Automated system and method to prioritize language model and ontology expansion and pruning

Abstract

A system and method for updating computerized language models is provided that automatically adds or deletes terms from the language model to capture trending events or products, while maximizing computer efficiencies by deleting terms that are no longer trending and use of knowledge bases, machine learning model training and evaluation corpora, analysis tools and databases.

Description

BACKGROUND

Field

Embodiments of the present invention relate to language models and ontologies, and more particularly, to a system and method for automatically prioritize language model and ontology expansion and pruning.

Background

In speech recognition, a language model (LM) is a graph of probabilities associated to word transitions from a known vocabulary, such as a word lattice. Word embedding is the collective name for a set of language modeling and feature learning techniques in natural language processing (NLP) where words or phrases from the vocabulary are mapped to vectors of real numbers. Some approaches to language model development include term frequency inverse document frequency (TF-IDF) and word similarity. For instance, vocabulary in the insurance domain is expected to differ greatly from vocabulary in the telecommunications domain. To create a LM for use in a specific domain, texts are gathered from various sources such as websites, chat logs, call logs, documentation, and other sources in that domain, but each such domain may use different terms or syntax for the same meaning. There is a need for a system and method to automatically prioritize language model and ontology expansion and pruning.

BRIEF SUMMARY OF THE DISCLOSURE

Accordingly, the present invention is directed to a system and method for a system and method for automatically prioritize language model and ontology expansion and pruning that obviates one or more of the problems due to limitations and disadvantages of the related art.

In accordance with the purpose(s) of this invention, as embodied and broadly described herein, this disclosure, in one aspect, relates to a computer product comprising computer executable code embodied in a non-transitory computer readable medium that, when executing on one or more computing devices performs a method of normalizing terminology and phrases within a language model for a language domain. The method includes receiving text from a plurality of platforms; determining whether the text includes a term in a stored data model; identifying if a term that does exist in the data model appears in a new context; passing the term in the new context to a human for determination if the term should be added to a training example in the new context for retraining the data model; if the term does not appear in the new context, checking the term for frequency of use in a known context and adding the term in the known context to the training example with a new priority if the frequency has reached a predetermined threshold; and recompiling the language model based on the term in context.

In another aspect, the disclosure relates to a method of adding terms to a language model based on use. The method includes receiving text from a plurality of platforms; determining whether the text includes a term in a stored data model; identifying if a term that does exist in the data model appears in a new context; passing the term in the new context to a human for determination if the term should be added to a training example in the new context for retraining the data model; if the term does not appear in the new context, checking the term for frequency of use in a known context and adding the term in the known context to the training example with a new priority if the frequency has reached a predetermined threshold; and recompiling the language model based on the term in context.

In another aspect, the disclosure relates to a method of removing terms from a language model based on use. The method includes receiving text from a plurality of platforms; determining whether the text includes a term in a stored data model; if the term is in the data model, determining a frequency of use of the term in the text in a context in which the term appears in the data model; deleting the term in context from a training example for the data model if the frequency of use falls below a predetermined threshold; and recompiling the language model based after removing the term in the context from the training example.

Further embodiments, features, and advantages of the system and method for a system and method for automatically prioritize language model and ontology expansion and pruning, as well as the structure and operation of the various embodiments of the system and method for a system and method for automatically prioritize language model and ontology expansion and pruning, are described in detail below with reference to the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying figures, which are incorporated herein and form part of the specification, illustrate the system and method for automatically prioritize language model and ontology expansion and pruning. Together with the description, the figures further serve to explain the principles of the system and method for automatically prioritize language model and ontology expansion and pruning described herein and thereby enable a person skilled in the pertinent art to perform and use the system and method for automatically prioritize language model and ontology expansion and pruning.

FIG. 1 is an example of a word lattice.

FIG. 2 illustrates an active learning process according to principles described herein.

FIG. 3 is a flowchart showing an example set of steps for performing a method as described herein.

DETAILED DESCRIPTION

Reference will now be made in detail to embodiments of the system and method for a system and method for automatically prioritize language model and ontology expansion and pruning with reference to the accompanying figures. The same reference numbers in different drawings may identify the same or similar elements.

It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention cover the modifications and variations of this invention provided they come within the scope of the appended claims and their equivalents.

To create a language model (LM) for use in a specific domain, texts are gathered from various sources such as websites, chat logs, call logs, documentation, and other sources in that domain. Once the texts are aggregated, LM construction toolkits such as the CMU [1], SRI[2], or IRST [3] are applied to the data. They extract the vocabulary used within the texts and the statistics of their use with other vocabulary, such as unigrams, bigrams, and trigrams. These statistics can then be used to calculate a priori statistics of sentences that can be formed using the known vocabulary, which are organized in a lattice. A word lattice is an acyclic directed graph with a single starting node and edges labeled with a word and its corresponding probability of following the current word in the source texts. By following a path through the lattice from the starting point to any particular node, the a priori probability of that series of words (i.e. a sentence) appearing in the domain specific texts can be calculated. In the case of FIG. 1, the subject phrase is “a conference is being recorded.

A different approach to modeling word usage in context is to construct vectors to represent each word in a N-dimensional vector space. These vectors are manipulated during training based on observing where terms occur in the context of the surrounding terms. Terms that occur in the same context are moved closer to alignment. Terms that do not occur in the same context are moved further away. Once trained, the set of vectors can be used to reason about the similarity of words by performing vector arithmetic, such as measuring the distance between two points in the vector space. This approach is known as word embeddings [4], and is a way to group similar terms in a corpus together. Both the LM and word embedding approaches are unsupervised in that they require no human effort to construct. The training algorithms are simply given large training corpora and they use term positions and statistics within the corpora to build a model.

In contrast to models showing the statistical relationship between terms in a training corpora, data modeling approaches seek to define deeper relationships between terms such as hierarchies and negations. For such models there are two common structures used. The simpler form is a taxonomy, which is simply a tree of entities that form a hierarchy. For example, you could create a taxonomy of food where the entities are individual food items such as cheddar cheese, peas, corn, apples, pork, skim milk, etc. You would then create low level classes of foods like red meat, white meat, all cheese, all milk, families of fruits and vegetables, etc. Then you group all of the specific individuals into the classes they belong. Next you create higher level classes such as meat, fish, dairy, fruit, vegetables, etc. and group the classes of foods into the higher level classes. Finally, you can create the top layers of animal products, and non-animal products and put them under the root node of food. In this way you have constructed a taxonomy of food that you can go from specific examples to more and more general classes by following the tree backwards. You can also do simple reasoning like parent-of or sibling-of relationships, and find the least common ancestor between two individuals, like animal products for milk and pork.

For many cases this tree structure is enough to model data and process it. But more complicated relationships, like multiple inheritance and applying logical assertions, require storing data and meta data in a graph form. This is where ontologies come in. An ontology is a directed graph with four primary components: individuals, classes, attributes, and relations. There are many more components possible like events and restrictions as well. Ontologies allow for very rich data modeling with complex relationships and logical inferences about the data. There are many ways to construct ontologies and several different syntaxes for expressing and storing them. Taxonomies and ontologies typically require some human effort to construct. They may be seeded by some statistical observations from corpora, but the relationships between terms are usually defined or refined by humans. These models are concerned with the logical inference that can be drawn from terms within them and therefore require at least some logical relations to be encoded within them by humans.

Human-in-the-loop (HITL) is a subfield of Machine Learning where the model requires some form of human interaction. A common HITL approach is known as Active Learning. With active learning an existing model is supplied with a large pool or stream of unlabeled samples. The model then chooses which samples it thinks would be most informative to know the label for based on a selection strategies, of which there are several commonly used. Human oracles are then shown the selected samples and give them labels. These labeled samples are added to the training data to retrain the model from. In this way the model will learn more quickly from less training data then given a large sample of labeled results that contain many duplicated features. This active learning process is shown in FIG. 2.

Language model and ontology refinement for Intelligent Virtual Assistants (IVAs) is described herein. It is not necessary for the IVA to understand every possible word in order to understand a user's intention in a query. Computational overhead of unbounded LMs and data models will increase understanding latency. However, some words, such as those referring to products or services, are important to understand.

Therefore it is desirable to monitor user communication channels for new terminology that should be understood. Personal communication texts such as emails, instant messaging, and social media are particularly challenging due to their open vocabulary that continuously grows. There are constantly new products, applications, devices, terminology, slang and abbreviations being created and used within such communication channels. In order to deal with the evolving nature of internet language in a timely fashion, automated methods to detect high value words for insertion into LMs and data models are needed. Conversely, words that fall out of use, such as discontinued products or cultural references that are no longer popular, should be removed to maintain the size of models and speed traversals through them.

Herein a system and method to detect and prioritize insertions of new terminology into language and data models. Also provided herein are a system and method to prioritize removal of unused terms from such models to limit their growth and searching latency. In order to discover new terms to add to the IVA's knowledge, several content streams may be monitored. For example, one stream may be trending topics in social media platforms. These may originate from Twitter, Facebook, Pinterest or similar sites where users are actively communicating around topics. As topics gain popularity, they begin to “trend” by rising to the top of the subjects or topics that people are communicating about.

For example, during the holiday season there are new products such as toys and electronic devices that are released to take advantage of the seasonal increase in consumer spending. Suppose one such product is a new smart phone device such as the Google Pixel. When the device is announced or released, there is a sudden emergence in conversations around the topic of the Pixel, where before the term was not related to the electronics domain at all, or may not have even existed if it is an original name. By monitoring trending topics we will observe a sudden appearance of an unknown term, or a term that is not previously associated to the context it appears in.

In a second example, suppose a tropical storm has appeared and is making landfall. Tropical storms are commonly named and if the storm is expected to have a large impact on a populated area many news sites and, therefore, social media sites will experience a sudden spike in conversations around this new name. For IVAs in the travel domain, it will be helpful that these events are understood quickly as many travelers will begin to ask about disruptions caused by the storm.

Additional content streams can be customer e-mail and live chat transcripts, or any other form of customer to company communication channels. For weather occurrences, feeds such as the NOAA and Weather Channel can be monitored. News sites and aggregates of new feeds can also be ingested. From these sources without the construct of trends, terms can be counted over a sliding time window such as a day or week to create a set of trending terms.

Term extraction may be performed on a large corpus of social media texts (and/or customer support emails, and/or other business related social media and/or emails) that are associated with a specific company. This results in a large set (or list) of terms, such as n-grams including unigrams, bigrams, trigrams, 4-grams, etc. that have a high likelihood to appear in these social media texts, emails, etc.

The set (or list) of terms is compared to the lexicon (vocabulary) and language model (e.g., ARPABO file which is based on the ARPA-format Back-Off Ngram language model) associated with the LVCSR (Large Vocabulary Continuous Speech Recognition) engine that is used to transcribe a call from a company's call center. This reveals out-of-vocabulary unigrams, and higher order n-grams that are missing in the language model or have low language-model weights.

Corrective action may then be automatically taken. Such corrective action may include adding the frequent out-of-vocabulary unigrams to the lexicon (and to the language model) and/or boosting the weights of the frequent n-grams in the language model.

In some implementations, it is contemplated that the prominent terms represent product names, department names, etc. that are specific to a company, and as such they are expected to occur in speech interaction as well. Therefore, augmenting the lexicon and language model will increase the accuracy of the LVCSR engine

Regardless of the source, when new terminology appears in trends we first consult the existing LM and data models used by the IVA. If the term is unknown to the LM it must be prioritized for addition. This prioritization can be based on any predetermined characteristic, e.g., frequency, topic, or source. For example, a frequency chart of mentions across multiple channels/source may be populated to determine the prevalence of the new term. For example, if a new term is only trending within a specific source such as Twitter, it may refer to some isolated phenomenon that may not be as important to be known by the IVA.

On the other hand, a new term may be important for the IVA to know to facilitate handling of incoming customer communication. An example of such phenomenon was the volcanic events that occurred in Iceland in 2010, which caused an ash cloud that grounded air transportation in Europe and caused a spike in the usage of terms such as “ash” and “volcano” that were previously very low ranked in IVAs for airlines. In the case of a tropical storm name or highly anticipated new product, the term might have been previously heard by the IVA (such as tropical storm name “Michael”), the term should be temporarily highly ranked across multiple sources, i.e., much higher and in a different way than the given name “Michael” might previously have been encountered in the existing ontology. That is, the context for an existing term has changed.

We can use this coverage of trending terms across multiple sources, e.g. external to the IVA or in conjunction with the IVA, to inform the usage of terms and to prioritize their addition to or usage in the data model/language model. For example, the frequency with which a term appears, context, and/or threshold counts can factor in the prioritization/addition of a term. In one aspect, a human could provide input for retraining the model based on new terms, including recommending a new term's priority and how that term should be deployed to the IVA. In the inverse, a term could be removed or priority reduced based on these factors, for example, when the volcano eruption is no longer a factor in air travel or the hurricane has passed, as discussed further below.

In some cases the term will be known, but the context it appears in is new. In the above Google Pixel example, the word “pixel” may already appear in the IVA's vocabulary, but in its ontology it is used for display or camera resolution. To detect usage, we can create embedding models from sliding windows of time using the text from the various input sources. If the existing term becomes embedded along with words that were not previously associated with the term, such as “iPhone”, “Apple”, or “Google” in the case of the Pixel, we can determine that the new usage of the existing term indicates a new object. In these cases the ontology or other data models will need to be updated to reflect the alternative usage and this update will be prioritized the same way as a new term. Once terms cross a predefined threshold, which is adjustable based on the availability of human annotators, they are given to a human along with example context to be added to the data models and/or LMs. This process of the model selecting its own training data is a special case of HITL known as Active Learning [5]. The human will then add the term to the ontology or update the ontology to reflect the new usage of the existing term. For new terms they will need to be added to the LMs as well so that speech recognition engines will be able to decode the word successfully. A flow chart for an exemplary implementation is provided in FIG. 3.

For an IVA, how a new term is being embedded can be studied. For example, if we look at previous versions of a word lattice or word embedding model, new text can be fed into the model to see how existing terms are being embedded with respect to new pairs or strings of terms. New usage of a term can therefore be identified and its new meaning/usage incorporated into the ontology/language model. This allows for the new usage to be disambiguated based on context. For example, the term can be updated in the ontology and its meaning updated. Its appearance in the language model/word lattice or word embedding model can therefore be changed to reflect the updated usage.

While monitoring streams of trending topics and terminology from various sources, a list of known vocabulary is maintained. This list reflects all terms known to the IVA through the LMs and data models. Each term in the vocabulary is associated to a timestamp of last mention. When a term is encountered in the input streams and it exists in the vocabulary, the timestamp associated to the term is updated. If a product is discontinued or a tropical storm passes over, we would expect the usage of such terms will diminish over time. Once a term has not been mentioned longer than some tunable cutoff period, over one year for example, it will be deleted from the LMs and data models of the IVA causing the IVA to “forget” the unused term. In this way terms that have a lifespan are not maintained in the LMs and data models, which will prevent unbounded growth and steadily increasing time to search and traverse the models. As the pruning requires no human intervention it can be done automatically, however human review can be used if desired to approve the modifications.

According to principles described herein, the time from when a new term enters use and its adoption by an IVA (or other language based system) can be reduced, thus causing an improvement in the functioning such a device.

According to principle described herein, a human may be involved in determining thresholds by which the system may then run automatically for addition of trending terms and removal of decreasing terms. In the case of addition, a human in the loop may improve retraining based on new terms because a human can provide context for use in retraining the model and recommend priority. Such human activity in conjunction with the automated system of language model and word embeddings described above, can increase the speed with by which the automated models can be retrained to account for trending and declining terms.

Throughout this application, various publications may have been referenced. The disclosures of these publications in their entireties are hereby incorporated by reference into this application in order to more fully describe the state of the art to which this invention pertains:

• [1] Ronald Rosenfeld. The cmu statistical language modeling toolkit and its use in the 1994 arpa csr evaluation. In Proceedings of the Spoken Language Systems Technology Workshop, pages 47{50, 1995.
• [2] Andreas Stolcke. Srilm-an extensible language modeling toolkit. In Seventh international conference on spoken language processing, 2002.
• [3] Marcello Federico, Nicola Bertoldi, and Mauro Cettolo. Irstlm: an open source toolkit for handling large scale language models. In Ninth Annual Conference of the International Speech Communication Association, 2008.
• [4] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013.
• [5] Burr Settles. Active learning. Synthesis Lectures on Artificial Intelligence and Machine Learning, 6(1):11114, 2012.

While various embodiments of the present invention have been described above, it should be understood that they have been presented by way of example only, and not limitation. It will be apparent to persons skilled in the relevant art that various changes in form and detail can be made therein without departing from the spirit and scope of the present invention. Thus, the breadth and scope of the present invention should not be limited by any of the above-described exemplary embodiments, but should be defined only in accordance with the following claims and their equivalents.

Claims

1. A non-transitory computer readable medium comprising instructions that, when executed by a processor of a processing system, cause the processor to perform a method of updating a language model for a language domain for an interactive virtual assistant, the method comprising: monitoring business related textual data from a plurality of platforms;collecting trending n-grams from the textual data over a sliding time window;comparing the n-grams to a vocabulary in a data model to identify terms comprising one of the trending n-grams existing in the data model;for a term comprising the one of the trending n-grams in the data model and appearing in a new context, wherein the new context is not associated with the trending n-gram in the data model, passing the term in the new context to a human for determination if the term should be added to a training example in the new context for retraining the data model;for a term comprising the one of the trending n-grams in the data model and not appearing in a new context, ignoring the term;for a term comprising the one of the trending n-grams and not in the data model, checking the term for frequency of use in the textual data and adding the term in context to the training example if the frequency of use has reached a predetermined threshold;recompiling the language model based on the term in context; andadopting the recompiled language model in an interactive virtual assistant.

2. The non-transitory computer readable medium of claim 1, the method further comprising, for the term existing in the data model and appearing in the new context, determining a frequency of appearance of the term and, when the frequency reaches a predetermined threshold, adding the term to the training example as the term appears in context.

3. The non-transitory computer readable medium of claim 1, the method further comprising ignoring terms that do not meet the predetermined threshold.

4. The non-transitory computer readable medium of claim 1, the method further comprising deleting known terms from the data model if the frequency of use of the known term in a known context falls below a predetermined threshold.

5. The non-transitory computer readable medium of claim 1, wherein determining the predetermined threshold for adding a term not in the data model to the training model is met is determined automatically based on predetermined values stored in a digital storage device.

6. The non-transitory computer readable medium of claim 1, wherein the term is passed to a human only when the term in the new context meets a predetermined threshold based on parameters of an existing ontology.

7. The non-transitory computer readable medium of claim 1, wherein the business-related textual data comprises social media texts, customer support emails, and/or other business related social media and/or emails.

8. The non-transitory computer readable medium of claim 1, wherein the n-grams comprise unigrams.

9. The non-transitory computer readable medium of claim 1, wherein the n-grams comprise bigrams.

10. The non-transitory computer readable medium of claim 1, wherein the n-grams comprise a contiguous sequence of n items in the business-related textual data, wherein n is a positive integer.

11. The non-transitory computer readable medium of claim 1, wherein the n-grams comprise a contiguous sequence of n items in the business-related textual data, wherein n is a positive integer greater than two.

12. A method of adding terms to a language model based on use, the method comprising: monitoring business related textual data from a plurality of platforms;collecting trending n-grams from the textual data over a sliding time window;comparing the n-grams to a vocabulary in a data model to identify terms comprising one of the trending n-grams existing in the data model;for a term comprising the one of the trending n-grams in the data model and appearing in a new context, wherein the new context is not associated with the trending n-gram in the data model, passing the term in the new context to a human for determination if the term should be added to a training example in the new context for retraining the data model;for a term comprising the one of the trending n-grams in the data model and not appearing in a new context, ignoring the term;for a term comprising the one of the trending n-grams and not in the data model, checking the term for frequency of use in the textual data and adding the term to the training example if the frequency of use has reached a predetermined threshold;recompiling the language model based on the term in context; andadopting the recompiled language model in an interactive virtual assistant.

13. The method of claim 12, further comprising, for the term existing in the data model and appearing in the new context, determining a frequency of appearance of the term and, when the frequency reaches a predetermined threshold, adding the term to the training example as the term appears in context.

14. The method of claim 12, further comprising ignoring terms that do not meet the predetermined threshold.

15. The method of claim 12, further comprising deleting known terms from the data model if the frequency of use of a known term in a known context falls below a predetermined threshold.

16. The method of claim 12, wherein determining the predetermined threshold for adding a term not in the data model to the training model is met is determined automatically based on predetermined values stored in a digital storage device.

17. The method of claim 12, wherein the term is passed to a human only when the term in the new context meets a predetermined threshold based on parameters of an existing ontology.

18. The method of claim 12, wherein the business-related textual data comprises social media texts, customer support emails, and/or other business related social media and/or emails.

19. The method of claim 12, wherein the n-grams comprise unigrams.

20. The method of claim 12, wherein the n-grams comprise bigrams.

21. The method of claim 12, wherein the n-grams comprise a contiguous sequence of n items in the business-related textual data, wherein n is a positive integer.

22. The method of claim 12, wherein the n-grams comprise a contiguous sequence of n items in the business-related textual data, wherein n is a positive integer greater than two.

23. A system for performing speech analytics, comprising: a memory comprising executable instructions;a processor configured to execute the executable instructions and cause the system to:monitor business related textual data from a plurality of platforms;collect trending n-grams from the textual data over a sliding time window;compare the n-grams to a vocabulary in a data model to identify terms comprising one of the trending n-grams existing in the data model;for a term comprising the one of the trending n-grams in the data model and appearing in a new context, wherein the new context is not associated with the trending n-gram in the data model, pass the term in the new context to a human for determination if the term should be added to a training example in the new context for retraining the data model;for a term comprising the one of the trending n-grams in the data model and not appearing in a new context, ignore the term;for a term comprising the one of the trending n-grams and not in the data model, check the term for frequency of use in the textual data and add the term in context to the training example if the frequency of use has reached a predetermined threshold;recompile the language model based on the term in context; andadopt the recompiled language model in an interactive virtual assistant.