We describe a deep neural model for pronominal anaphora resolution in Amharic, a Semitic language of Ethiopia. In Amharic, pronouns may appear as standalone words or implicit ones, i.e. morphemes of other words. Both carry morphological features that can be used to determine their potential antecedents among all the mentions of entities in a text. To achieve this, the model computes span embeddings that combine context-dependent boundary representations with a head-finding attention mechanism and prunes the space of possible pronoun-antecedent pairs in a coarse-to-fine and heuristic approach. To facilitate a proper treatment of implicit pronouns the model combines character-based morpheme embeddings with pre-trained word embeddings by means of a BiLSTM recurrent neural network that captures a given word's morphological and semantic features. These representations are contextualized by applying another BiLSTM neural network. We also include an attention mechanism over the words in each span to indicate its headword. The model achieves F1 scores of 50.27% and 50.03% on the test data when using a heuristic or the coarse-to-fine approach, respectively.

In information retrieval (IR), documents that match the query are retrieved. Search engines usually conflate word variants into a common stem when indexing documents because queries and documents do not need to use exactly the same word variant for the documents to be relevant. Stemmers are known to be effective in many languages for IR. However, there are still languages where stemmers or morphological analyzers are missing; this is the case for Amharic which is the working language of Ethiopia. Morphological analysis is the key to derive stems, roots (primary lexical units) and grammatical markers of words such as person, tense and negation markers. This paper presents morphologically annotated Amharic lexicons as well as stem-based and root-based morphologically annotated corpora which could be used by the research community as benchmark collections either to evaluate morphological analyzers or information retrieval for Amharic. Such resources are believed to foster research in Amharic IR.

Automatic Speech Recognition (ASR) is one of the most important technologies to support spoken communication in modern life. However, its development benefits from large speech corpus. The development of such a corpus is expensive and most of the human languages, including the Ethiopian languages, do not have such resources. To address this problem, we have developed four large (about 22 hours) speech corpora for four Ethiopian languages: Amharic, Tigrigna, Oromo and Wolaytta. To assess usability of the corpora for (the purpose of) speech processing, we have developed ASR systems for each language. In this paper, we present the corpora and the baseline ASR systems we have developed. We have achieved word error rates (WERs) of 37.65%, 31.03%, 38.02%, 33.89% for Amharic, Tigrigna, Oromo and Wolaytta, respectively. This results show that the corpora are suitable for further investigation towards the development of ASR systems. Thus, the research community can use the corpora to further improve speech processing systems. From our results, it is clear that the collection of text corpora to train strong language models for all of the languages is still required, especially for Oromo and Wolaytta.

In this paper, we describe an attempt towards the development of parallel corpora for English and Ethiopian Languages, such as Amharic, Tigrigna, Afan-Oromo, Wolaytta and Ge’ez. The corpora are used for conducting a bi-directional statistical machine translation experiments. The BLEU scores of the bi-directional Statistical Machine Translation (SMT) systems show a promising result. The morphological richness of the Ethiopian languages has a great impact on the performance of SMT specially when the targets are Ethiopian languages. Now we are working towards an optimal alignment for a bi-directional English-Ethiopian languages SMT.

Semantic networks have become key components in many natural language processing applications.This paper presents an automatic construction of Amharic semantic networks using Amharic WordNet as initial knowledge base where intervening word patterns between pairs of concepts in the WordNet are extracted for a specific relation from a given text.For each pair of concepts which we know the relationship contained in Amharic WordNet, we search the corpus for some text snapshot between these concepts.The returned text snapshot is processed to extract all the patterns having n-gram words between the two concepts.We use the WordSpace model for extraction of semantically related concepts and relation identification among these concepts utilizes the extracted text patterns.The system is designed to extract "part-of" and "type-of" relations between concepts which are very popular and frequently found between concepts in any corpus.The system was tested in three phases with text corpus collected from news outlets, and experimental results are reported.

This paper presents an approach for recognition of Ethiopic braille documents. We have used direction field tensor for noise removal and isolation of braille dots from their background. A new skewness correction technique, which exploits the horizontal direction nature of braille dots, is designed. A half character derection method, which differentiates braille dots from noise, is applied. Having recognized the half characters through analysis of their dot positions, the braille cells are formulated by way of examining horizontal distances between the half characters. The system is also designed to determine braille dot sizes automatically so that it works for different braille documents with different braille dot sizes. We have finally used a table lookup for translating each braille cell into its corresponding Ethiopic print character.

In this paper we propose algorithms and methods that address critical issues in developing a general Amharic text-to-speech synthesizer. Converting grapheme to phoneme in Amharic is a very challenging task because of the two necessary and yet orthographically unrepresented components of the language -- epenthesis and gemination. Modeling prosodic features of various speaking styles is also the other challenging task in developing Amharic TTS. Making use of orthographic property of verbs in their perfect form, this work introduces set of rules that can be used to locate phones that need to be stressed and algorithm that inserts epenthetic vowel wherever necessary. This paper also presents methods to represent intonations that vary according to punctuation marks ("?" and "#") and phoneme location. In addition to these issues we also introduce transliteration rule for numbers and phone labeling scheme that captures allophonic variations in regard to gemination, epenthesis, and declarative versus judgment request intonations. The system is evaluated for naturalness and intelligibility by ten fluent speakers of the language and experimental results are reported.

This paper presents a novel approach to Amharic document image retrieval by taking the morphology of the language into account. In addition to the general problems and issues concerning document image retrieval systems, Amharic poses further difficulties in modeling retrieval systems due to its complex morphology. We encode the morphological characteristics of the language to improve query formulation and image database indexing. In this work, morphological generator is used to automatically synthesize surface words from a lexicon containing Amharic root forms resulting in surface word image features coded with their respective root forms. Using this morphological coding, document word images and query terms are processed to be represented by their root forms. In the process of indexing and query formulation, cosine similarity is used for comparing word image features extracted from vertical projection, upper bound profile and lower bound profile. The proposed system is tested by using real-life Amharic documents collected from various sources and experimental results are reported.

Amharic is the official language of Ethiopia and uses Ethiopic script for writing. In this paper, we present writer-independent HMM-based Amharic word recognition for offline handwritten text. The underlying units of the recognition system are a set of primitive strokes whose combinations form handwritten Ethiopic characters. For each character, possibly occurring sequences of primitive strokes and their spatial relationships, collectively termed as primitive structural features, are stored as feature list. Hidden Markov models for Amharic words are trained with such sequences of structural features of characters constituting words. The recognition phase does not require segmentation of characters but only requires text line detection and extraction of structural features in each text line. Text lines and primitive structural features are extracted by making use of direction field tensor. The performance of the recognition system is tested by a database of unconstrained handwritten documents collected from various sources.

This paper describes an offline handwriting recognition system for Amharic words based on lexicon. The system computes direction fields of scanned handwritten documents, from which pseudo-characters are segmented. The pseudo-characters are organized based on their proximity and direction to form text lines. Words are then segmented by analyzing the relative gap between subsequent pseudo-characters in text lines. For each segmented word image, the structural characteristics of pseudo-characters are syntactically analyzed to predict a set of plausible characters forming the word. The most likelihood word is finally selected among candidates by matching against the lexicon. The system is tested by a database of unconstrained handwritten Amharic documents collected from various sources. The lexicon is prepared from words appearing in the collected database.

In real life, documents contain several font types, styles, and sizes. However, many character recognition systems show good results for specific type of documents and fail to produce satisfactory results for others. Over the past decades, various pattern recognition techniques have been applied with the aim to develop recognition systems insensitive to variations in the characteristics of documents. In this paper, we present a robust recognition system for Ethiopic script using a hybrid of classifiers. The complex structures of Ethiopic characters are structurally and syntactically analyzed, and represented as a pattern of simpler graphical units called primitives. The pattern is used for classification of characters using similarity-based matching and neural network classifier. The classification result is further refined by using template matching. A pair of directional filters is used for creating templates and extracting structural features. The recognition system is tested by real life documents and experimental results are reported.