uttered speech such as an interjection to be presented to express the utterance intention of the processing device, the intonation is used. Prosody such as accents and accents are appropriately controlled.
[0081]
FIG. 5 shows an example of the configuration of the language generation unit 4.
[0082]
The syntactic structure generation unit 26 receives the response content information output from the response content determination unit 3 and determines the syntax structure of the response sentence for realizing the content of the response content information with reference to the grammar rule 28. Further, the language generating unit 4 outputs information indicating the status of the operation such as the success / failure of the syntax structure generation as generation operation status information.
[0083]
The surface sentence generation unit 30 generates a response surface sentence by referring to the language dictionary 32 or the like based on the determined syntax structure of the response sentence and the response content information, and performs operations such as success / failure of the surface sentence generation. Information indicating the situation is output from the language generation unit 4 as generation operation situation information.
[0084]
Based on the determined response surface information and response content information, the voice synthesizer 34 generates a voice output as a response by voice synthesis with reference to the phonological rules 36 and the like, presents the response to the user, and presents the response voice output. The language generating unit 4 outputs information indicating an operation status such as success or failure of generation as generation operation status information.
[0085]
In addition, the language generating unit 4 also performs simultaneous utterances such as interruption of response output during output and interruption or a response to a user's input in natural language in accordance with control from the discourse management unit 5.
[0086]
(5) Discourse management unit 5
The discourse management unit 5 receives all or one of the analysis operation status information output from the language analysis unit 1, the response determination operation status information output from the response content determination unit 3, and the generation operation status information output from the language generation unit 4. Based on the section, all or a part of the language analysis section 1, the response content determination section 3, and the language generation section 4 are controlled.
[0087]
Since the discourse management unit 5 is a component that plays a central role in the present embodiment, it will be described in further detail below.
[0088]
FIG. 6 shows an example of the configuration of the discourse management unit 5.
[0089]
The discourse control section 5a has a discourse state register RS for holding a conversation state therein, receives analysis operation state information, generation operation state and response determination operation state information, and refers to the contents of the discourse control rule storage section 5b. Thus, the language analysis unit 1, the language generation unit 4, and the response content determination unit 3 are controlled through the language analysis unit control information, the language generation unit control information, and the response content determination unit control information.
[0090]
The discourse control rule storage unit 5b stores rules for determining the operation of the discourse control unit 5a prepared in advance, and is referred to by the discourse control unit 5a.
[0091]
FIG. 7 shows an example of the contents of the discourse control rule storage unit 5b.
[0092]
In each entry of the discourse control rule storage unit 5b, information on the discourse control rule corresponding to the entry Qi includes discourse state information A, attention target information B, situation condition information C, discourse state transition information D, and control content information E. Etc. and recorded. F is storage address information. The information of A, B, C, D, E, and F of the entry Qi is collectively referred to as a discourse control rule Ri.
[0093]
In the column of the discourse state information A, restrictions on the contents of the discourse state register RS when the discourse control rule Ri corresponding to the entry Qi is applied are recorded.
[0094]
In the column of the target information B, “analysis”, “generation”, or “response” representing the type of the operation situation information Xj that is a condition for applying the discourse control rule Ri is recorded. For example, in the entry of Q11 in FIG. 7, since the content of the target information B is "analysis", the condition when the discourse control rule Ri stored in Q11 is applied relates to the analysis operation status information. It has been shown.
[0095]
In the column of the situation condition information C, a condition relating to the content of the operation situation information Xj when the discourse control rule Ri is applied is recorded. For example, in the entry of Q11 in FIG. 7, since the content of the situation condition information C is "(ambiguity> threshold T1)", in order to apply the discourse control rule R11 corresponding to the entry Q11, It is shown that the contents of the operation status information must indicate that the ambiguity has exceeded the threshold value T2.
[0096]
In the column of the discourse state transition information D, the discourse state after applying the discourse control rule Ri corresponding to the entry Qi is recorded. For example, in the entry Q11, since the content of the discourse state transition information D is “S1”, the discourse state after applying the discourse control rule R11 corresponding to the entry Q11 is set to S1, and the discourse state register RS of the discourse control unit is set. Should be set to the value “S1”.
[0097]
In the column of control content information E, information on a control procedure to be performed when application of the corresponding discourse control rule Ri is determined is recorded.
[0098]
Examples of the contents of the individual procedures are shown in the following control procedure examples P1 to P7.
[0099]
Control procedure examples P1 to P7
P1 When the application of the discourse control rule in which the content of the control content information E is “procedure (aizu)” is determined
The discourse management unit 5 controls the language generation unit 4. For example, the user is presented with an interjection such as “Yes” or “Eh” which indicates consent.
[0100]
P2 When the discourse control rule in which the content of the control content information E is “procedure (interruption)” is determined to be applied
The discourse management unit 5 controls the language generation unit 4. For example, an interjection such as "Ah" or "Eh" is presented to the user, and an interruption to the user's natural language input utterance is executed.
[0101]
P3 When the content of the control content information E is “procedure (disambiguation dialogue)” and application of the discourse control rule is determined
The ambiguity is eliminated by performing the same processing as the dialog for eliminating the ambiguity of a natural language input having ambiguity with respect to the dependency (see Japanese Patent Application No. 5-10071).
[0102]
P4 When the content of the control content information E is “procedure (contradiction confirmation dialogue)”, the application of the discourse control rule is determined.
The language generation of response content information "(confirmation (true / false (K))") for confirming to the user whether the content K of the area knowledge storage unit 2 is inconsistent with the analysis result of the natural language input from the user. And sends the response content information. For example, by presenting to the user a question of natural language expression such as "Is not K?", The error of the user or the processing device is eliminated.
[0103]
P5 When application of the discourse control rule in which the content of the control content information E is “procedure (resumption of dialogue)” is determined
The control of a series of interruptions or discourse about the conversation is ended, and the control of the processing device is returned to the normal dialogue.
[0104]
P6 When the discourse control rule in which the content of the control content information E is “procedure (interruption interruption)” is determined to be applied
A series of interruptions, or interrupting the control of the discourse about the chat, is interrupted, and the control of the dialogue system is returned to the normal dialogue.
[0105]
Next, processing performed by the discourse control unit 5a based on the analysis operation status information, the generation operation status information, the response determination operation status information, and the contents of the discourse control rule storage unit 5b will be described.
[0106]
The basic operation of the discourse control unit 5a is performed based on the discourse control procedure A described below.
[0107]
Discourse control procedure A
A1 “S0” is recorded as an initial value in the discourse state register RS of the discourse control unit 5a.
[0108]
A2 The contents of the analysis operation status information, the response determination operation status information, and the generation operation status information output from the language analysis unit 1, the response content determination unit 3, and the language generation unit 4 as the dialog between the processing device and the user progresses. Based on the content of the discourse state register RS of the discourse control unit 5a, an applicable rule is searched from the discourse management rules recorded in the discourse control rule storage unit 5b.
[0109]
A3 As soon as the relevant rule is found, the rule is applied.
[0110]
A4 Go to A2.
[0111]
The above is the configuration of the present apparatus and its functions.
[0112]
Concrete example
The operation of the natural language processing device will be described in further detail with reference to the drawings, taking a dialogue information processing device that performs a natural language dialogue for a user's question based on a database and word processing relating to a word processor as an example.
[0113]
First, in order for the user to know "the model number of a portable word processor with a transfer type printer", the answer to the processing unit is "Please answer the model number of the portable word processor with a transfer type printer." A description will be given by taking the case of input as an example.
[0114]
FIG. 8 shows the contents of the utterance of the user (U) and the utterance of the processing device (S) at each time point (T0 to T9) as an example of the interactive processing in the processing device.
[0115]
T0 In accordance with A1 of the discourse control procedure A, the discourse status register RS of the discourse control unit 5a is initialized with the value “S0”.
[0116]
The utterance of U1 from the T1 user is started, and the language analysis unit 1 sequentially analyzes the utterance.
[0117]
However, the utterance of U1 that is making an utterance has an ambiguity regarding the dependency as shown in FIG.
[0118]
For this purpose, when the user utters, for example, up to the “to word processor” portion of the utterance U1, the number of candidates having ambiguity regarding dependency in the language analysis in the language analysis unit 1 is set in advance. It is assumed that the threshold value T1 has been exceeded and the analysis operation status information X1 as shown in FIG. 3B has been output. At this point, it is assumed that a candidate for an analysis result having ambiguity in the analysis is held in the language analysis unit 1.
[0119]
Analysis operation status information X1:
[Syntactic analysis: (ambiguity> threshold T1)]
T2 The discourse management unit 5 receives the analysis operation status information X1.
[0120]
The discourse management unit 5 compares the discourse state value “S0” recorded in the discourse information register RS of the discourse control unit 5a, the content of the analysis operation status information X1, and the content of the discourse management rule storage unit 5b.
[0121]
The content of the discourse state information A of the entry at the storage address Q21 in FIG. 7 is “S0”, the content of the attention target information B is “analysis”, and the content of the situation condition information C is “ambiguity> threshold T1”. This matches the content of the analysis operation status information X1 of interest.
[0122]
Therefore, it is detected that the discourse management rule R21 recorded at the storage address Q21 is applicable.
[0123]
Discourse management rule R21:
[S0, analysis, (ambiguity> threshold T1), S3, procedure (interrupt)]
T3 The discourse control unit 5a executes the process according to the control procedure example P2 when the content of the control content information E of the discourse management rule R21 is “procedure (interruption)”, and interrupts the user's utterance through the language generation unit 4. To achieve. For example, an interjection such as “Ah” of the utterance S1 is presented to the user, and the value “S3” is recorded in the discourse state register RS based on the discourse state transition information D of the discourse management rule R21.
[0124]
T4 In response to the interruption from the processing device at T3, the user indicates acceptance of the interruption of the processing device. For example, an interjection indicating an acknowledgment such as "Yes" of the utterance U2 is uttered, and this is detected by the language processing unit 1, and analysis operation status information X2 indicating acceptance of an interrupt as shown in FIG. 4C is output. .
[0125]
Analysis operation status information X2:
[Syntactic analysis: (interrupt receipt = T)]
T5 The discourse control rule storage unit 5b detects the discourse control rule R31 for the entry at the storage address Q31 in FIG. 7 based on the analysis operation status information X2 and the content "S3" of the discourse status register RS by the same processing as at the time T2. Is done.
[0126]
Discourse management rule R31:
[S3, analysis, (interrupt acceptance = T), S5, procedure (disambiguation dialog)]
T6 Since the value of the control content information E of the discourse control rule R31 is “procedure (disambiguation dialogue)” by the same processing as that at the time point T3, the dialogue is similar to the dialogue for dependency disambiguation according to the control procedure P3. (See JP-A-5-10071), and the value "S5" is recorded in the discourse state register RS based on the discourse control rule R2 applied.
[0127]
Example of utterance of question from processing device at time T6
[Is the "portable" person involved a "word processor"? "
In response to the question from the processing device at T7 and T6, an input indicating a response of the user's confirmation, such as "Yes", is made.
[0128]
The ambiguity is resolved by the dependency analysis that has occurred in the language analysis unit 1 through the dialogue for confirming the utterance S2 to the utterance U3.
[0129]
The language analysis unit 1 outputs analysis operation status information X3 as shown in FIG.
[0130]
The discourse control rule R51 stored at the storage address Q51 in FIG. 7 is detected by the same processing as at the time T2.
[0131]
By the same processing as at the time point T3, the control procedure P5 is started, and a series of interruptions or discourse about the quiz is ended, and the control is returned to the normal dialogue.
[0132]
Analysis operation status information X3:
[Syntactic analysis: (Confirmation completed)]
Discourse management rule R51:
[S5, analysis, (confirmation completed), S0, procedure (resumption of dialogue)]
T8 Returning to the normal dialogue, the following input from the user such as "show the model number of the word processor" is performed as follows, for example, by a method of performing an anaphoric solution by using the same noun that appears most recently. Note that parentheses indicate a dependency relationship.
[0133]
The expression "that word processor" was disambiguated by the above processing
"(((Transfer-type (printer)) (portable (word processor)))"
By analyzing that the user request indicated by a series of dialogues is, for example,
"(Request (output (model number (((transfer type (printer)))" (portable (word processor))))))
Described in a semantic expression such as
It is analyzed that the information means "information provision of the word processor A model number" indicated by the expression.
[0134]
T9 The contents of the storage addresses P11 to P22 in the example of the contents of the area knowledge storage unit 2 shown in FIG. 4 and the inference processing in the response content determination unit 3 determine that the response content is “JWP-XX” to be the response content. You.
[0135]
An output response such as “the model number is JWP-XX” is presented to the user through the language generation unit 4.
[0136]
According to the present device configured as described above, the following effects can be obtained.
[0137]
During processing of a natural language input from a user, an interrupt from the processing device can be performed as needed.
[0138]
During processing of natural language input from the user, if ambiguity that exceeds the analysis capability of the processing device occurs, interjections etc. are output immediately, thereby interrupting the user's utterance and causing unnecessary utterance. The burden on the user due to continuing can be eliminated.
[0139]
During processing of natural language input from the user, after an ambiguity exceeding the analysis capability of the processing device occurs and an interrupt is generated, a dialog for resolving the ambiguity is performed, so that the user and the processing device can communicate with each other. And a flexible and smooth dialogue can be conducted.
[0140]
When the processing device interrupts the user's utterance due to an interjection or the like, the user interrupts the utterance, or the user accepts the interrupt of the processing device by uttering the verbal or the like during the comprehension. Only in such a case can a dialog for disambiguation or inconsistency be invoked.
[0141]
The above embodiment monitors the number of ambiguities occurring in the analysis of the natural language input from the user in the language analysis unit 1, detects excessive ambiguity by comparing with a preset threshold, and utters the user. We are conducting a confirmation dialogue to interrupt ambiguity and to resolve ambiguity. However, instead of this, the following modifications can be made.
[0142]
Modification example 1
It can be extended so that it can be used to resolve errors and inconsistencies contained in the user's speech.
[0143]
For example, the response content determination unit 3 checks the consistency between the (partial) analysis result of the user's utterance obtained from the language analysis unit 1 and the information recorded in the area knowledge storage unit 2 as needed. When a contradiction occurs between the two, an interrupt process to the utterance of the user can be performed.
[0144]
Modification example 2
In checking the number of ambiguities in the analysis of the natural language input from the user, the second threshold value T2 and the third threshold value T3 are set so as to satisfy “threshold value T3 <threshold value T2 <threshold value T1”. In the discourse state S0, when the ambiguity becomes “threshold T2 <ambiguity <threshold T1”, the discourse state is set to S1, and the ambiguity is resolved by the analysis of the natural language input of the structure from the user afterwards. When gender = 0, or when the ambiguity becomes sufficiently small and “ambiguity <threshold T3”, a message such as “Eh” is presented to the user through the language generation unit 4.
[0145]
As a result, a complex natural language input that may cause ambiguity exceeding the analysis capability of the processing device is made by the user, and the processing device can analyze the content of the input by processing the subsequent input or the like. In this case, it is possible to eliminate anxiety about an increase in burden due to the user's restatement when the analysis fails.
[0146]
This function can be realized by the discourse control rules recorded at the storage addresses Q11 and Q12 in the example of the contents of the discourse control rule storage unit 5b in FIG.
[0147]
Modification 3
In the above-described embodiment, the presentation of the message by the control procedure P1 and the execution of the interruption of the control procedure P2 by the language generation unit 4 are performed immediately when the discourse control unit 5a determines the execution.
[0148]
By expanding this, words appearing immediately before the point at which interrupts can occur in natural language utterances (for example, words such as "-ne", the last particle that indicates the end of a sentence or a phrase, or words such as "a", which are the last sentence expression) It is prepared in advance in the interruptable position recording unit newly provided in the language analysis unit 1. When the discourse control unit 5a interrupts or requests for an utterance to be uttered, the language analysis unit 1 analyzes the utterance currently performed by the user and the information recorded in the interruptable position storage unit. The discourse management unit 5 outputs the analysis operation status information as shown in, for example, FIG. At this point, a message is sent through the language generator 4 or an interrupt is executed.
[0149]
This makes it possible to determine the end position of the utterance or the position where interruption is possible in a natural language dialogue and interrupt at an appropriate timing when an interrupt from the processing device is necessary for the user's utterance It is.
[0150]
Modification 4
7
It is possible to record the conversation history between the user and the processing device, extract the part of the utterance immediately before the utterance end position of the user from the recorded conversation history, and add it to the interruptable position recording unit. .
[0151]
Modification 5
A reliability point, which is a numerical value representing the reliability of the information recorded in the entry, is added to each entry of the interruptable position recording unit, and a response is made depending on whether or not the interrupt performed by the processing device has been received from the user. By adjusting the value of the reliability point to be performed and controlling the application of the information recorded in the interruptable position information storage unit according to the value of the reliability point, flexible interruption according to the utterance habit of each user can be performed. Can be performed.
[0152]
Second embodiment
Hereinafter, a rule speech synthesizer 100 as one embodiment of the natural language processing device of the second invention will be described with reference to FIGS.
[0153]
FIG. 10 is a configuration diagram of the rule speech synthesizer 100 of the present embodiment.
[0154]
The rule-based speech synthesizer 100 is a device that performs a syntax analysis of a text and applies speech information obtained as a result of applying the phoneme information according to rules prepared in advance as speech data to perform speech synthesis. That is, the rule speech synthesizer 100 includes the following.
[0155]
The syntax analysis unit 101 performs a morphological analysis and a syntactic analysis on the input sentence according to a grammar rule prepared in advance, and generates a syntax tree.
[0156]
The similar sound search unit 102 searches for a homonym or a noun whose pronunciation is easy to hear for each noun in the input sentence based on the accent pattern and a list of phonemes that are easy to hear.
[0157]
The similar sound determining unit 103 restricts the result of the similar sound determining unit 103 based on co-occurrence information with other words in the sentence.
[0158]
The paraphrase candidate generation unit 104 generates a paraphrase candidate of a noun having a homonym or a noun whose pronunciation is easy to hear.
[0159]
The paraphrase candidate determining unit 105 selects the paraphrase candidates for each noun and rewrites the syntax tree according to the selected paraphrase candidates so that the paraphrase candidates for different nouns do not have a misleading relationship.
[0160]
The phoneme processing unit 106 generates speech data by applying phoneme information to a syntax tree according to a prosody rule prepared in advance.
[0161]
The synthesizer 107 outputs the audio data generated by the phoneme processing unit 106.
[0162]
The tone change selection unit 108 causes the synthesizer 107 to generate an actual voice from the voice data and to select whether or not to change the tone of the part replaced by the paraphrase candidate.
[0163]
Here, the definition of words used in this embodiment, the rules for accents, and the like will be described.
[0164]
(1) Pronunciation distance between words
The “pronunciation distance between words” is defined as the number of times a certain word is converted to another word by exchanging phonemes included in the similar sound list 109.
[0165]
For example, if “detergent” is decomposed into phonemes, it becomes “sensai”, but “sensai”, in which one sound is replaced by “z → s” in the similar sound list, has a pronunciation distance of one.
[0166]
As shown in FIG. 11, the similar sound list 109 is a list of certain sounds and sounds whose sounds are likely to change. For example, “s → z” means that the consonant “s” becomes “z” (“sa (sa)” becomes “za (za)”, and “pj → hj” becomes “pj” becomes “hj” (“ "Pju" is easily misunderstood by "hju".
[0167]
▲ 2 ▼ “Synonyms”, “Synonyms Nouns”
If the accent pattern is the same between any two words, and the pronunciation distance is equal to or less than a predetermined threshold, the two words are said to have a similar sound relationship, and one word is referred to as a "similar word" of the other word. , And a noun is defined as a "similar noun."
[0168]
(3) Accent rules
The accent is represented as shown in FIG. In the national polity, "kutai" is pronounced high, and "significance" indicates that it is lowered by the sound following "i".
[0169]
As shown in FIG. 13, the accent dictionary 110 is a list of words and accent patterns, and the accent pattern of “Kokutai” is “Ko H Kutai”, which is higher than the sound after “Ko”. This means that the accent pattern “I Lgi” of “significance” is lower than the sound next to “i”.
[0170]
▲ 4 ▼ Others
The selection as to whether or not to change the tone of the voice output of the portion where the paraphrase candidate has been replaced is given in advance by the user as an input. The state is held in the tone change selection unit 108. If it is selected to change the tone, the paraphrase determination unit adds markers to the start and end points of the paraphrase to indicate the start and end of the paraphrase. The phonemic processing unit 106 changes the tone based on the marker added by the paraphrase determination unit and outputs the changed tone.
[0171]
Next, the function of each component will be described in order.
[0172]
(1) Syntax analyzer 101
The syntax analysis unit 101 receives an input sentence input from a keyboard or an interactive device, performs morphological analysis and syntactic analysis according to a grammar rule prepared in advance, and generates a syntax tree as an output.
[0173]
When the input sentence is input, a word tone change selection input as to whether or not to change the output voice of the paraphrased part is also input, and the state is held by the word tone change selection unit 108.
[0174]
(2) Sound search unit 102
As shown in the processing flow of FIG. 19, the similar sound search unit 102 refers to the similar sound list 109, the dictionary 113, and the accent dictionary 110 for each noun in the syntax tree, and searches for the presence or absence of a similar sound noun for each noun. I do. In other words, for each noun, a sequence of sounds in which each sound in the noun is replaced by a number of sounds within the threshold value that is easily misunderstood included in the similar sound list 109 is included as a noun in the dictionary, and the accent pattern is not included. In the same case (that is, in the case of a similar noun), a pair of the corresponding noun and the similar noun is registered in the paraphrase target list 114. At this time, it is determined that nouns having different accent patterns or nouns whose pronunciation distance is equal to or larger than the threshold are not erroneously recognized even if the pronunciation is the same.
[0175]
As shown in FIG. 17, the dictionary 113 stores the frequency of appearance of parts of speech, synonyms, synonyms, and words of the same sound. For example, “detergent” has a reading “senzai”, a part of speech “common noun”, is synonymous with “detergent”, and has an appearance frequency of 0.7.
[0176]
The paraphrase target list 114 stores a pair of a noun in the input sentence having a synonymous noun and a similar sound noun, and the subsequent processing is performed on the noun included in the paraphrase target list 114. .
[0177]
(3) Similar sound determination unit 103
The similar sound determination unit 103 receives the output of the similar sound search unit 102, and refers to the co-occurrence dictionary 111 for the noun, as shown in the processing flow of FIG. If it exists, it is determined that there is no misunderstanding from the co-occurrence relationship, and the pair of the noun and the similar noun is deleted from the paraphrase target list 114.
[0178]
As shown in FIG. 14, the co-occurrence dictionary 111 indicates a co-occurrence relationship between a certain word and a certain word, that is, a relationship that is likely to appear at the same time. For example, “diet” and “country-country”, “stadium” and “national body”, “detergent” and “washing”, “pre-plant” and “garden shop” have a co-occurrence relationship.
[0179]
However, if the appearance frequency of the noun with reference to the dictionary 113 is equal to or lower than the threshold, the noun is determined to need to be modified, and is not deleted from the paraphrase target list.
[0180]
(4) Paraphrase candidate generation unit 104
As shown in the processing flow of FIG. 21, the paraphrase candidate generation unit 104 paraphrases the noun in the input sentence included in the paraphrase target list 114 by the following processing for the preset maximum number of generations of the paraphrase candidate. Generate candidates. First, referring to the part of speech information of the dictionary for the noun, a paraphrase candidate generation process is performed in each of the following three cases. The pair of the original noun and the paraphrase candidate is registered in the paraphrase candidate list 115.
[0181]
1. Abbreviations that omit some nouns
With reference to the dictionary, the noun (formal name or the like) before the noun is omitted is a paraphrase candidate. Alternatively, a combination of the noun before the abbreviation before or after the noun is a paraphrase candidate.
[0182]
For abbreviations of foreign nouns such as English, in addition to the above-described processing, the Japanese translation of the noun before abbreviation, abbreviations in the Japanese translation, or parallel writing similar to the above-described processing are paraphrase candidates.
[0183]
"National Athletic Meet" → "National Athletic Meet"
"Country" → "Diet Measures Committee"
"JAS" → "Japan Food Standard" "Japan Air System"
"WASP" → "Protestant Anglo-Saxon Caucasians"
2. For proper nouns
The proper noun is a noun representing a name limited to one such as a personal name, a company name, and a place name. The paraphrase generation means for such proper noun refers to the thesaurus 112 and adds a word located above and below the proper noun as a paraphrase candidate.
[0184]
As shown in FIG. 15, the thesaurus 112 expresses a conceptual hierarchical relationship between words. In the case of FIG. 15, the thesaurus 112 expresses information relating to a connection between place names such as a region, a prefecture, and a municipal. In this figure, "Kitamachi", "Kitacho" and "Kidacho" are "Kitamachi", "Minamicho" "Minamicho" and "Minamicho" are "Minamimachi", "Nakamachi" and "Nakamachi" are "Nakamachi" And the same reading.
[0185]
Here, when considering "Kita-cho" in the figure, a form to which information such as "B city" and "X prefecture" located at a higher level is added, a form such as "K city of B city", and "Kita town of X prefecture B city" Is a paraphrase candidate.
[0186]
"Kitamachi" → "Kitamachi A City" "Kitamachi A City X Prefecture"
"Kita-cho" → "Kita-cho, B city" "Kita-cho, X city B city"
"Minamicho" → "Minamicho, City A" "Minamicho, City A"
"Minami Town" → "Minami Town in C City" "Minami Town in C City in Y Prefecture"
3. For common nouns
Common nouns are used for solids belonging to one class. One of the paraphrase candidate generation means for common nouns is a method of referring to the same thesaurus 112 as the means for proper nouns. For example, it is assumed that a thesaurus for “Ko” and “Performance” having the same reading and the same accent of “Ko H” is prepared as shown in FIG. The paraphrase candidate at this time is
"Park" → "park of place"
"Performance" → "Performance of event"
Is generated as follows.
[0187]
Another means for generating a paraphrase candidate for a common noun is to refer to the dictionary 113 and set a synonym / synonym of the word as a paraphrase candidate.
[0188]
"Detergent" → "soap"
"Pre-planting" → "Gardening"
(5) Paraphrase candidate determination unit 105
The paraphrase candidate determination unit 105 refers to the paraphrase candidate list 115 and paraphrases these paraphrase candidates if the paraphrase candidates for the different nouns are the same or have a similar sound relationship, as shown in the processing flow of FIG. It is deleted from the candidate list 115.
[0189]
Then, the following rules are applied in order to replace the noun in the syntax tree with a paraphrase candidate.
[0190]
1. The replacement process is performed from left to right, that is, from the beginning to the end.
[0191]
2. The abbreviation is replaced only once at the beginning, and is not replaced when it appears after the second time. If the abbreviation appears only once, it is replaced with a paraphrase candidate with a formal name, and if it appears more than once, it is replaced with a paraphrase candidate with a formal name.
[0192]
3. The proper noun is replaced only once at the first time, and is not replaced even if it appears after the second time.
[0193]
4. The replacement of the common noun with the additional information type paraphrase candidate is performed only once at the first time, and is not replaced even if it appears after the second time.
[0194]
5. Substitution of common nouns with synonyms or synonyms is performed for all occurrences.
[0195]
6. When the mode for changing the tone of the replacement part is selected, a marker indicating the replacement part is added before and after the paraphrase candidate.
[0196]
(6) Phoneme processing unit 106
The phoneme processing unit 106 converts the syntax tree, which has been replaced with the paraphrase candidate by the paraphrase candidate determination unit 115, into speech data according to the phoneme rules. At this time, if the mode of changing the tone of the replacement part is selected, the replacement part is cut out according to the marker added by the replacement operation of the paraphrase candidate determination unit, and if the attached word immediately follows, Is converted into voice data so that the tone of the replacement part changes as a tone change section including the attached word.
[0197]
(7) Synthesizer 107
The synthesizer 107 outputs the audio data as actual audio by using, for example, a formant synthesizer given in advance.
[0198]
The above is the configuration of the present apparatus and its functions.
[0199]
Concrete example
Next, the operation of the present embodiment will be described based on a specific input example. It is assumed that the input sentence in the specific example is "Today, the National Assembly has discussed the significance of the national polity."
[0200]
(1) Syntax analyzer 101
The syntax analyzer 101 performs a morphological analysis and a syntactic analysis on the input sentence, and outputs a syntax tree. The syntax tree as a result of the syntax analysis unit 101 greatly changes depending on the content of the grammar used by the syntax analysis unit 101. In any case, the nouns in the sentence to be paraphrased are “today”, “diet” It is analyzed and output if there are five types: "national body", "significance", and "discussion".
[0201]
(2) Sound search unit 102
This syntax tree is input to the similar sound search unit 102. As shown in FIG. 18, the similar sound search unit 102 searches for a similar sound noun as “national body” and “country pair” as homophonic accents, and “significance” and “opposition” as homophonic accents. It is registered in the list 114.
[0202]
Also, for example, when the threshold of the pronunciation distance related to similar sounds is 1, “Washing (Sensai)” and “Sensai” with the same accent and pronunciation distance as “Detergent (Sensai)” are the synonyms of “Detergent”. Searched. "Appetizer" is not a synonymous noun because the pronunciation distance is 2. “Tenchi”, “Tenchi” and “Tenchi” and “Tenchi” have the same sound, but are different from each other in their accent patterns, so they are not synonyms.
[0203]
FIG. 23 shows the contents of the paraphrase target list 114 at this time.
[0204]
(3) Similar sound determination unit 103
The similar sound determination unit 103 searches for co-occurrence words in the sentence regarding the "national body" and the "significance" in which the similar sound noun is present in the similar sound search unit 102. In this example, “national polity” and “significance” do not have a co-occurrence relationship in the text. When the threshold of the appearance frequency is 0.5, it is assumed that the appearance frequency of “national body” is 0.4 and the appearance frequency of “significance” is 0.6 from the dictionary information. In this case, since neither of them has a co-occurrence relationship, no search is performed on the paraphrase target list 114, and the next paraphrase candidate generation unit 104 generates a paraphrase candidate.
[0205]
(4) Paraphrase candidate generation unit 104
The paraphrase candidate generation unit 104 performs paraphrase candidate generation processing according to the type of the noun and the similar phonetic noun. In this example,
"National Athletic Meet" → "National Athletic Meet" "National Athletic Meet (National Athletic Meet)" "National Athletic Meet (National Athletic Meet)"
“Country” → “Diet Task Force” “Diet Task Force (country)” “Country Action (Diet Task Force)”
"Meaning" → "Meaning" "Value"
"Objection" → "objection" "appeal" "opposition"
Are generated. FIG. 24 shows the state of the paraphrase candidate list 115 at this time.
[0206]
(5) Paraphrase candidate determination unit 105
The paraphrase candidate determining unit 105 deletes paraphrase candidates having a similarity to each other from among the paraphrase candidates for different words.
[0207]
For example, "meaning" and "dissatisfaction" each have a similar noun,
"Meaning" → "Meaning"
“Dissatisfied” → “Objection”
When the paraphrase candidate is generated, the paraphrase candidates “significance” and “objection” are deleted from the paraphrase candidate list 115 because they have a similar sound relation with the same sound and the same accent. This is shown in FIG.
[0208]
Then, the noun in the syntax tree is replaced with a paraphrase candidate according to the replacement rule. In this example, since “national athletic” appears only once, “national athletic meet” is selected, and “meaning”, which is the head of synonyms and synonyms, is selected as a paraphrase candidate for “meaning”, and replacement is performed. When the mode for changing the tone of the replacement part is selected, a marker '△' indicating the head and end of the replacement part is added to each of the head and the tail of the replacement part.
[0209]
As a result of the above, after the paraphrase / replacement operation, the input sentence is paraphrased as "the debate has been held today in the Diet regarding the {meaning} of the {National Sports Festival}}. '△' indicates the beginning and end of the replacement part.
[0210]
(6) Phoneme processing unit 106
The phoneme processing unit 106 converts the speech data into speech data in accordance with the phoneme rules. The adjunct immediately after the replacement part is included in the tone change part, and the input sentence is discussed in the Diet today. Has been performed "(" ... "indicates a tone change portion) and is converted into voice data.
[0211]
(7) Synthesizer 107
The speech data is converted into actual speech by the synthesizer 107.
[0212]
Thus, according to the present device configured as described above, a plurality of nouns having the same or similar pronunciation in pronunciation and having the same or similar pronunciation are used to convert a noun that is easily misunderstood to information addition or a synonymous synonym. Is performed, regular speech synthesis that generates a speech output that is less likely to be misunderstood becomes possible.
[0213]
The internal structure of each dictionary and specific numerical values of each threshold value are not limited to the above examples.
[0214]
Modification example 1
For example, when it is desired to explicitly output “Mr. Sato” to which “Mr. Sato” points, the rule-based speech synthesizer 100 outputs information about which “Mr. Sato” points to from a sentence that is an input of the device. I couldn't get it.
[0215]
However, in the embodiment as a spoken dialogue device, the knowledge expression for output from the spoken dialogue device is given as an input of the speech output part, and the knowledge information of the spoken dialogue device can be referred to. It is possible to know which "Mr. Sato" is pointing, and for example, it is possible to paraphrase "Mr. Sato in Kobe".
[0216]
Modification example 2
Each word is represented by a set of finer concepts (hereinafter referred to as micro-features), and in the generation of paraphrase candidates, micro-features that make up each of the phonetic nouns are compared and information that clarifies the difference By adding and paraphrasing to synonyms and synonyms, the efficiency of paraphrase candidate generation can be improved, and the method can be changed to a method that can generate more accurate paraphrases.
[0217]
For example, in the same reading of "scissors", "sandwich (document pinching)" and "scissors (cutting scissors)" are described by the following set of micro features.
[0218]
[Table 1]

When "clamp" is output, the micro feature of the difference is clearly specified and "clamp" is fixed. In the case of "scissors", homonyms such as "knife scissors" or "cutting scissors" are used. A clear paraphrase candidate can be generated with additional information that minimizes the difference from a word.
[0219]
Modification 3
This embodiment can be regarded as a voice output portion of the voice interaction device, and the voice output portion of the voice interaction device can be replaced with this embodiment as shown in FIG.
[0220]
As a result, in the spoken dialogue apparatus, when recognizing the voice input from the user, when there are a plurality of recognition candidates having the same recognition rate, it is determined that the dialogue processing unit asks the user again. Then, when actually re-inquiring, for the candidates having a similar sound relation to each other among the recognition candidates, the processing in the paraphrase generation unit and the paraphrase processing unit is performed, so that the recognition candidates having the similar sound relation do not have the similar sound relation. In other words, it is possible to ask a question.
[0221]
Example of operation: Spoken dialogue device as interface means for encyclopedia database
User: “I want to know about Senzai”
“Senzai” recognition candidates include “detergent” and “maizeki” with the same sound and accent, and “war damage” with a pronunciation distance of 1. The apparatus decides to re-inquire which of these “detergent”, “pre-plant”, and “war disaster” the user refers to.
[0222]
And since "detergent", "planting" and "war disaster" are related sounds, it is easy to misunderstand the user if asked as it is, so we generate and select "cleaning agent", "garden" and "war disaster" as paraphrase candidates, The user asks which of these points he points to.
[0223]
As a re-querying method, if the candidate with the highest probability among the recognition candidates is selected and re-queryed,
Equipment: "Are you a cleaning agent?"
Or, if you ask each recognition candidate side by side and ask again,
Equipment: "Are you a cleaning agent, a garden, a war disaster?"
A query output of the device is generated.
[0224]
【The invention's effect】
According to the natural language processing device of the first invention, an interrupt from the processing device can be performed as necessary during processing of a natural language input from a user. This allows for a natural, smooth and flexible dialogue.
[0225]
According to the natural language processing apparatus of the second invention, a word having a homonym or a word having a pronunciation that is easily misunderstood is subjected to a paraphrase processing to a form that is not misunderstood, thereby obtaining a natural voice output. A great effect can be obtained in practical use, for example.
[Brief description of the drawings]
FIG. 1 is a block diagram of a natural language processing apparatus according to a first embodiment.
FIG. 2 is a block diagram of a language analysis unit 1;
FIG. 3 is an example of analysis operation status information.
FIG. 4 is an example of contents of an area knowledge storage unit 2;
FIG. 5 is a block diagram of a language generation unit 4.
6 is a block diagram of the discourse management unit 5. FIG.
FIG. 7 is an example of the contents of a discourse control rule storage unit 5b.
FIG. 8 is an example of an interaction process in the natural language processing device.
FIG. 9 is an example of an analysis result candidate having ambiguity regarding the dependency of an example sentence.
FIG. 10 is a block diagram of a rule speech synthesizer 100 according to a second embodiment.
11 is an example of a similar sound list 109. FIG.
FIG. 12 is a notation example of a general accent.
FIG. 13 is an example of an accent dictionary 110.
FIG. 14 is an example of a co-occurrence dictionary 111.
FIG. 15 is an example of a thesaurus 112.
FIG. 16 shows another example of the thesaurus 112.
FIG. 17 is an example of a dictionary 113.
18 shows an example of searching and determining a similar sound in a similar sound searching unit 102 and a similar sound determining unit 103. FIG.
FIG. 19 is a processing flow of a similar sound search unit 102.
20 is a processing flow of a similar sound determination unit 103. FIG.
21 is a processing flow of the paraphrase generation unit 104. FIG.
FIG. 22 is a processing flow of a paraphrase determination unit 105;
FIG. 23 is an example of a paraphrase target list 114.
24 is an example of a paraphrase candidate list 115. FIG.
25 shows an example of deletion from the paraphrase candidate list 115. FIG.
FIG. 26 is a block diagram of an example of a voice interaction device.
[Explanation of symbols]
1 Language analysis department
2 Area knowledge storage
3 Response content determination unit
4 Language generator
5 Discourse Management Department
101 Syntax analysis unit
102 Sound Searching Unit
103 Like sound judgment unit
104 Paraphrase candidate generation unit
105 Paraphrase candidate judging unit
106 phoneme processing unit
107 synthesizer
108 Tone change selector
109 Sound List
110 Accent Dictionary
111 Co-occurrence Dictionary
112 Thesaurus
113 dictionaries
114 Paraphrase List
115 Paraphrase candidate list

Claims (14)

A language analyzing means for sequentially analyzing a natural language input from a user to output analysis result information of syntactic information, prosodic information or semantic information, and outputting analysis operation status information indicating the status of the analysis operation; ,
In response to the analysis result information from the linguistic analysis means, response content information is determined according to the content thereof, and response determination operation status information indicating the status of the determination operation from receiving the analysis result information to determining the response content information Response content determining means for outputting
In response to the response content information from the response content determination means, the content is converted into a natural language output and sequentially presented to the user, and generation operation status information indicating the status of the natural language output generation operation by the conversion is generated. A language generating means for outputting,
According to at least one of the analysis operation status information, the response determination operation status information, and the generation operation status information, a simultaneous utterance such as a response to a user's natural language input or an interrupt is given to the language generation unit. Having discourse management means for controlling to be performed,
The language analysis means,
According to the number of candidates having ambiguity in the analysis result generated in the process of analyzing the natural language input, a process of outputting analysis operation status information including information on ambiguity is performed ,
The discourse management means,
At the time of receiving the analysis operation status information, by outputting a natural language for realizing an interrupt composed of interjections through the language generating means, a response to the user's utterance or a process of interrupting is performed. A natural language processing apparatus characterized in that: A language analyzing means for sequentially analyzing a natural language input from a user to output analysis result information of syntactic information, prosodic information or semantic information, and outputting analysis operation status information indicating the status of the analysis operation; ,
In response to the analysis result information from the linguistic analysis means, response content information is determined according to the content thereof, and response determination operation status information indicating the status of the determination operation from receiving the analysis result information to determining the response content information Response content determining means for outputting
In response to the response content information from the response content determination means, the content is converted into a natural language output and sequentially presented to the user, and generation operation status information indicating the status of the natural language output generation operation by the conversion is generated. A language generating means for outputting,
According to at least one of the analysis operation status information, the response determination operation status information, and the generation operation status information, a simultaneous utterance such as a response to a user's natural language input or an interrupt is given to the language generation unit. A discourse management means controlling to be performed,
Having domain knowledge storage means for retaining domain knowledge information which is knowledge related to the content of the dialogue;
The language analysis means,
And obtained analysis result information, sequentially performs a comparison between the domain knowledge information recorded in the domain knowledge storage means, in case of conflict between the two has occurred is output as the analysis operation status information including the inconsistent content Do the processing ,
The discourse management means,
Upon receiving the analysis operation status information, a natural language output for realizing an interrupt composed of interjections is output through the language generating means, and a process of interrupting the user's utterance is performed. Natural language processor. The language analysis means,
In addition to the above-described processing, the user interrupts the utterance or utters a verbal utterance during comprehension, and performs processing for outputting analysis operation status information when performing analysis that the user has accepted an interrupt ,
The discourse management means,
When receiving the analysis operation status information from the language analysis means, the ambiguity at the time of analysis of the natural language input from the user that caused the interrupt, or inconsistency in the meaning of the user's natural language input 3. The natural language processing device according to claim 1, wherein a dialogue for eliminating the problem is performed. The language analysis means,
In addition to the above processing, when interrupting the user by the interjection under the control of the discourse management means, the user interrupts the utterance, or generates an interjection during understanding and interrupts the dialogue device. Detects whether the user has accepted it, outputs analysis operation status information indicating the result,
The discourse management means,
4. The natural language processing apparatus according to claim 3, wherein control is performed to activate a dialog for ambiguity resolution and inconsistency resolution in accordance with the analysis operation status information. An interruptable position feature storing means for storing in advance information relating to the end position of the utterance in the natural language dialogue, the syntactic feature of the partner utterance immediately before the interruptible position, or the prosodic feature of the partner utterance. ,
The discourse management means,
In addition to the above processing, when an interrupt is required, the analysis result information obtained from the language analyzing means is compared with the contents of the interruptable position feature storage means. 3. The natural language processing device according to claim 1, wherein control is performed such that interruption is performed by the interjection through the generation unit. Having a dialog history storage means for holding a history of the dialog with the user,
From the contents of the dialog history storage means, the position where the utterance change occurs or the position of the silent section is extracted as the utterance end position or the interruptable position,
Extract syntactic features or prosodic features immediately before the position,
6. The natural language processing apparatus according to claim 5, further comprising an interruptable position extracting unit that additionally records the feature in the interruptable position feature storage unit. The interruptable position feature storage unit holds information on the reliability of the interruptable position information,
The language analysis means,
An interrupt is performed based on the information recorded in the interruptable position feature storage means, and it is determined whether or not the user has interrupted the utterance or generated an interjection during the response to the interrupt and accepted the interrupt. Output analysis operation status information including the result,
The discourse management means,
6. The natural language processing device according to claim 5, wherein the reliability of the interruptable position information stored in the interruptable position feature storage means is adjusted according to the analysis operation status information. A language analyzing means for sequentially analyzing a natural language input from a user to output analysis result information of syntactic information, prosodic information or semantic information, and outputting analysis operation status information indicating the status of the analysis operation; ,
In response to the analysis result information from the linguistic analysis means, response content information is determined according to the content thereof, and response determination operation status information indicating the status of the determination operation from receiving the analysis result information to determining the response content information Response content determining means for outputting
In response to the response content information from the response content determination means, the content is converted into a natural language output and sequentially presented to the user, and generation operation status information indicating the status of the natural language output generation operation by the conversion is generated. A language generating means for outputting,
According to at least one of the analysis operation status information, the response determination operation status information, and the generation operation status information, a simultaneous utterance such as a response to a user's natural language input or an interrupt is given to the language generation unit. Having discourse management means for controlling to be performed,
The language analysis means,
Outputs analysis operation status information including information on ambiguity according to the number of candidates having ambiguity of the analysis result generated in the process of analyzing natural language input,
The discourse management means,
At the time of receiving the analysis operation status information, the language generating means outputs a natural language for realizing an interrupt made up of an interjection, thereby responding to the utterance of the user or interrupting the user. A natural language processing method characterized by the following. A language analysis step of sequentially analyzing a natural language input from a user to output analysis result information of syntactic information, prosodic information or semantic information, and outputting analysis operation status information indicating the status of the analysis operation; ,
In response to the analysis result information from the language analysis step, response content information is determined according to the content thereof, and response determination operation status information indicating the status of the determination operation until the analysis result information is received and the response content information is determined. Response content determining step of outputting
With respect to the response content information from the response content determination step, the content is converted into a natural language output and sequentially presented to the user, and the generation operation status information indicating the status of the natural language output generation operation by the conversion is generated. A language generation step to output,
According to at least one of the analysis operation status information, the response determination operation status information, and the generation operation status information, a simultaneous utterance such as an answer to a user's natural language input or an interrupt to the language generation step is performed. A discourse management step to control what to do,
Having a domain knowledge storage step of retaining domain knowledge information that is knowledge related to the content of the dialogue;
The language analysis step includes:
The obtained analysis result information is sequentially compared with the area knowledge information recorded in the area knowledge storage step, and when inconsistency occurs between the two, it is output as analysis operation status information including the inconsistent contents. ,
The discourse management step includes:
Upon receiving the analysis operation status information, a natural language output for realizing an interrupt consisting of an interjection is output through a language generation step to interrupt a user's utterance. Method. Input means for inputting natural language information;
In the natural language information, similar sound detection means for detecting a similar sound phrase that is a phrase having the same pronunciation or similar pronunciation phonetically,
Paraphrase candidate generation means for generating paraphrase candidates for the detected similar sound phrase;
When the detected similar sound phrases are plural, delete the paraphrase candidates having a similar sound relationship among the paraphrase candidates of each of the similar sound phrases, and replace the non-deleted paraphrase candidates with the respective similar sound phrases. A natural language processing device comprising: a paraphrase candidate determining unit for performing a replacement process. The paraphrase candidate determining means includes:
At the time of replacement of the synonymous phrase and the target paraphrase candidate, to add a control marker for changing the tone of the output voice before and after the target paraphrase candidate,
Phoneme processing means for changing the tone of the output voice by the control marker;
11. The natural language processing apparatus according to claim 10, further comprising a tone change selecting means for selecting whether or not to actually change the tone of the phoneme processing means. The paraphrase candidate determining means includes:
11. The natural language according to claim 10 , wherein, in addition to the processing, a history of the replacement operation to the paraphrase candidate is held, and the replacement operation is controlled according to a predetermined substitution rule for the paraphrase candidate corresponding to the history. Processing equipment. 4. The apparatus according to claim 1, further comprising a similar sound determining unit configured to remove unnecessary similar sound phrases from the similar sound phrases detected by the similar sound detecting unit based on word co-occurrence relation and appearance frequency information. 10. The natural language processing device according to 10. Detects similar phonetic phrases that are phonetically identical or similar in the input natural language information,
Generate paraphrase candidates for the detected similar phrases,
When the detected similar sound phrases are plural, delete the paraphrase candidates having a similar sound relationship among the paraphrase candidates of each of the similar sound phrases, and replace the non-deleted paraphrase candidates with the respective similar sound phrases. A natural language processing method characterized by replacing each of them.

Images