Device and method for processing temporal expression belonging to unstructured text for inputting data to knowledge database

Abstract

To provide a method and device for processing a temporal expression belonging to an unstructured text for inputting data to a knowledge database (102).SOLUTION: A method and device disclosed herein involve determining a temporal expression in a text, determining a type of the temporal expression as a function of the text, and mapping the temporal expression and the type on prediction of a value of the temporal expression in a context-free representation of the temporal expression.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an apparatus and method for processing temporal expressions belonging to unstructured text for populating a knowledge database. [Background technology]

[0002] For the purpose of populating a knowledge database, one can envisage the extraction and normalization of temporal expressions belonging to unstructured text.

[0003] Both subtasks can be solved together, as described, for example, in "SUTime: A library for recognizing and normalizing time expressions" by Chang & Manning, 2012 (https: / / aclanthology.org / L12-1122 / ) or "Multilingual and cross-domain temporal tagging" by Stroetgen & Gertz, 2013 (https: / / link.springer.com / article / 10.1007 / s10579-012-9179-y).

[0004] Both subtasks can be handled independently of each other, as described in "A Synchronous Context Free Grammar for Time Normalization" by Bethard, 2013 (https: / / aclanthology.org / D13-1078 / ) or "Context-dependent Semantic Parsing for Time Expressions" by Lee et al., 2014 (https: / / aclanthology.org / P14-1135 / ).

[0005] These normalization approaches, as well as those described in "CogCompTime: A Tool for Understanding Time in Natural Language" by Ning et al., 2018 (https: / / aclanthology.org / D18-2013 / ) and "Automatic rule generation for time expression normalization" by Ding et al., 2021 (https: / / aclanthology.org / 2021.findings-emnlp.269), are rule-based. [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] Chang & Manning, 2012, “SUTime: A library for recognizing and normalizing time expressions” (https: / / aclanthology.org / L12-1122 / ) [Non-patent document 2] Stroetgen & Gertz, 2013, “Multilingual and cross-domain temporal tagging” (https: / / link.springer.com / article / 10.1007 / s10579-012-9179-y) [Non-patent document 3] Bethard, 2013, “A Synchronous Context Free Grammar for Time Normalization” (https: / / aclanthology.org / D13-1078 / ) [Non-patent document 4] Lee et al., 2014, “Context-dependent Semantic Parsing for Time Expressions” (https: / / aclanthology.org / P14-1135 / ) [Non-Patent Document 5] Ning et al., 2018, “CogCompTime: A Tool for Understanding Time in Natural Language” (https: / / aclanthology.org / D18-2013 / ) [Non-patent document 6] Ding et al., 2021, “Automatic rule generation for time expression normalization” (https: / / aclanthology.org / 2021.findings-emnlp.269) Summary of the Invention [Means for solving the problem]

[0007] Disclosure of the Invention A computer-implemented method according to independent claim 1 for normalizing time expressions belonging to unstructured text for populating a knowledge database includes a deep learning-based method for normalizing time expressions.

[0008] The advantage of deep learning-based normalization methods over rule-based normalization is their transferability to new languages ​​and text genres, which would otherwise only be possible with greater effort in creating new rules.

[0009] Unlike rule-based normalization, deep learning-based normalization methods can normalize temporal expressions that are not covered by static rule systems. That is, deep learning-based normalization methods are more flexible than rule-based approaches with respect to sources of imperfections in the text, such as relatively minor orthographic errors or unknown words, whose meaning can be inferred from the context.

[0010] Deep learning-based normalization methods can be used in conjunction with various extraction models that do not need to be explicitly trained. In exemplary extraction models, extraction can be performed using a rule-based approach or a neural network. Extraction can be performed by the user, who sets the boundaries for the expressions to be normalized.

[0011] A method for processing time expressions, in particular those belonging to unstructured text, for populating a knowledge database is configured such that time expressions in the text are determined, a type of the time expression is determined in dependence on the text, and the time expressions and the type are mapped to a prediction of the value of the time expression in a non-contextual representation of the time expression, where the time expression and / or the type is determined in dependence on the text, for example based on rules or by means of a neural network trained therefor.

[0012] Preferably, a model is configured to map the time representation and type to the prediction, and the time representation and type are mapped to the prediction by this model. The value of the time representation is determined, for example, by a transformer trained for this purpose as a model, the transformer being configured, i.e., trained, to map the time representation and type to the prediction. Instead of a transformer, it is also possible to provide another model trained for this purpose, for example a deep learning model, in particular a recurrent neural network or a convolutional neural network, which is configured, i.e., trained, to map the time representation and type to the prediction.

[0013] Preferably, a text including a time expression is prepared, a reference date is determined or set for the text, time characteristics are determined depending on the context-independent expression and depending on the reference date, and the time characteristics are input into a knowledge database, thereby inputting time characteristics associated with the time context previously given by the text into the knowledge database.

[0014] Preferably, a temporal representation is provided, a non-contextual representation is provided, and a model is trained based on the temporal representation and the non-contextual representation to map the temporal representation to a prediction for the non-contextual representation. The resulting converter is flexible with respect to sources of impairment in the text, such as orthographic errors or unknown words, whose meaning can be inferred from the context. The resulting model, e.g., a converter or deep learning model, and the determination of the non-contextual representation, can be applied to multiple languages.

[0015] The non-context-dependent representation preferably includes a plurality of slots, the temporal representation is represented context-independently by the content of the slots, at least one of the plurality of slots is masked, and a model is trained in dependence on the temporal representation to map the temporal representation to a prediction for the masked content of the at least one masked slot, thereby training the model very efficiently as less training data is required compared to the context-dependent representation.

[0016] Preferably, in a first training phase, in which at least one slot is unmasked and at least one slot is masked, a model is trained to map the time representation and the unmasked content of the at least one unmasked slot to a prediction for the masked content of the at least one masked slot, which allows the model to be trained very efficiently.

[0017] Preferably, the model trained in the first training phase is trained to map temporal representations to predictions for slot contents in a second training phase in which all slots are specifically masked, thereby training the model very efficiently to learn perfect predictions for non-contextual representations.

[0018] Preferably, the model is trained in multiple iterations, with fewer slots masked during the first iteration than during the second iteration, and especially during the second iteration performed after the first iteration, which allows the model to be trained very efficiently, especially when very little training data exists, for example in various languages.

[0019] Preferably, the model determines a number or a numeric value or an operator, in particular an arithmetic operator, or a character or a word part, or a word or a wildcard from a predetermined vocabulary as a prediction for a slot of the non-context-dependent expression. Different vocabularies can be used for different non-context-dependent expressions. The vocabulary is predetermined according to the desired non-context-dependent expression. This makes the predictions by the model very reliable.

[0020] The model determines a probability distribution over the vocabulary, preferably for each masked content of one slot, and determines the content of the slot depending on this probability distribution, allowing the model to be trained very efficiently and to perform subsequent analysis or alternative predictions based on the probability distribution.

[0021] It is envisaged here that an excerpt containing a time expression is determined from the text based on rules and depending on the text, and that a prediction is then determined by a model depending on the excerpt containing the time expression. The excerpt is, for example, a single sentence. The prediction is based on the time expression in its context. This additionally improves the prediction.

[0022] What is envisaged here is that, in particular, an excerpt text containing a time expression is determined from the text based on a rule and in a text-dependent manner, and the model is trained in dependence on the excerpt text containing the time expression.

[0023] An apparatus for processing time expressions, in particular those belonging to unstructured text, for populating a knowledge database comprises a computing device configured to implement the above-described method, which has corresponding advantages.

[0024] A computer program comprising computer readable instructions which, when executed by a computer, causes the computer to carry out the above-described method has corresponding advantages.

[0025] Further advantageous embodiments can be seen from the following description and drawings. [Brief explanation of the drawings]

[0026] [Figure 1] 1 is a schematic diagram of an apparatus for processing time expressions belonging to unstructured text for populating a knowledge database; [Figure 2] FIG. 1 illustrates a training step in a method for processing temporal expressions belonging to unstructured text for populating a knowledge database. [Figure 3] FIG. 1 illustrates an inference step in a method for processing temporal expressions belonging to unstructured text for populating a knowledge database. DETAILED DESCRIPTION OF THE INVENTION

[0027] FIG. 1 shows a schematic diagram of an apparatus 100 for processing temporal expressions, particularly those belonging to unstructured text, for populating a knowledge database 102 .

[0028] Examples of time expressions are date expressions, such as January 2022, and time expressions, such as 13:00, and durations, such as 5 months.

[0029] Examples of applications for which data is envisaged for populating the knowledge database 102 are the creation or extension of knowledge graphs, or knowledge bases, or databases with temporal characteristics, for example, entries can be given validity periods based on time expressions, and the events they represent can be associated with specific points in time.

[0030] The apparatus 100 includes a computing device 104 configured to implement the methods described below for processing temporal expressions belonging to unstructured text for populating the knowledge database 102.

[0031] The computing device 104 includes, for example, at least one processor 106 and at least one memory 108. The at least one memory 108, in this example, is configured to store the knowledge database 102. The at least one memory 108, in this example, is configured to store instructions, e.g., computer programs, that, when executed by the at least one processor 106, perform the methods described below.

[0032] This method is explained below with an example of extraction of time expressions and subsequent normalization, which is part of "time tagging".

[0033] Extraction, in this example, involves type determination. Given a text, the task of extraction is to find the boundaries of the time expression. Such boundaries are usually so-called character offsets, which define the unambiguous boundaries of the expression within the text by means of two numbers. For example, an expression 11 characters long might start at character number 20 and end at character number 30 (inclusive).

[0034] The extraction may be performed, for example, based on rules, e.g., if an expression matches a pattern for a time expression according to a predetermined rule, the expression is identified as a time expression.

[0035] The extraction is performed, for example, using approaches for deep learning, in particular by solving a sequence labeling task or a Seq2Seq problem.

[0036] "From Characters to Time Intervals: New Paradigms for Evaluation and Neural Parsing of Time Normalizations" by Laparra et al. (2018) (https: / / aclanthology.org / Q18-1025 / ) and "Adversarial Alignment of Multilingual Models for Extracting Temporal Expressions from Text" by Lange et al. (2020) (https: / / aclanthology.org / 2020.repl4nlp-1.14 / ) describe extraction by solving sequence labeling tasks.

[0037] This method will be explained using an example of a text containing a time expression.

[0038] Typing is performed, for example, by HeidelTime, as described, for example, in "A Baseline Temporal Tagger for all Languages" by Stroetgen & Gertz, 2015 (https: / / aclanthology.org / D15-1063 / ).

[0039] The determination of the type of time expression is for example made in one of four classes (DATE, TIME, DURATION, SET), although other, more or fewer classes may be envisaged.

[0040] In this embodiment, the determination of the type of the time expression is performed simultaneously with the extraction of the time expression. It is also envisaged that the time expression may be extracted and then the type may be determined.

[0041] As an output of the extraction and type determination, in this embodiment, for each time expression, the annotation boundaries and types of that expression in the text are determined.

[0042] In the case of the exemplary excerpt text "The Eta Aquariiden Meteorschauer wird im Mai erscheinen." ("The Eta Aquarids meteor shower will appear in May"), for example, the time expression Mai (May) is extracted and its type is determined by the type DATE.

[0043] In this text excerpt, in this example, the time expression Mai is extracted by HeidelTime and the text excerpt itself is annotated with the type DATE. The resulting annotated text excerpt is <timex3 type=""DATE”"> Mai< / timex3> erscheinen.”

[0044] The time expressions in the normalization process are supplemented in the Mai example by a non-contextual representation of the time expression, which in this example is used in place of the time expressions in the excerpt text, along with the time expressions in the TIMEX 3 representation, resulting in the following excerpt text in this example: “Der Eta Aquariiden Meteorschauer wird im <timex3 type=""DATE”" value=""UNDEF-year-05”"> Mai< / timex3> erscheinen.”

[0045] TIMEX 3 is described, for example, in "The specification language TimeML" by James Pustejovsky et al. (The language of time: a reader, pages 545-557, Oxford University Press).

[0046] In this example, in HeidelTime, the time expression Mai having the position of the letter M in the text: START = 42 and the position of the letter i in the text: END = 44 is assigned the context-independent representation kuRepr of the time expression, kuRepr = "UNDEF-year-05".

[0047] For the context-independent representation KuRepr, in this example there is provided one value in the TIMEX 3 representation.

[0048] A text excerpt containing a TIMEX 3 expression is used in this example as the input quantity for normalization.

[0049] Normalization, in this example, involves Masked Language Modeling, which means that normalization is modeled as a Masked Language Modeling task, which involves a transformer with a language model, such as BERT.

[0050] BERT is described, for example, in "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding" (https: / / aclanthology.org / N19-1423 / ), published in 2019 by Devlin et al.

[0051] Instead of a transformer, it is also possible to provide other models trained for this purpose, for example deep learning models, in particular recurrent neural networks or convolutional neural networks.

[0052] In the Masked Language Modeling task, specific regions in a non-contextual representation are masked by a mask, and a transducer is trained to fill the mask.

[0053] In training the transformer, the transformer is trained to determine a prediction for a context-independent representation of a time expression in dependence on the time expression.

[0054] For the prediction of context-independent expressions by the transducer, multiple slots are used in this embodiment.

[0055] The context-independent representation kuRepr of the value of the TIMEX 3 expression is represented in this example by the contents of the slot.

[0056] In this embodiment, the content of one slot includes a part of a non-context-dependent expression or a wildcard. In this embodiment, multiple slots have a predetermined, invariant length. In this embodiment, multiple slots have the same length. Slots of different lengths may be defined. In this embodiment, each of these slots has a different semantic meaning.

[0057] For example, for the year-month-date representation, one slot each for the day representation with two digits and the month representation with two digits, and two slots for the year representation, are provided. For example, for the year representation, two slots with two digits for the ones and tens digits of the year and two slots with two digits for the hundreds and thousands digits of the year are provided. In this embodiment, the order of the slots is defined. The contents of the slots are unambiguously associated according to their meaning with respect to the time representation. Time representations less than one day are stored, for example, in three other slots: a slot with two digits for the hour representation, a slot with two digits for the minute representation, and a slot with two digits for the second representation. It is assumed here that at least one further slot is provided. The content of the at least one further slot can be a function call. The function call can be expressed by an operator such as "plus" or "minus." The content of the at least one further slot can be an argument for the function call. The content of at least one further slot may contain additional information, such as a marker indicating whether the date is in BC or AD, or whether at least one further slot is a period or a holiday. For example, the content BC or AD indicates the year in BC or AD. For example, the content of four slots could be this-day-plus-2 for the day after tomorrow or this-day-minus-1 for yesterday. For example, the content of a slot could indicate a holiday. No slots are provided for other expressions, such as UNDEF, because this expression is implicitly adopted when processing time expressions containing specific words such as "next" or "last".

[0058] The non-context-dependent representation can be reconstructed by a predetermined mapping to the contents of the slot and a portion of the non-context-dependent representation.

[0059] In this embodiment, unused slots are provided with wildcards: unused slots are slots that do not contain a value in the non-context-sensitive expression or that the transducer predicts will not contain a value in the non-context-sensitive expression.

[0060] In this example, 11 slots are used. It is contemplated that other numbers of slots may be used. The transducer is designed to predict the contents of a slot depending on a time expression. In this example, the transducer is designed to predict the contents of a slot depending on a text excerpt that includes a time expression. In this example, the transducer is designed to predict the contents of a slot depending on a text excerpt that includes a TIMEX 3 expression.

[0061] During training, one or more slots are masked with a mask, and the transformer is trained to determine a prediction that matches as closely as possible for the masked slot or for the masked slots.

[0062] In this embodiment, the text excerpt using the TIMEX 3 representation is fed to a transformer, and a probability distribution over the likely insertion content for each mask is calculated as the output. In this embodiment, the transformer includes weights, e.g., an artificial neural network with weights. In this embodiment, backpropagation is used to adapt the transformer weights.

[0063] In this example, a non-context-dependent representation of a time expression is predicted for normalization by the converter. It is assumed that the converter is used to determine a specific value in the TIMEX 3 representation, i.e., the non-context-dependent representation kuRepr. In this example, for the time expression Mai (May), the content of the slot for the month notation, e.g., 05, is predicted, while the contents of the other slots are wildcards. In this case, the content 05 is mapped to the non-context-dependent representation kuRepr = "UNDEF-year-05" based on the rule.

[0064] The predicted context-free representation for the temporal representation constitutes the normalized representation in this embodiment.

[0065] Populating the knowledge database includes, in this embodiment, two sub-aspects:

[0066] 1. When a time expression is a relative expression in the text, this expression is linked to a predetermined base date. For example, the time expression "yesterday" is semantically determined using a predetermined base date.

[0067] 2. If the time expression is an implicit expression belonging to the text, such as Christmas Eve, this expression is resolved by the semantic knowledge that this is the evening of December 24th.

[0068] In this embodiment, the knowledge database includes entities and relations, and these relations connect the entities to each other, preferably in pairs. An example of a knowledge database is a knowledge graph, which represents entities as nodes and relations as edges between the nodes. In one embodiment, the knowledge database includes 100 or more entities. In one embodiment, the knowledge database includes 1000 or more entities. In one embodiment, the knowledge database includes 10,000 or more entities. In one embodiment, the knowledge database includes 100 or more relations. In one embodiment, the knowledge database includes 1000 or more relations. In one embodiment, the knowledge database includes 10,000 or more relations.

[0069] The text, in this embodiment, includes at least one time expression. The text, in one embodiment, includes 100 or more time expressions. The text, in one embodiment, includes 1000 or more time expressions. The text, in one embodiment, includes 10,000 or more time expressions. The text, in this embodiment, includes one reference date. The text, in one embodiment, includes 100 or more reference dates. The text, in one embodiment, includes 1,000 or more reference dates. The text, in one embodiment, includes 10,000 or more reference dates.

[0070] The method for processing time expressions, particularly those belonging to unstructured text, for populating the knowledge database 102 includes a training step and an inference step. The purpose of the training step is to train a transducer to predict a context-free representation of a time expression, or multiple time expressions, belonging to the text. The purpose of the inference step is to use the transducer to predict a context-free representation of a time expression, or multiple time expressions, belonging to the text.

[0071] The goal of populating the knowledge database is to use one or more context-free expressions determined by an inference step on temporal expressions in the text to add one or more entities or one or more relations or their attributes to the knowledge database.

[0072] The training step can be performed separately from the inference step. The inference step can be performed separately from the training step, particularly using an already trained transformer.

[0073] The training steps are shown in Figure 2.

[0074] In step 202, a time representation is prepared.

[0075] In one embodiment, a text excerpt, for example a sentence, is determined from the text that includes the time expression.

[0076] The time expression is determined in the text based on, for example, rules.

[0077] In step 204, a context-independent representation kuRepr is prepared.

[0078] The context-independent representation kuRepr, in one embodiment, includes multiple slots: the time expression is represented context-independently by the contents of the slots.

[0079] In one embodiment, a text excerpt is prepared that includes a TIMEX 3 representation for a time expression and a non-context-sensitive representation of that expression.

[0080] In step 206, at least one of the slots of the non-context-dependent representation kuRepr is masked.

[0081] The transformer is trained in one embodiment over multiple iterations, with fewer slots masked during the first iteration than during the second iteration, which in one embodiment is performed after the first iteration.

[0082] In step 208, the non-contextual representation kuRepr and the time representation are mapped by the converter to a prediction for the non-contextual representation kuRepr. In one embodiment, a text excerpt containing a time representation is mapped to a prediction by the converter. In one embodiment, a text excerpt containing a TIMEX 3 representation for the time representation is mapped to a prediction by the converter.

[0083] It is assumed that in a first training phase, at least one slot is unmasked and at least one slot is masked, and during the first training phase, the temporal representation and the content of the at least one unmasked slot are mapped by a transformer to a prediction for the content of the at least one masked slot.

[0084] It is assumed that the slots are all specifically masked during the second training phase, during which the temporal representations are mapped to predictions for the slot contents by the transformers trained during the first training phase.

[0085] The converter in one embodiment determines as a prediction a number or a numeric value or an operator, in particular an arithmetic operator, or a letter or a word part, or a word or a wildcard from a predefined vocabulary.

[0086] Extracted text “Der Eta Aquariiden Meteorschauer wird im <timex3 type=""DATE”" value=""UNDEF-year-05”"> Mai< / timex3>erscheinen.” in this example contains the context-independent expression kuRepr = “UNDEF-year-05”. The transducer is trained in this example to predict the number 05 as content for the slot defined for it by the output of the transducer provided for this slot.

[0087] Further prediction examples are given below.

[0088] The excerpt text "...the day after tomorrow..." is represented, for example, by the non-context-dependent expression "UNDEF-this-day-PLUS-2", for which the transducer predicts the word "this" as the content in the transducer output for the first slot, the content "day" in the transducer output for the second slot, the operator PLUS as the content in the transducer output for the third slot, and the number 2 as the content in the transducer output for the fourth slot.

[0089] In one embodiment, the converter determines a lexical probability distribution for each masked content of a slot and determines the content of the slot depending on this probability distribution.

[0090] If the non-contextual representation includes multiple slots, the non-contextual representation and the temporal representation are mapped by the transducer to a prediction for the content of at least one masked slot.

[0091] Steps 202 to 208 are performed for a number of temporal expressions belonging to text in this embodiment during one training period.

[0092] In step 210, at least one weight of the converter is determined. The weight is determined, for example, depending on the deviation between a prediction for the non-contextual representation and the non-contextual representation. If the non-contextual representation includes multiple slots, the weight is determined depending on the deviation between a prediction for the masked content of at least one of the multiple slots and the masked content of the at least one slot. In this embodiment, the weight is determined by backpropagation depending on multiple deviations determined for multiple temporal representations belonging to the text.

[0093] Step 210 is performed over multiple training periods in this embodiment.

[0094] The inference step is illustrated in Figure 3. For inference, a transformer is configured to map a time expression and type to a prediction for a context-free representation of the time expression.

[0095] In step 300, a text containing a time expression is prepared.

[0096] In step 302, a reference date is determined or set for the text, for example, the reference date is determined from the text specifically based on rules.

[0097] In step 304, time expressions in the text are determined, in particular based on rules.

[0098] In one embodiment, the excerpt text is determined from the text containing the time expression in a text-dependent manner based on particular rules.

[0099] In step 306, the type of time expression is determined depending on the text, in particular based on rules.

[0100] In step 308, the time representation and type are mapped by a converter to a prediction for a non-contextual representation.

[0101] The converter in one embodiment determines as a prediction a number or a numeric value or an operator, in particular an arithmetic operator, or a letter or a word part, or a word or a wildcard from a predefined vocabulary.

[0102] In one embodiment, the transducer determines a probability distribution over the vocabulary for each slot and determines the content of that slot depending on this probability distribution.

[0103] In one embodiment, the predictions made by the transducer are determined in dependence on the excerpt text, which includes the time expression.

[0104] In step 310, a time characteristic is determined depending on the non-context-dependent expression and depending on the reference date.

[0105] In step 312, the knowledge database is populated with time characteristics.

[0106] Step 304 is then performed.

[0107] Steps 304 through 312 are performed in this embodiment until all time expressions present in the text have been processed.

Claims

1. In a computer-implemented method for processing temporal representations, particularly those belonging to unstructured text, for inputting data into a knowledge database (102), The time representation in the text is determined (304), The type of time expression is determined depending on the text (306), The time representation and the type are mapped to the prediction of the value of the time representation in the non-context-dependent representation of the time representation (308) A method characterized by the following:

2. The model is configured to map the time representation and the type to the prediction, so that the time representation and the type are mapped to the prediction by the model (308). The method according to claim 1.

3. A text containing the aforementioned time expression is prepared (300), A reference date is determined or set for the aforementioned text (302), The temporal characteristics are determined depending on the non-context-dependent expression and the reference date (310), The aforementioned time characteristics are entered into the knowledge database (312). The method according to claim 1.

4. Time representation is prepared (202), A context-independent expression is prepared (204), The model is trained dependent on the temporal representation and the non-context-dependent representation so as to map the temporal representation to predictions for the non-context-dependent representation (208). The method according to claim 2.

5. The non-context-dependent representation includes a plurality of slots, the time representation is represented independently of the context by the contents of the slots, and at least one of the plurality of slots is masked (206), The model is trained in a time representation-dependent manner so as to map the time representation to predictions for the masked content of at least one masked slot (208). The method according to claim 4.

6. In a first training phase in which at least one slot is not masked and at least one slot is masked, the model is trained to map the time representation and the unmasked content of the unmasked at least one slot to predictions for the masked content of the masked at least one slot (208). The method according to claim 5.

7. The model trained in the first training phase is trained in a second training phase in which the slots are all masked in particular to map the time representation to predictions for the contents of the slots (208). The method according to claim 6.

8. The model is trained over multiple iterations, and during the first iteration, fewer slots are masked than during the second iteration, and especially during the second iteration performed after the first iteration (206). The method according to claim 4.

9. The aforementioned model determines, as predictions, numbers or numerical values ​​or operators, particularly arithmetic operators, or characters or word components, or words or wildcards from a predetermined vocabulary (208, 308). The method according to claim 1.

10. According to the model, a probability distribution relating to the vocabulary is determined for each masked content in a slot, and the content of the slot is determined depending on the probability distribution (208, 308). The method according to claim 9.

11. In particular, based on the rules and dependent on the aforementioned text, an excerpt text including the time expression is determined from the aforementioned text (304), The prediction is determined by the model, depending on the excerpt text including the time expression (308). The method according to claim 1.

12. In particular, based on rules and depending on the text, an excerpt text containing the time expression is determined from the text (202), and the model is trained depending on the excerpt text containing the time expression. The method according to claim 1.

13. A device (100) for processing time representations, particularly those belonging to unstructured text, for inputting data into a knowledge database (102), The apparatus (100) is characterized by including a computing device configured to carry out the method described in any one of claims 1 to 12.

14. A computer program, wherein the computer program includes computer-readable instructions, and when the computer-readable instructions are executed by a computer, the computer performs the method described in any one of claims 1 to 12.

Images