International Business Machines Corporation

Yang Yu

Ming Tan

Saloni Potdar

Haoyu Wang

Ladislav Kunc

Haode Qi

Navneet N Rao

Wei Zhang

Murray Scott Campbell

Sadhana Kumaravel

Ravi Nair

Dakuo Wang

Robert Leslie Yates

Shiyu Chang

Mo Yu

Panos Karagiannis

Shasha Lin

An artificial intelligence (AI) platform to support random action replay for natural language (NL) learning. A NL conversation is subject to exploration to train a neural network. One or more tuples are leveraged for the training, with each tuple representing an input action, a vector, an output action, and a reward value. An action is sampled from the vector, with the sampling configured to assess a corresponding first gradient. The first gradient is applied to selectively adjust the neural network. As NL input is received and applied to the selectively adjusted neural network, an output corresponding to the NL input is identified and a corresponding action is subject to be executed.

Random action replay for reinforcement learning

An approach to determine the quality of encodings assigned to a word by a word embedding model. The approach may include determining the asymmetry of two embeddings associated with two words from a word embedding model. The asymmetry of the two words from a preexisting evocation dataset may be determined. The asymmetry of the two embeddings may be compared to the asymmetry from the evocation dataset to generate an encoding quality score.

Word embedding quality assessment through asymmetry

A computer-implemented method includes using an embedding network to generate prototypical vectors. Each prototypical vector is based on a corresponding label associated with a first domain. The computer-implemented method also includes using the embedding network to generate an in-domain test vector based on at least one data sample from a particular label associated with the first domain and using the embedding network to generate an out-of-domain test vector based on at least one other data sample associated with a different domain. The computer-implemented method also includes comparing the prototypical vectors to the in-domain test vector to generate in-domain comparison values and comparing the prototypical vectors to the out-of-domain test vector to generate out-of-domain comparison values. The computer-implemented method also includes modifying, based on the in-domain comparison values and the out-of-domain comparison values, one or more parameters of the embedding network.

Out-of-domain encoder training

Domain specific model compression by providing a weighting parameter for a candidate operation of a neural network, applying the weighting parameter to an output vector of the candidate operation, performing a regularization of the weighting parameter output vector combination, compressing the neural network model according to the results of the regularization, and providing the neural network model after compression.

Domain specific model compression

A system includes a memory having instructions therein and at least one processor configured to execute the instructions to: receive a natural language question; determine, from a chat log comprising a plurality of chat session logs, a set of chat session logs most relevant to the natural language question; determine a respective plurality of non-overlapping text spans most relevant to the natural language question within each of a respective plurality of conceptual pseudo-documents; determine a conceptual pseudo-document most relevant to the natural language question; extract a question-answer pair most relevant to the natural language question from the most relevant pseudo-document; and convey the most relevant question-answer pair to a user. Each one of the conceptual pseudo-documents corresponds to a respective one of the most relevant chat session logs.

Contextual question answering using human chat logs

A mechanism is provided to implement an abnormal entity detection mechanism that facilitates detecting abnormal entities in real-time response systems through weak supervision. For each first intent from an entity labeled workspace that matches a second intent in labeled chat logs, when the entity score associated with each first entity or second entity is above a predefined significance level the first entity or the second entity is recorded. For each first intent from the entity labeled workspace that matches the second intent in the labeled chat logs: responsive to the first entity being recorded and the second entity failing to be recorded, that first entity is removed from the training data as being mistakenly included; or, responsive to the second entity being recorded and the first entity failing to be recorded, that second entity is added as a potential business case to the training data.

Weak supervised abnormal entity detection

A mechanism is provided to implement suggestion of new entity types with discriminative importance analysis. The mechanism obtains a list of predefined intents from a chatbot designer. The mechanism receives an input sentence having a target intent within the list of predefined intents. The mechanism performs intent-specific importance analysis on the input sentence to generate an importance score for each token in the input sentence. The mechanism ranks the tokens in the input sentence by importance score and outputs a token with a highest importance score as a candidate entity type.

Suggestion of new entity types with discriminative term importance analysis

A method of text classification includes generating a text embedding vector representing a text sample and applying weights of a regression layer to the text embedding vector to generate a first data model output vector. The method also includes generating a plurality of prototype embedding vectors associated with a respective classification labels and comparing the plurality of prototype embedding vectors to the text embedding vector to generate a second data model output vector. The method further includes assigning a particular classification label to the text sample based on the first data model output vector, the second data model output vector, and one or more weighting values.

Hybrid model for short text classification with imbalanced data

A mechanism is provided for implementing an intent segmentation mechanism that segments intent boundaries for multi-intent utterances in a conversational agent. For each term of a set of terms in the utterance from a real-time chat session, a set of adversarial utterances is generated for the utterance. An influence of changing each term is determined so as to identify a term importance value. Utilizing the term importance value, one or more of a change in ranking of the intent of the utterance or a change in confidence with regard to the intent of the utterance is identified. An entropy-based segmentation of the utterance into a plurality of candidate partitions is performed. An associated intent and entropy value are then assigned. Based on a segment with minimum entropy, a call associated with the real-time chat session is directed to an operation associated with an intent of the segment with minimum entropy.

Intent boundary segmentation for multi-intent utterances

A computing device receives a natural language input from a user. The computing device routes the natural language input from an active domain node of multiple domain nodes of a multi-domain context-based hierarchy to a leaf node of the domain nodes by selecting a parent domain node in the hierarchy until an off-topic classifier labels the natural language input as in-domain and then selecting a subdomain node in the hierarchy until an in-domain classifier labels the natural language input with a classification label, each of the plurality of domain nodes comprising a respective off-topic classifier and a respective in-domain classifier trained for a respective domain node. The computing device outputs the classification label determined by the leaf node.

Routing text classifications within a cross-domain conversational service

A mechanism is provided for implementing a bias detection mechanism that mitigates unintended bias in a conversational agent by leveraging conversational agent definitions, a conversational agent chat logs, and user satisfaction statistics. One or more protected attributes are identified within an utterance from the conversational agent chat logs. Using the identified protected attributes, a replacement utterance with a replacement term is generated for at least one of the identified protected attributes in the utterance. A score is generated for the utterance and the replacement utterance using utterance level relative term importance for protected attributes and regular terms in the utterance and the replacement utterance. Utilizing the scoring, a determination is made as to whether unintended bias exists within the utterance. Responsive to unintended bias being detected, an action is implemented that causes a change to a machine learning model used by the conversational agent.

Unintended bias detection in conversational agent platforms with machine learning model

A mechanism is provided to implement a text classifier training augmentation mechanism for incorporating unlabeled data into the generation of a text classifier. For each term of a plurality of terms in each document of a plurality of documents in a set of unlabeled data, a term frequency value is determined. The term is normalized by dividing the term frequency value by a total number of terms in the document. An inverse document frequency (idf) value is determined for each term based on the term frequency value. A subset of terms is filtered from the plurality of terms based the determined idf values. The idf values for the remaining terms are transformed into feature weights. Terms from a set of labeled data are re-weighted based on the feature weights determined from the set of unlabeled data. The text classifier is then generated using the re-weighted labeled data.

Feature reweighting in text classifier generation using unlabeled data

An online version of a sentence representation generation module updated by training a first sentence representation generation module using first labeled data of a first corpus. After training the first sentence representation generation module using the first labeled data, a second corpus of second labeled data is obtained. The second corpus is distinct from the first corpus. A subset of the first labeled data is identified based on similarities between the first corpus and the second corpus. A second sentence representation generation module is trained using the second labeled data of the second corpus and the subset of the first labeled data.

Updating an online multi-domain sentence representation generation module of a text classification system

A computer-implemented method includes obtaining a training data set including a plurality of training examples. The method includes generating, for each training example, multiple feature vectors corresponding, respectively, to multiple feature types. The method includes applying weighting factors to feature vectors corresponding to a subset of the feature types. The weighting factors are determined based on one or more of: a number of training examples, a number of classes associated with the training data set, an average number of training examples per class, a language of the training data set, a vocabulary size of the training data set, or a commonality of the vocabulary with a public corpus. The method includes concatenating the feature vectors of a particular training example to form an input vector and providing the input vector as training data to a machine-learning intent classification model to train the model to determine intent based on text input.

Weighting features for an intent classification system

A method includes providing input text to a plurality of multi-task learning (MTL) models corresponding to a plurality of domains. Each MTL model is trained to generate an embedding vector based on the input text. The method further includes providing the input text to a domain identifier that is trained to generate a weight vector based on the input text. The weight vector indicates a classification weight for each domain of the plurality of domains. The method further includes scaling each embedding vector based on a corresponding classification weight of the weight vector to generate a plurality of scaled embedding vectors, generating a feature vector based on the plurality of scaled embedding vectors, and providing the feature vector to an intent classifier that is trained to generate, based on the feature vector, an intent classification result associated with the input text.

Growing community of inventors

Acton, MA, United States of America

Yang Yu