System and method for detection and correction of a query
Systems and methods are provided for systematically finding and fixing automatic speech recognition (ASR) mistranscriptions and natural language understanding (NLU) misinterpretations and labeling data for machine learning. High similarity of non-identical consecutive queries indicates ASR mistranscriptions. Consecutive queries with close vectors in a semantic embedding space indicates NLU misinterpretations. Key phrases and barge-in also indicate errors. Only queries within a short amount of time are considered.
1. A method for detecting an inaccurate response to a query, the method comprising:
receiving a first query from a user, the first query having an end time and a first transcription;
providing a response to the first query;
receiving a second query from the user within a short time period of receiving the first query, the second query having a second transcription,
comparing the second transcription to the first transcription to determine the similarity between the first transcription and the second transcription, wherein the step of comparing includes:
transforming a sequence of tokens of the first transcription into a first embedding vector that represents the meaning of the first query;
transforming a sequence of tokens of the second transcription into a second embedding vector that represents the meaning of the second query; and
calculating the query similarity as the distance between the first embedding vector and the second embedding vector; and
determining that the response is inaccurate when the first query is similar to the second query based on determination of similarity between the first transcription and the second transcription.
2. The method of claim 1 , wherein the second query interrupts the step of providing the response.
3. The method of claim 1 , wherein the step of comparing includes analyzing the first query relative to the second query by calculating the edit distance.
4. The method of claim 1 , wherein the second query is received within thirty (30) seconds of the end time of the first query.
5. A method of identifying a mistranscription of any query, the method comprising:
receiving a query at a first time, the query having an end time;
deriving a first transcription for the query received at the first time;
providing a response to the query;
receiving the query at a second time;
deriving a second transcription for the query at the second time, the second transcription being different from the first transcription;
comparing, when a time lapse between the first time and the second time is less than thirty (30) seconds, the first transcription with the second transcription to determine if a level of similarity between the first transcription and the second transcription exceeds a threshold; and
determining that there is a mistranscription when the level of similarity between the first transcription and the second transcription exceeds the threshold, wherein determining that there is a mistranscription depends on determining that the first transcription of the query at the first time includes sufficient semantic similarity to the second transcription of the query at the second time based on a distance between semantic embedding vectors.
6. The method of claim 5 , wherein determining that there is a mistranscription depends on determining that the query at the second time interrupts the response.
7. The method of claim 5 , wherein determining that there is a mistranscription depends on determining that the query at the second time includes a key phrase that indicates the first transcription was wrong.
8. The method of claim 5 wherein determining that there is a mistranscription depends on determining that the second transcription is derived from a text query.
9. A method of identifying a mistranscription of any query, the method comprising:
receiving a query at a first time, the query having an end time;
deriving a first transcription for the query received at the first time;
providing a response to the query;
receiving the query at a second time;
deriving a second transcription for the query at the second time, the second transcription being different from the first transcription;
comparing, when a time lapse between the first time and the second time is less than thirty (30) seconds, the first transcription with the second transcription to determine if a level of similarity between the first transcription and the second transcription exceeds a threshold, wherein the step of comparing includes:
transforming a sequence of tokens of the first transcription into an embedding vector that represents the first query;
transforming a sequence of tokens of the second transcription into an embedding vector that represents the second query; and
calculating the query similarity as the distance between the embedding vectors; and
determining that there is a mistranscription when the level of similarity between the first transcription and the second transcription exceeds the threshold.
10. A method of labeling query data comprising:
determining, for a query, a first transcription;
identifying, for the query at a first time, a first end time;
determining, for the query at a second time, a second transcription, the second transcription being non-identical to that of the first transcription;
calculating a query similarity between the first transcription and the second transcription based on edit distance; and
labeling the first transcription in response to the query similarity calculation exceeding a threshold value;
transforming a sequence of tokens of the first transcription into an embedding vector that represents the query at the first time;
transforming a sequence of tokens of the second transcription into an embedding vector that represents the query at the second time; and
calculating the query similarity as the distance between the embedding vectors.
11. The method of claim 10 further comprising the step of generating a response, wherein the labeling depends on determining that the query at the second time barges in during the response.
12. The method of claim 10 wherein the query at the first time and the query at the second time have high query similarity and small edit distance and the first transcription is labelled as a mistranscription.
13. The method of claim 10 wherein the query at the first time and the query at the second time have high query similarity and large edit distance and the first transcription is labelled as an intended transcription.
14. The method of claim 10 wherein labeling depends on determining that the query at the second time includes an indicator with respect to at least one word in the query.
15. The method of claim 14 wherein the indicator is at least one of emphasis, loudness, tone, and word duration.
16. The method of claim 10 further comprising the step of:
calculating a phoneme edit distance between the first transcription and the second transcription; and
in response to the edit distance being below a threshold, labeling the first query as a mistranscription.
17. The method of claim 10 further comprising the steps of:
calculating a query similarity between the transcription of the first query and the transcription of the second query; and
in response to the query similarity being above a threshold, labeling the first query as a natural language understanding (NLU) misinterpretation.
18. A method of unsupervised labeling of query errors, the method comprising:
reading, for a first query, a first transcription and identifying an end time;
reading, for a second query, a second transcription that is non-identical to the first query and identifying a start time, the start time being less than 30 seconds after the end time;
calculating an edit distance between the first transcription and the second transcription;
calculating a first sentiment vector for the first transcription;
calculating a second sentiment vector for the second transcription;
calculating a vector distance between the first sentiment vector and the second sentiment vector; and
computing each of a transcription error probability and a natural language understanding (NLU) misinterpretation probability using trained models,
wherein each probability is related to the vector distance, each probability is inversely related to the edit distance, and the transcription error probability exceeds the NLU misinterpretation probability for large vector distances.