Automatic classification and translation of written segments

A non-transitory machine-readable medium having instructions stored thereon, the instructions executable by the one or more processors to perform:

• receiving, by a processor of a first computer, a first training set of segments, including a first plurality of segments in a source language that have been classified as non-translatable to a target language and a held-out set;
• creating and storing in memory, by the processor, a first statistical model from the first training set for classifying segments based on a translatability from the source language to the target language,
• the first statistical model accepting an input segment and producing a score for the input segment that indicates a translatability of the segment;
• determining a set of criteria that converts the score into a classification as non-translatable to a target language or translatable to the target language,
• the set of criteria including a first threshold on scores produced by the first statistical model for classifying segments as non-translatable from the source language to the target language,
• the first threshold maximizing recall while maintaining a certain level of precision with respect to the held-out set;
• receiving a document comprising a list of segments in the source language from a second computer over a communication network;
• determining a score and a classification for each of the list of segments using the first statistical model and the set of criteria;
• in response to the classification of one of the list of segments as non-translatable into the target language, causing the second computer to display the one segment as the translation in the target language.

A system for automatically classifying written segments and translating written segments between human-understandable languages, comprising:

• one or more processors;
• a non-transitory machine-readable medium having instructions stored thereon, the instructions executable by the one or more processors to perform:
• receiving a first training set of segments, including a first plurality of segments in a source language that have been classified as non-translatable to a target language and held-out set;
• creating and storing in memory a first statistical model from the first training set for classifying segments based on a translatability from the source language to the target language,
• the first statistical model accepting an input segment and producing a score for the input segment that indicates a translatability of the segment;
• determining a set of criteria that converts the score into a classification as non-translatable to the target language or translatable to the target language,
• the set of criteria including a first threshold on scores produced by the first statistical model for classifying segments as non-translatable from the source language to the target language,
• the first threshold maximizing recall while maintaining a certain level of precision with respect to the held-out set;
• receiving a document comprising a list of segments in the source language from a remote computer over a communication network;
• determining a score and a classification for each of the list of segments using the first statistical model and the set of criteria;
• in response to the classification of one of the list of segments as non-translatable into the target language, causing the second computer to display the one segment as the translation in the target language.